DEV Community: Karl Wirth

Better team decisions with AI

Karl Wirth — Thu, 06 Aug 2026 15:00:00 +0000

Peter Drucker said "Making good decisions is a crucial skill at every level." AI can help your team with this.

This blog lays out a four-phase framework for a teams to make smarter, more collaborative decisions with AI.

Your team could follow this framework synchronously or asynchronously, whether in a team meeting or across async conversations. If its a small decision, you could move through the whole process (skipping some steps) in 20 minutes. If its a large decision, you could move through it over months.

My Principles

Don't be lazy and have AI make the decision for you
Do use AI to help you do things you know you should do but probably wouldn't do, to brainstorm and add to your thinking, and to formalize your work / flesh out your notes.

1. Frame your decision

What is the best question to answer?

Write down the question you are deciding
What kind of decision is this? (Go / No Go, Decide among alternatives, Prioritize among alternatives)
- Prompt: I'm working on a decision and need help clarifying what type of decision it actually is. The decision I’m facing is: [INSERT DECISION DESCRIPTION]. Please help me determine whether this is best framed as a Go / No Go decision (a binary choice to proceed or not), a decision among multiple alternatives (selecting the best option from a set), or a prioritization decision (ranking or sequencing alternatives based on importance, timing, or impact). If the decision doesn't cleanly fit into one of these categories, suggest how to best frame it for effective analysis and planning. Also help me understand how the decision type might affect the process, tools, or criteria I should use.
Refine the decision question (What question you choose to answer is key to the outcome)
- Make explicit the assumptions behind the question
- Possibly widen the question, narrow it, state it in reverse, or switch the kind of decision
- Prompt: I'm working on a decision and want to refine how I'm framing the question. My current decision question is [INSERT YOUR QUESTION HERE]. Please help me explore this decision from multiple angles by making the question more precise, widening the question to consider broader implications, narrowing it into specific sub-questions, reversing it to consider what I should avoid or the opposite path, switching decision types (e.g., from yes/no to multiple choice or priority ranking), and challenging any assumptions in the original phrasing. For each of these approaches, provide 2-3 specific reworded versions of my question and suggest which reframing might help most in reaching a clear and informed decision.
Finalize the question you will be deciding

What is the context?

Write down your goals and your values
Describe the context of the decision
- Prompt: I'm working on a decision and need help completing the context. My goal is [INSERT GOAL], and here’s what I currently know: [INSERT CURRENT CONTEXT DESCRIPTION]. Please help me fill in missing context by suggesting key environmental factors, relevant stakeholders, important constraints, historical precedents, market or industry influences, and any timeline considerations I should be factoring in.
Assess the decision's importance and urgency on a 2x2
- Prompt: I need to assess where this decision fits relative to my other recent decisions. Here’s my current decision: [INSERT DECISION DESCRIPTION]. These are my other recent major decisions: [LIST OTHER RECENT DECISIONS]. Please help me place this decision and the others on a 2x2 matrix of Importance (High/Low) vs. Urgency (High/Low), suggest how much time and resources to allocate based on its position, and identify whether I'm misallocating effort relative to its significance.
State the amount of time, effort, resources you will spend on the decision

What is your plan for making the decision?

Create a plan for making the decision
- Could be simple and you make the decision in the next 20 minutes. Could be a months long decision
- Prompt: I need to create a plan for making this decision: [INSERT DECISION DESCRIPTION]. The deadline is [INSERT DEADLINE OR "flexible"], and the complexity level is [Simple/Moderate/Complex]. Please generate a decision-making plan including key phases, timelines for each phase, who should be involved at each step, what information or analysis needs to be gathered, decision checkpoints, and fallback options in case more time or information is needed.
Generate a list of stakeholders and input from them (consider how AI might reshape your org structure and who the right decision-makers are)
- Prompt: "I need help creating a comprehensive RACI chart for [INSERT YOUR PROJECT/PROCESS NAME HERE]. Please identify all relevant stakeholders who would be involved in this project (including internal teams, external partners, customers, regulators, etc.), then for each stakeholder provide their typical input regarding their roles, responsibilities, what they need to be accountable for, what they need to be consulted on, and what they need to be informed about. Using this stakeholder analysis, create a detailed RACI matrix showing all key activities/deliverables as rows, all stakeholders as columns, and clear R/A/C/I assignments for each intersection with brief explanations for any complex assignments. Project context: [ADD YOUR SPECIFIC PROJECT DETAILS, SCOPE, AND CONSTRAINTS HERE]."

2. Work on your decision

If it's a big or complex decision, break down the decision into smaller parts
- Prompt: I'm facing a big or complex decision and need help breaking it down into more manageable components. The overall decision is: [INSERT DECISION DESCRIPTION]. Please help me identify the smaller, more specific sub-decisions that make up this larger decision. These might include sequencing choices, resource allocation, stakeholder involvement, timing considerations, or parallel decisions that need to be addressed independently. For each sub-decision, suggest how it could be tackled on its own and how it connects to the overall decision. This will help me approach the larger decision in a clearer, more structured way.
If choosing one of several alternatives:
- List alternatives
- Prompt: I'm working on this decision: [INSERT DECISION] and need help brainstorming alternatives. I've already considered these options: [LIST CURRENT ALTERNATIVES]. Please suggest new and creative alternatives, including radically different approaches, hybrids of current ideas, out-of-the-box options that challenge assumptions, phased or pilot strategies, and alternatives with different resource, timeline, or risk profiles. Aim for 5–10 new alternatives.
- List criteria
- Prompt: For this decision: [INSERT DECISION], I need help developing thorough evaluation criteria. I’ve already identified these criteria: [LIST CURRENT CRITERIA]. Please suggest additional criteria, including financial (e.g., cost, ROI), strategic (e.g., goal alignment), risk-based (e.g., operational or market risks), timeline-related (e.g., time to value), resource-related (e.g., skills or tech needed), stakeholder-related (e.g., customer or employee impact), and quality-related (e.g., scalability or reliability).
- Make a table comparing the alternatives against the criteria
- Prompt: I need to create a comparison table for a decision. My alternatives are: [LIST ALTERNATIVES], and the criteria are: [LIST CRITERIA]. Please build a table with criteria as columns and alternatives as rows. For each criteria, make two columns, one with a text explanation, and one with a rating. Rate each alternative consistently from 1 to 5 with 5 being good and 1 being weak. Highlight the best performer per criterion, and identify clear winners or weak options.
- Make a reduced table with the top alternatives against the most important criteria
- Prompt: Based on a full comparison of all alternatives and criteria, I now want to narrow the focus to the most promising options. Please help me create a reduced comparison table that includes only the top-performing alternatives and the most important criteria. Use the initial table to identify which alternatives scored highest overall or stood out in key areas, and which criteria had the most influence on decision quality or tradeoffs. Build a new table that includes these selected alternatives and criteria, showing both text explanations and consistent 1–5 ratings (with 5 being strong). Highlight the strongest performers and flag any remaining tradeoffs to support a final decision.

If prioritizing alternatives:

List alternatives
- Prompt: I'm working on this decision: [INSERT DECISION] and need help brainstorming alternatives. I've already considered these options: [LIST CURRENT ALTERNATIVES]. Please suggest new and creative alternatives, including radically different approaches, hybrids of current ideas, out-of-the-box options that challenge assumptions, phased or pilot strategies, and alternatives with different resource, timeline, or risk profiles. Aim for 5–10 new alternatives.
List criteria
- Prompt: For this decision: [INSERT DECISION], I need help developing thorough evaluation criteria. I’ve already identified these criteria: [LIST CURRENT CRITERIA]. Please suggest additional criteria, including financial (e.g., cost, ROI), strategic (e.g., goal alignment), risk-based (e.g., operational or market risks), timeline-related (e.g., time to value), resource-related (e.g., skills or tech needed), stakeholder-related (e.g., customer or employee impact), and quality-related (e.g., scalability or reliability).
Make a table evaluating the alternatives based on the criteria
- Prompt: I need to create a comparison table for a decision. My alternatives are: [LIST ALTERNATIVES], and the criteria are: [LIST CRITERIA]. Please build a table with criteria as columns and alternatives as rows. For each criteria, make two columns, one with a text explanation, and one with a rating. Rate each alternative consistently from 1 to 5 with 5 being good and 1 being weak. Highlight the best performer per criterion, and identify clear winners or weak options.
Your team prioritizes the alternatives directly or develop a scoring mechanism
- Prompt: Please help me create a scoring mechanism that assigns appropriate weights to each criterion based on its importance, defines a consistent scoring scale (e.g., 1–5 or 1–10), and applies it across all alternatives. Then calculate total weighted scores for each option, rank them from highest to lowest, and identify any close calls or ties that might require further consideration. Also include a brief explanation of how each score was derived and any insights about the strongest or weakest options.
If a go / no go decision:
- List pros/cons of the decision.
- Turn the pros/cons into list of reasons to go and not go
- Prompt: I’ve listed pros and cons for this decision: [INSERT DECISION]. Please convert these into more actionable and specific lists: "Reasons to proceed" (benefits and positive outcomes) and "Reasons not to proceed" (risks and downsides). Add any reasons I may have missed, group similar points, and make them measurable where possible.
- Rank the reasons to go by importance (Critical, Essential, Important, Unimportant)
- Generate a Go / No Go comparison
- Prompt: I'm considering a Go / No Go decision: [INSERT DECISION DESCRIPTION]. Please help me generate a structured comparison by organizing the reasons to proceed ("Go") and the reasons to hold back ("No Go") into four priority levels: Critical, Essential, Important, and Unimportant. For each side, group and label the reasons accordingly, making sure they reflect both the strength of the argument and the potential impact on the outcome. Where possible, make each reason specific, measurable, and actionable. This breakdown should help clarify the weight of each consideration and guide a more confident decision.

3. Make the Decision

Ask team members to write down their gut for the decision and share it
Consider a potential decision and write it down
Get people on the team's feedback.
- And/Or use AI to simulate the feedback of a variety of roles, stakeholders
- Prompt: I'm considering this decision: [INSERT DECISION DESCRIPTION] and would like simulated stakeholder feedback. Please think through the right set of roles that should provide feedback for this decision given our context. For each, offer both supportive and critical feedback. Optionally, add a specific role and their perspective if I insert one.
Play devil's advocate on the decision and tell you why it's a bad idea
- Prompt: I'm considering this decision: [INSERT DECISION] and want you to play devil’s advocate. Please argue against the decision by highlighting risks, questioning assumptions, identifying what could go wrong, flagging resource or timing issues, and challenging whether it aligns with my goals. Provide strong critical reasoning to stress-test my thinking.
Do a pre-mortem or write a press release on the decision
- Prompt: I'm about to make this decision: [INSERT DECISION] and want to conduct a pre-mortem. Please imagine the decision failed badly and write a scenario describing what went wrong, when it started, who was affected, what was missed, and the root causes. Then outline preventive steps I can take. Afterward, write a “press release” describing the successful outcome of this decision to help me visualize a positive future.
Check the decision against your values and your goals
- Prompt: Please evaluate this decision: [INSERT DECISION] in relation to my values and goals. My values are: [LIST VALUES], and my goals are: [LIST GOALS]. Provide an alignment score for each value/goal (1–10), highlight conflicts, identify opportunities to strengthen alignment, suggest adjustments, and assess overall consistency with who I want to be and what I aim to achieve.
Expand on possible second and third order effects of the decision
- Prompt: I’m considering this decision: [INSERT DECISION] and want an analysis of its ripple effects. Please identify second-order effects like team or process impacts, third-order effects like culture or market responses, and fourth-order or beyond effects such as unintended consequences or systemic changes. Include both positive and negative potential impacts across short- and long-term horizons.
Have team members express their ideas and their emotions about the potential decision. Encourage honesty and take the emotions seriously.
Iterate on this step and working on the decision until you are satisfied or time is up.
Make the decision tentatively, take it to superiors, stakeholders etc...

4. Record, Act on, and Learn from the Decision

Record the decision and the reasons for it
- Prompt: Please create a decision record for this decision: [INSERT DECISION]. Include the decision made, date, decision makers, and stakeholders consulted. For context, explain the problem or opportunity, alternatives considered, evaluation criteria, and key decision factors. For implementation, list next steps, timeline, resources needed, and success metrics. Finally, outline a review plan with a future review date, evaluation focus, and triggers for revisiting the decision.
Write down what success will look like and how you will measure it
Write down when you will revisit the decision or do a retro and do so

Next Steps

Try this framework on a team decision using an AI
Or, even better, download Nimbalyst to collaborate together with AI on your decisions.
See how AI team tools compare for collaborative decision-making.

Better team meetings with AI

Karl Wirth — Thu, 06 Aug 2026 13:00:00 +0000

Meetings are essential to teamwork, yet they're often the source of frustration, wasted time, and missed opportunities. AI presents an opportunity to reimagine how teams come together, collaborate, and make decisions in meetings.

By thoughtfully integrating AI into our meeting workflows, we can shift from information-heavy status sessions to more productive, informed conversations that drive better decisions.

Intelligent Pre-Meeting Preparation

AI excels at homework and preparation tasks that traditionally burden meeting participants. Before teams gather, AI can:

Synthesize relevant background materials and pre-reading
Analyze previous meeting notes and action items
Prepare briefing documents tailored to each participant's role

Examples:

Product Status Update: AI reviews last week's meeting notes, the plan, and the commits and conversations in the intervening weeks, and prepares a summary of status.
Client Presentation Prep: AI researches the client's recent news, financial performance, and industry challenges, then creates talking points and potential objections with suggested responses for each sales team member.

This preparation ensures everyone arrives informed and ready to contribute meaningfully, rather than spending meeting time getting up to speed.

Eliminating Status Update Theater

Too many meetings exist simply to communicate status updates, progress reports, and routine information sharing. AI can handle much of this communication asynchronously, freeing up meeting time for genuine collaboration.

Examples:

Weekly Standups: Instead of spending 30 minutes having each person report their progress, AI automatically pulls updates from project management tools, code repositories, and calendars, then generates a visual dashboard highlighting blockers, achievements, and upcoming deadlines.
Project Status Communications: AI monitors Slack, email, and project tools to automatically generate stakeholder updates, escalate issues that need attention, and schedule check-ins only when human intervention is actually needed.

When teams use AI to share status updates, track project progress, and communicate routine information outside of meetings, they can reserve their face-to-face time for activities that truly require human interaction: creative brainstorming, complex problem-solving, strategic planning, and relationship building. Reducing this coordination overhead is one of the most impactful ways AI reshapes how teams work.

From Information Asymmetry to Shared Intelligence

During the meeting, AI can address information asymmetry by providing rapid fact checks, research, analysis, and summarization as discussions unfold. Instead of taking action items to look up data that is key to a strategic question and instead of relying on hearsay or the loudest person's assertions, teams can access relevant information instantly, keeping the conversation flowing and ensuring everyone operates from the same foundation of knowledge.

Examples:

Market Size Discussions: When someone mentions "the growing cybersecurity market," AI instantly provides current market size, growth rate and key segments, allowing the team to make data-driven decisions without derailing the conversation.
Competitive Analysis: During a product feature debate, AI quickly pulls up competitor feature comparisons from your internal documents and recent customer reviews (from the web), giving everyone the same baseline information to evaluate strategic options.

This during-the-meeting information depth and rigor can reduce one of the greatest risks of meetings: the tradeoff between the need to make progress while everyone is gathered and the need to have sufficient and accurate data to inform the discussion.

Better Meeting Notes

Taking good meeting notes is essential to not lose key points, questions, context, decisions, and action items. It is hard to do this well while also participating fully in the discussion. AI can free up all to participate while it takes a full transcript and provides the summary notes.

Examples:

Decision Tracking: AI identifies when the team decides to "pivot the marketing strategy" and automatically creates action items with owners, deadlines, and success metrics, while flagging that the budget implications weren't fully discussed.
Conflict Resolution: During a heated debate about resource allocation, AI notes that one person advocated for hiring developers while another pushed for marketing spend, and flags this as an unresolved tension requiring follow-up discussion
Action Items: AI tracks when team members make commitments or are assigned actions and automatically creates action items and follow-up check-ins, ensuring nothing falls through the cracks.

The person who takes meeting notes has traditionally held significant power in that they shape the narrative, highlight certain decisions, and influence follow-up actions. AI can democratize this process by taking a full transcript and by linking summary points to that transcript.

AI as Meeting Contributor vs. Secretary

The most powerful applications of AI in team meetings go beyond having a digital secretary. Instead of just taking notes and scheduling follow-ups, AI can participate as a collaborative team meeting contributor:

Suggesting relevant frameworks or methodologies
Offering alternative perspectives
Identifying potential blind spots in team thinking
Simulating the perspectives of different stakeholders

Examples:

Strategic Planning: When the team focuses on expanding to new markets, AI suggests considering the "Jobs to Be Done" framework and points out they haven't discussed customer acquisition costs in emerging markets or regulatory barriers.
Product Development: During feature prioritization, AI plays devil's advocate by highlighting that the proposed AI chatbot feature might cannibalize their premium support tier revenue, forcing the team to think through business model implications.
Crisis Response: When discussing a PR crisis, AI suggests the team consider second and third-order effects they haven't mentioned, such as impact on employee morale and partner relationships, while recommending crisis communication frameworks from similar industry situations.
Facilitated Decision-Making: AI facilitates team decisions by anonymously collecting input from all members, identifying common themes and outlier perspectives, and presenting options without revealing who suggested what, reducing hierarchy bias and groupthink.

Today, this kind of help from AI can be unlocked by asking the AI questions like "What are we missing in this strategic plan so far?" or "Play devil's advocate?" or "What else should we consider in dealing with this PR crisis?" or "What would you add to the discussion so far?"

Better Meetings

The goal isn't to replace human meetings with AI, but to eliminate the mundane aspects that prevent teams from reaching their full collaborative potential. When AI handles information gathering, note-taking, aspects of preparation, and routine communication, human participants can focus thinking, discussion, strategy, decisions, and relationship. Removing this friction from your process lets teams spend time on what actually matters.

While AI can augment each team member's work individually, the greatest value comes from AI that empowers the whole team together.

Next Steps

Try one use case or example above for your next meeting.
Download Nimbalyst to collaborate together with your team members and AI

ChatGPT vs Claude: Features & Pricing Compared

Karl Wirth — Wed, 05 Aug 2026 21:00:00 +0000

Let's compare the features and pricing tiers for the leading AI vendors, Claude and ChatGPT with a particular focus on the the two tiers they each provide for teamwork: Teams and Enterprise.

What are the differences between Claude and ChatGPT, what additional features do these tiers provide over the ChatGPT Plus / Claude Pro tiers, and how much do they help with teamwork? For a deeper dive on ChatGPT's team offering, see our complete guide to ChatGPT Teams.

ChatGPT and Claude Pricing Tiers, Part 1 of 2

| | | ChatGPT | | | | | Claude | | | |
| Area | SubArea | Free | Plus | Teams | Enterp | | Free | Pro | Teams | Enterp |
| Price | Pricing ($/person/month) | $0 | $20 | $30 | ~$60 | | $0 | $20 | $30 | ~$60 |
| Inputs | Text, Images, PDFs | ✓ | ✓ | ✓ | ✓ | | ✓ | ✓ | ✓ | ✓ |
| Web | ✓ | ✓ | ✓ | ✓ | | ✓ | ✓ | ✓ | ✓ |
| Speech | ✓ | ✓ | ✓ | ✓ | | ✗ | ✗ | ✗ | ✗ |
| Google Drive, Microsoft OneDrive | ✓ | ✓ | ✓ | ✓ | | ✗ | ✓ | ✓ | ✓ |
| Record mode (meeting record) | ✗ | ✗ | ✓ | ✓ | | ✗ | ✗ | ✗ | ✗ |
| MCP and Internal Systems | ✗ | ✓- | ✓ | ✓ | | ✗ | ✓ | ✓ | ✓ |
| Deep Research | ✓- | ✓ | ✓ | ✓ | | ✗ | ✓ | ✓ | ✓ |
| Models | Latest Models | ✓-- | ✓ | ✓ | ✓ | | ✓-- | ✓ | ✓ | ✓ |
| Tokens | ✓-- | ✓- | ✓ | ✓+ | | ✓-- | ✓- | ✓ | ✓+ |
| Response time | ✓- | ✓ | ✓ | ✓ | | ✓- | ✓ | ✓ | ✓ |
| Context Window | 8K | 32K | 32K | 128K | | Less | 200K | 200K | 200K |
| Memory | ✓- | ✓ | ✓ | ✓ | | ✗ | ✓ | ✓ | ✓ |
| Workspace | Chats | ✓ | ✓ | ✓ | ✓ | | ✓ | ✓ | ✓ | ✓ |
| Canvas / Artifacts | ✓ | ✓ | ✓ | ✓ | | ✓ | ✓ | ✓ | ✓ |
| Projects | ✗ | ✓ | ✓ | ✓ | | ✗ | ✓ | ✓ | ✓ |

Blue highlights difference between Teams and ChatGPT Plus / Claude Pro
X indicates the feature is not offered or supported
The Check Minus or Plus indicates the feature is offered but limited or enhanced

Observations

Claude and ChatGPT offer strikingly similar functionality
Claude and ChatGPT have chosen to name and price the same
There is very little difference between both vendors' Teams offering and the ChatGPT Plus and Claude Pro tiers
In this section, only ChatGPT's Record mode and Connection to Internal systems are related to team collaboration.

ChatGPT and Claude Pricing Tiers, Part 2 of 2

| | | ChatGPT | | | | | Claude | | | |
| Area | SubArea | Free | Plus | Teams | Enterp | | Free | Pro | Teams | Enterp |
| Price | Pricing ($/person/month) | $0 | $20 | $30 | ~$60 | | $0 | $20 | $30 | ~$60 |
| Outputs | Speech | ✓- | ✓ | ✓ | ✓ | | ✗ | ✗ | ✗ | ✗ |
| Text, Tables, Charts, Graphs, PDFs | ✓- | ✓ | ✓ | ✓ | | ✓ | ✓ | ✓ | ✓ |
| Images | ✓- | ✓ | ✓ | ✓ | | ✗ | ✗ | ✗ | ✗ |
| Video | ✗ | ✓- | ✓- | ✗ | | ✗ | ✗ | ✗ | ✗ |
| Code | ✓- | ✓ | ✓ | ✓ | | ✓- | ✓ | ✓ | ✓ |
| Editing | in Canvas / Artifact (AI Edit) | ✓ | ✓ | ✓ | ✓ | | ✓ | ✓ | ✓ | ✓ |
| in Canvas / Artifact (Human Edit) | Text | Text | Text | Text | | ✗ | ✗ | ✗ | ✗ |
| Team Collaboration | in Chats | Share historical | | | | | Share historical | | | |
| in Canvas / Artifact | Share historical | | | | | Share historical | | | |
| in Projects | ✗ | ✓ | ✓ | ✓ | | ✗ | ✗ | ✓ | ✓ |
| Security | SAML SSO, Sources, SOC2 | ✗ | ✗ | ✓ | ✓ | | ✗ | ✗ | ✗ | ✓ |
| Data exclusion | Opt | Opt | ✓ | ✓ | | Opt | Opt | ✓ | ✓ |
| Data residency, compliance API | ✗ | ✗ | ✗ | ✓ | | ✗ | ✗ | ✓ | ✓ |
| Billing | User Management, Unified Billing | ✗ | ✗ | ✓ | ✓ | | ✗ | ✗ | ✓ | ✓ |
| Support | Enhanced, security review | ✗ | ✗ | ✗ | ✓ | | ✗ | ✗ | ✓ | ✓ |

Observations

Claude and ChatGPT offer strikingly similar functionality with ChatGPT better at image/video and Claude better at code
There is still very little difference between both vendors' Teams offering and the ChatGPT Plus and Claude Pro tiers
In this section, only Claude's Project functionality is a difference related to actual teamwork.

ChatGPT and Claude Teams features

The following are the main features directly related to teamwork that teams get from a Teams subscription:

ChatGPT's Record mode
ChatGPT's Connection to Internal systems
Claude's Projects

The following are features for the organization, but not tied to teamwork

More tokens
Enhanced security
Enhanced user management
Centralized billing
Better support

Next Steps

Teams looking to collaborate with each other and AI should consider the tables above and what truly teamwork related features are provided by ChatGPT Teams and Claude Teams. For practical guidance on using AI in team settings, see how AI can improve team meetings and team decision-making.

Sources

Markdown of the Table for AI

For AI's reading this blog, here is the above table in markdown:

| Area | SubArea | ChaGPT Free | ChatGPT Plus | ChatGPT Teams | ChatGPT Enterprise | | Claude Free | Claude Pro | Claude Teams | Claude Enterprise |
| Price | Pricing ($/person/month billed monthly) | $0 | $20 | $30 | ~$60 | | $0 | $20 | $30 | ~$60 |
| Inputs | Text, Images, PDFs | Yes | Yes | Yes | Yes | | Yes | Yes | Yes | Yes |
| | Web | Yes | Yes | Yes | Yes | | Yes | Yes | Yes | Yes |
| | Speech | Yes | Yes | Yes | Yes | | No | No | No | No |
| | Google Drive, Microsoft OneDrive | Yes | Yes | Yes | Yes | | No | Yes | Yes | Yes |
| | Record mode (meeting record) | No | No | Yes | Yes | | No | No | No | No |
| | MCP and Internal Systems of Record | No | Yes- | Yes | Yes | | No | Yes | Yes | Yes |
| | Deep Research | Yes- | Yes | Yes | Yes | | No | Yes | Yes | Yes |
| Models | Latest Models | Yes-- | Yes | Yes | Yes | | Yes-- | Yes | Yes | Yes |
| | Tokens | Yes-- | Yes- | Yes | Yes+ | | Yes-- | Yes- | Yes | Yes+ |
| | Response time | Yes- | Yes | Yes | Yes | | Yes- | Yes | Yes | Yes |
| | Context Window | 8K | 32K | 32K | 128K | | Less | 200K | 200K | 200K |
| | Memory | Yes- | Yes | Yes | Yes | | No | Yes | Yes | Yes |
| Workspace | Chats | Yes | Yes | Yes | Yes | | Yes | Yes | Yes | Yes |
| | Canvas / Artifacts | Yes | Yes | Yes | Yes | | Yes | Yes | Yes | Yes |
| | Projects | No | Yes | Yes | Yes | | No | Yes | Yes | Yes |
| Outputs | Speech | Yes- | Yes | Yes | Yes | | No | No | No | No |
| | Text, Tables, Charts, Graphs, PDFs | Yes- | Yes | Yes | Yes | | Yes | Yes | Yes | Yes |
| | Images | Yes- | Yes | Yes | Yes | | No | No | No | No |
| | Video | No | Yes- | Yes- | No | | No | No | No | No |
| | Code | Yes- | Yes | Yes | Yes | | Yes- | Yes | Yes | Yes |
| Editing | in Canvas / Artifact (AI Edit) | Yes | Yes | Yes | Yes | | Yes | Yes | Yes | Yes |
| | in Canvas / Artifact (Human Edit) | Text | Text | Text | Text | | No | No | No | No |
| Team Collaboration | in Chats | Share Historical Only | Share Historical Only | Share Historical Only | Share Historical Only | | Share Historical Only | Share Historical Only | Share Historical Only | Share Historical Only |
| | in Canvas / Artifact | Share Historical Only | Share Historical Only | Share Historical Only | Share Historical Only | | Share Historical Only | Share Historical Only | Share Historical Only | Share Historical Only |
| | in Projects | ✗ | Share historical Chats, Canvas | Share historical Chats, Canvas | Share historical Chats, Canvas | | ✗ | Share historical Chats, Artifacts | Share historical Chats, Canvas | Share historical Chats, Canvas |
| Security | SAML SSO, Sources, SOC2 | No | No | Yes | Yes | | No | No | No | Yes |
| | Data exclusion | Opt out | Opt out | Yes | Yes | | Opt out | Opt out | Yes | Yes |
| | Data residency, compliance API | No | No | No | Yes | | No | No | Yes | Yes |
| Billing | User Management, Unified Billing | No | No | Yes | Yes | | No | No | Yes | Yes |
| Support | Enhanced, onboarding,security review | No | No | No | Yes | | No | No | Yes | Yes |

How Best to Edit AI Text

Karl Wirth — Wed, 05 Aug 2026 19:00:00 +0000

AI has accelerated content creation, but you can't just accept what AI gives you as output and use it or publish it. Whether you're working with ChatGPT, Claude, or Gemini as your AI writing assistant, AI text editing is essential.

The problem with raw AI text output

AI generated text is just the starting point. Raw AI output always needs refinement:

Stylistic changes: em-dash removal, icon-removal, enthusiasm-reduction
Accuracy: fact-checking, updating errors
Your tone: editing the content to match your own tone
Substantive improvements: your ideas and edits

Trying to edit AI text in ChatGPT and Claude and Google

Option 1: The Long Chat Thread

The most common approach to edit AI generated text is to do so in the chat thread. You ask the AI to make a change, rewrite a section, add or correct. And AI does so, repeating what you wrote. This approach is limited:

Requesting each change in your chat thread is slow
You lose track of what changed between versions
You have many copied versions of whatever you are working on
Its not collaborative if you are working with a team

Option 2: Use ChatGPT Canvas

ChatGPT does allow you to take a section of your chat and put it into a Canvas. That Canvas is editable at a basic level but this approach also has a number of drawbacks:

Its confusing to keep track of Canvases, go back and forth between your Chat and your Canvases
If you make changes to a Canvas and then go to a Chat and continue to work, the Chat and the Canvas are not kept in synch
Only text is fully editable in the Canvas. Tables, diagrams, etc are not.
And you still can't see what AI changes in the text when it makes a change. It just does it.
Its not collaborative if you are working with a team

Option 3: Copying back and forth to a Google Doc

Because of the native limits of working in ChatGPT and Claude, many people take a section of an AI chat and paste it into a Google Doc, where they edit it for a while and then they paste back into the AI chat a section for further questions, work and refinement. This approach is also problematic:

You are working in two different workspaces
The AI doesn't have the full context of what you are doing and have changed
You must maintain two copies of everything and will lose your place
While Google Doc's is collaborative

Option 4: Full AI text editing in Nimbalyst

An AI editor like Nimbalyst can transform how you work with AI and text. Instead of the tedious chat-based editing process and instead of copying and pasting, you get direct control over your AI-generated content integrated with AI.

Direct AI Text Editing

With Stravu, you can directly edit AI output yourself. No more describing changes in chat prompts. You can simply click and edit the text as you would in any professional editor. This immediate control over AI text makes the editing process intuitive and efficient.

Continued use of AI with full context

And, Nimbalyst lets you connect that edited content to your AI for ongoing work and context. Its the best of both worlds. The power of AI and your ability to directly work and edit your content.

Approve every change

One of the biggest frustrations with AI editing is losing track of AI modifications. AI can add a paragraph, rewrite a whole section, change many parts of a document. If you don't have a way of seeing and reviewing this, then errors or just changes you didn't want could slip by you.

Nimbalyst solves this by letting you see and approve what changed. Additions to your text are marked in green. Deletions are marked in red. You can edit what has changed before finally approving or rejecting the change. You maintain complete visibility into the editing process, ensuring nothing gets lost or overlooked.

Streamline you workflow

Instead of opening multiple canvases or chat threads to manage your content, modern AI editors like Nimbalyst provide a unified workspace where you can:

Generate initial AI text or generate an AI response to your initial text
Make direct edits and improvements
Keep your editing and the AI context united
Track and approve all changes

Best Practices for AI Text Editing

Start with your own ideas: While sometimes, starting with what AI generates can overcome writer's block, its best to start with your own thoughts, brainstorm, or outline and let AI flesh this out.

Always edit AI output: Never publish AI content without review. Even the most sophisticated AI writing tools produce text that benefits from human editing and refinement.

Use the right tools: Choose an AI editor that offers direct editing capabilities rather than relying solely on chat-based revision requests.

Maintain change visibility and approve output: Ensure your AI editing workflow includes clear tracking of what modifications were made with the option to edit those changes before approving or rejecting them.

Iterate efficiently: The best AI text editing happens through multiple quick iterations rather than trying to perfect everything in one pass.

The Future of AI Writing and Editing

As AI writing continues to evolve, the distinction between content generation and content editing is blurring. The most effective content creators are those who bring their own ideas, leverage AI text generation, and do thorough, iterative, collaborative editing of AI text.

Modern AI editors, like Nimbalyst, are enabling this by providing:

Seamless integration between AI generation and human editing
Collaboration between AI and human intelligence
Change tracking and version control for AI-generated content
Intuitive interfaces that make AI text editing as natural as traditional writing

If you're looking to easily edit and improve AI-generated content without the frustration of slow, repetitive chat-based revisions, it's time to explore modern AI editing solutions.

Check out Nimbalyst today and discover how direct AI text editing can improve your content creation process. Regardless, do not accept raw AI output. Edit it!

How Best to Edit AI Tables

Karl Wirth — Wed, 05 Aug 2026 17:00:00 +0000

Raw AI-generated tables like what you make in Claude and ChatGPT are just a starting point. You need to be able to edit your tables, organize them, update them, and collaborate about them. Nimbalyst (formerly Stravu) has introduced editable, collaborative, unified AI tables.

The problem with raw AI table output

AI-generated tables require significant refinement before they're ready for professional use:

Data accuracy: Fact-checking values, updating outdated information, correcting errors
Structure optimization: Reorganizing columns, adjusting row priorities, refining categorization
Completeness: Filling gaps, adding missing data points, expanding analysis
Context integration: Connecting table data to your specific business needs and existing knowledge

The limited options for editing AI tables in ChatGPT, Claude, and other platforms

Option 1: The Tedious Chat Request Loop

The most common approach is asking AI to modify tables through chat commands. You request changes like "add a column for pricing" or "sort by revenue" and hope the AI regenerates the table correctly. This approach has major limitations:

Each modification requires a new chat prompt and full table regeneration
You lose previous versions and can't track what changed
Complex edits often result in AI misunderstanding your intent
No way to collaborate with teammates on table refinement

A typical long chat, duplicating the table for each change you need

Option 2: Copy and Paste to a Spreadsheet

You can copy an AI table to Excel or Google Sheets for editing. While this provides familiar editing tools, it creates significant workflow problems:

You lose the AI context and can't easily get AI assistance with further analysis
Tables become disconnected from your original research and decision-making process
Collaboration happens in isolation from the AI insights that generated the data
No integration between your edited tables and ongoing AI research

Option 3: Canvas

ChatGPT offers a Canvas option which allows you to use AI to update the table (a bit in some cases), but it also has major shortcomings:

Human can only edit the text in a cell and can't work with the table as a table (adding rows, adding columns, filtering, sorting)
The Canvas is an awkward side path from the rest of the AI chat

The problem with startup tools that are only AI table generators

There are a class of AI startups offering focused sheets or table solutions. These have the following problems:

Context Isolation Problem: Table-specific tools operate in isolation from your broader research and analysis workflow. When you switch out of your AI chat or document to work in a table-specific tool, you lose the crucial context that created that table.
Synchronization Nightmare: Keeping tables synchronized with your evolving research or analysis creates constant friction.

Introducing AI table editing in Nimbalyst

Nimbalyst transforms how you work with AI-generated tables by providing powerful tables, direct editing, and team collaboration, while maintaining full AI integration and collaboration features.

Table integrated with text and diagrams: AI can leverage the surrounding notebook context for its work in the table and the human can also relate the text and diagrams to the table.

Direct Cell Editing: With Nimbalyst, you can click into any table cell and edit it directly, just like in Excel or Google Sheets.

Direct Table Changes: The human and AI can both directly add rows, add columns, filter, sort

Stravu AI Tables - add column aI

The Future of AI Table Creation and Analysis

The future of data analysis lies in seamless integration between AI research capabilities and human data editing. The most effective analysts are those who can leverage AI for initial data gathering while maintaining direct control over data refinement and validation.

Modern AI table editors, like Nimbalyst, are advancing this vision by providing:

Unified tables, text, and diagrams for human and AI context -- with review and approval workflows built in
Real-time collaboration between humans and AI on the table
Direct editing capabilities that preserve AI context
Integration between table data and broader decision-making workflows

How Best to Edit AI Diagrams

Karl Wirth — Wed, 05 Aug 2026 15:00:00 +0000

The Problem with AI Diagrams

You've been there. You ask ChatGPT or Claude to create a diagram for your team's process flow. It spits out something decent, but then you realize you need to:

Edit a few boxes because the AI misunderstood your process
Add your team's feedback from three different people who all have opinions
Iterate on the design as your understanding evolves
Reference the diagram in other documents and presentations

But here's the problem: AI-generated diagrams are dead ends. You can't edit them collaboratively. You can't iterate on them with your team. And if you want changes, you're back to prompting the AI to rebuild everything from scratch.

Non-Solutions

As a result, most teams end up in one of these frustrating workflows:

The Copy-Paste Dance: Generate diagram → screenshot → paste into slides → manually edit slides → repeat when changes needed
The Rebuild Cycle: Ask AI to remake the entire diagram every time you need a small change
The Tool-Switching Marathon: Use AI for ideas → recreate manually in Lucidchart or Powerpoint → lose the AI context → start over

The Solution: Editable AI Diagrams with Nimbalyst

What if your entire team could collaborate on AI-generated diagrams in real-time, iterating together with AI until you get exactly what you need?

What if you could work on text iteratively with AI and text and leverage that into AI generated diagrams and keep your diagrams and text in synch?

That's exactly what Nimbalyst's editable AI diagrams deliver.

Here's How It Works with Nimbalyst

Build your context for the diagram with AI and your team

In this example, we are looking to understand the steps a feature team goes through to build a product feature. We begin by typing some bullets.

We then use AI to complete the rest editing and approving what AI added, iterating in the text.

Use AI to create the initial diagram

We then use Nimbalyst to create an initial diagram inline, right in the same notebook, based on the full context.

AI creates a solid starting point, but unlike other tools, this is just the beginning.

Edit the diagram directly, collaborate with your team

Click into any element and edit it directly. Move boxes, change text, adjust connections—all while your teammates see changes in real-time. No more screenshots, no more rebuilding, no more tool-switching.

Keep iterating with AI

Need to add a new process step? Ask the AI to modify the diagram while preserving your team's edits. The AI understands your existing work and builds on it rather than replacing it.

Connect Everything

Since diagrams are first-class objects in Nimbalyst, you can reference them in your text, pull insights into tables, ask the AI questions about the diagram itself, and update the text based on the diagram. Everything stays connected and contextual.

Why This Matters for Teams

For Feature Teams: Map out your architecture diagrams and user flows collaboratively, then iterate as you learn. The diagram grows with your understanding.

For Strategy Teams: Create decision trees and process flows that the whole team can edit and approve tied to the text and table aspects of your strategy.

For Anyone Building with AI: Stop accepting AI diagram outputs as final. Start crafting it into exactly what you need. Base your diagrams on your full context.

Harness the full power of AI by iterating on editable diagrams that are tied to your context

With Nimbalyst's editable AI diagrams, you get:

AI-assisted creation that gives you a strong starting point
Real-time collaboration so your whole team can contribute
Iterative refinement without losing previous work
Integrated workflow that connects diagrams to your other work and context.

Stop generating diagrams you can't edit. Stop rebuilding from scratch. Start crafting visual knowledge that grows with your team.

Try Nimbalyst today!

How AI Coding Is Reshaping Feature Teams

Karl Wirth — Wed, 05 Aug 2026 13:00:00 +0000

AI-powered coding tools like Cursor, GitHub Copilot, Crystal, and Claude Code are shortening the development cycle which makes planning and process delays more salient as the ratio of planning to development increases. Feature Team work needs to adapt.

AI changes the Planning to Development Ratio

More rapid development breaks the waterfall process

In traditional waterfall, before AI, you would spend 2 to 10x more time on development that than in planning. This made it feasible to have long, process-oriented planning cycles.

This ratio changes with AI coding where for every hour of planning you might only spend a single hour on development moving the process bottleneck from engineering to planning. To address this, Andrew Ng recently proposed a 1 PM to 1 Engineer or 1 to 0.5 ratio.

More rapid development has implications for Agile as well

The point of Agile is to break the problem apart in order to:

Ship more often
Iterate faster
Get to customer feedback more quickly

Before AI, Agile development looked like this:

As development cycles are shortened by AI, agile development looks more like this:

Again, planning becomes the bottleneck. This shift is one reason AI coding increases friction for feature teams that haven't adapted their process.

I've been interviewing many engineering and PM leaders about this dynamic. Some have said their AI-first teams are following standard process and are frustrated. Others have thrown process to the wind and are following a process like this:

Neither is right. To work effectively with AI as a feature team, there needs to be a lighter-weight, faster process.

AI enables PM to mock up and prototype

AI coding platforms are enabling PMs and UX Designers to prototype full experiences more quickly. Tools like Lovable, Bolt, and v0 let PMs get from idea to customer feedback as quickly as possible.

A PM can help initiate an AI-speed appropriate process with two things:

The Why and What of a feature (see our guide on how to plan features with AI agents)
A robust prototype of that feature that has been iterated on based on customer feedback

Its not clear that these PM mockups / prototypes are helpful as code for the developers. So far, for Nimbalyst (formerly Stravu), they have not been. But, they are helpful as input into the Feature definition.

Our process for a feature team iterating rapidly with AI: a unified plan

A more lightweight, but still coordinated/managed process for a feature team working, prototyping, and coding with AI looks like this. This is how we are trying to do it at Nimbalyst:

At the center of this process is a unified Plan that contains the requirements, prototypes, architecture, technical approach, and to dos for the Feature. These words are paired with the code that is being developed. The plan and the code are kept updated and in synch as the team iterates rapidly. This approach enables coordination, rapid iteration, change based on everyone's feedback and learning, and everything staying in synch

Mitigating the downside of AI-powered development speed: bad code

In general, AI code increases speed but decreases quality. Some organizations are willing to take this tradeoff. Some are not. Many spend some of their saved developer time fixing the issues in AI's code. The following 4 inputs help to improve the quality of AI Code's output but do not mitigate it entirely:

A detailed description of the feature: This is the plan document described above. This needs to be as accurate (and kept accurate) and specific as possible.
A technical implementation: Also in the plan, with architecture decisions and diagrams
Context about your product and company: This includes the code-base, the company's vision, differentiators, other product features.
Instructions about how to code well at your company: This includes best practices, prompts and learnings from previous attempts to code with AI

Learning together

We are all figuring this out as it happens. What changes has your team made to adapt to AI-powered development?

Nimbalyst provides a way for feature teams to work together on unified, collaborative plans as they iterate and develop with AI.

If you are interested in exploring this with us, download Nimbalyst today!

Why AI Coding Increases Friction (and the Fix)

Karl Wirth — Tue, 04 Aug 2026 21:00:00 +0000

If you're leading a product or engineering team using AI coding tools, you've probably noticed something paradoxical: while individual developers are more productive, team coordination is increasingly slowing you down ... the friction increasing. Why?

The viscous friction problem for your feature team

Think about fluid dynamics for a moment. When you move a spoon slowly through honey, there's resistance but it's manageable. But try to move quickly through that same honey, and the resistance becomes overwhelming— this is viscous friction (which is directly, inversely a function of the velocity of movement).

The same thing is happening to your product development process. AI coding is increasing your development velocity, but your developers and whole feature team is moving through the same environment (the honey) and so the friction has increased. Your coordination processes and tools are still built for the old, slower world.

Let's consider three sources of friction and context loss:

Handoffs between people: Traditional feature development, even in "agile" environments, involves many handoff points between Product Management, UX, and Engineering. Each handoff happens between different people and different systems, creating opportunities for context loss and delays. For example:

Product defines big picture one pager feature in a Google Doc → UX researches flows → PM aligns on features in Aha → UX creates wireframes in Figma → PM reviews design via a meeting and Slack → PM writes detailed requirements in Confluence → Engineering provides feedback on requirements and designs in Confluence → UX finalizes specs in Figma → Engineering plans architecture in separate Confluence → Engineering defines task and estimates in JIRA → Engineering builds prototype and code lives in Github → UX reviews internally → Engineering builds final version → QA testing → PM prepares launch docs in Google Docs → Deploy

Tool Fragmentation: Your context is scattered across many different tools. One pagers in Google Docs, features in Aha, requirements in Confluence, designs in Figma, tickets in Linear or Jira, code in GitHub, and developer notes in random text files that disappear when the feature ships. The more tools the more friction and opportunity for context loss.

Manual Syncs: Someone, usually a project manager or scrum master or engineering leaders, spends their days trying to keep everyone aligned on what's being built, what's done, and what's next. A human developer might spend time tracking down the product manager to ask clarifying questions.

This is the viscous honey that you are moving through and it is what is resulting in friction and context loss.

How AI makes viscous friction worse and context loss more of an issue

Communication Becomes the Bottleneck

Great product development is highly collaborative and iterative. But iteration requires constant communication about trade-offs, priorities, and changes. When development cycles shrink from weeks to days, this communication overhead becomes crushing. This is why AI coding and prototyping are reshaping feature team work at every level.

Teams find themselves spending more time in status meetings than building features. Project managers become air traffic controllers, frantically trying to coordinate work that's happening faster than they can track.

AI won't do the work a human will to get the right context

AI-coding and prototyping makes it all the more important for the full context to be available and up-to-date for the developer. A human may be willing and able to spend the time finding out what is accurate or not and asking around for updates and context. An AI will not.

Using Claude Code to MCP to JIRA and Confluence is powerful but only a partial solution as you are leaving too much interpretation up to the AI.

Documentation Debt Explodes

AI coding encourages experimentation. Developers can quickly try four different approaches, keep the one that works best, and throw away the other three. But where do you track those experiments? When context is the new code, how do you capture what you learned? How do you prevent the next person from repeating the same failed approaches?

Traditional project management tools weren't built for this kind of rapid iteration and experimentation.

A New Way of Working

The solution isn't to slow down development. We need to collaborate differently.

Shared, accurate context is king

This means:

Developers and AI that understand full project context, not just individual tasks
Requirements, plans, and todos available where the actual work happens
AI tools and human team members accessing the same context simultaneously
Updates flowing between specs and implementation
Collective learning from AI experiments and iterations

Lightweight, continuous processes are needed

This means moving away from formal handoffs, ticket management, and heavyweight documentation toward:

More frequent, lightweight check-ins with working prototypes
Living documents that evolve with the code rather than getting outdated
Iterative prototypes, experiments, and finished code

What This Looks Like in Practice

For my own team at Nimbalyst, instead of separate PRDs, design specs, and technical documentation, we work from unified planning documents for each feature. (Of course we use Nimbalyst on Nimbalyst to do so!)

These documents are leveraged simultaneously by Claude (our AI coding assistant), our developers, and me as PM. When someone discovers that a requirement needs to change, or when we learn something new during development, it gets updated in one place and flows to everyone who needs it. We are integrating this with some of the standard project management tools.

The result? We're building better features faster, with less coordination overhead and fewer miscommunications.

It is a lot more fun to swim in water than to try to swim in honey.

Learning together

AI coding isn't going to slow down. If anything, it's going to get faster and more capable. The teams that thrive will be those that redesign their collaboration practices around this new reality.

We are all figuring this out as it happens. What coordination challenges are you facing with AI coding? What changes has your team made to adapt to AI-powered development?

Nimbalyst provides a way for feature teams to work together on plans as they iterate and develop with AI.

If you are interested in exploring this with us, download Nimbalyst today!

How I review and approve AI output

Karl Wirth — Tue, 04 Aug 2026 19:00:00 +0000

I use AI constantly as we build Nimbalyst. It is an invaluable assistant. But, its output needs editing, review, and approval. Here is how I work with AI and edit/review/approve its output.

The Dangers of Unreviewed AI Output

Hallucinations are common. AI systems confidently generate information that sounds authoritative but is completely fabricated.
Inaccuracies creep in constantly. Even when AI isn't hallucinating outright, it frequently gets things wrong. Yesterday, it matched the
AI takes your content in unexpected directions. I had AI use a detailed outline for this blog post (I will include it below) to write a first version for me. It talked about all kinds of AI for all use cases, whereas I wanted the blog focused on working with AI on business content. I rejected this blog post, gave more context, and tried again.
AI's length is off. The first blog AI wrote was a book. I like things shorter.
Important! You and your team need to understand what you are writing and what is happening.Seriously.

Best Individual AI Workflow

The most effective individual workflow treats AI as a sophisticated drafting partner, not a final authority:

Start with your own thinking. Before engaging AI, write down you goals, some key points, and open questions. Here is the initial goal I wrote down.
Or sometimes get kickstarted with AI: Often AI brainstorming can help get me started as it did for this blog post, but I usually then throw away everything it wrote and move to the next step.
Build your detailed outline. Build a full outline or bullet points or a prompt for what you want. Be clear about your goals. The more specific guidance you provide, the more useful the output will be. The example outline I used for this blog post is in the appendix.
Have AI generate a first response.
Accept what is good. Reject what is bad. Edit ruthlessly.
Iterate until it's right. Use multiple rounds of AI generation and human editing to create content that truly meets your standards.

Scaling Review for Team Success

Team workflows require more coordination but produce better results:

Harness diverse perspectives early. Have each team member brainstorm independently—treat AI as another team member contributing ideas, not the primary source. This prevents groupthink and ensures multiple viewpoints inform your content.
Centralize and evaluate all inputs. Bring all brainstorms together in one shared space. Evaluate ideas collectively, identifying the strongest concepts and approaches.
Build consensus on structure. Create a detailed outline that the whole team agrees on before any AI generation begins. This prevents scope creep and ensures everyone works toward the same goals.
Use AI for expansion, not direction. Let AI flesh out your agreed-upon outline, but maintain human control over strategy and key messaging.
Enable collaborative editing. Team members should add their expertise, refine AI-generated sections, and ensure the content reflects collective knowledge.
Implement review processes. Have team members review each other's work, then use AI to identify potential issues or improvements you might have missed.
Iterate as a team. Use feedback loops where human edits and AI suggestions inform each other, gradually improving the content through multiple rounds.

Practical Tools for Effective Review

Red/Green Diff tracking with Approve/Reject makes it easy to see exactly what changed between versions, especially what edits AI made. This applies to text, tables, and diagrams alike. This visual comparison helps teams understand the evolution of content and ensures important changes don't get lost.
Undo functionality provides safety nets for experimentation. Teams can try AI suggestions knowing they can quickly revert if the changes don't work.
History with Revert. If things get too messed up, revert to an earlier version.

These tools transform iterative collaboration with AI from a burden into an efficient process that improves both quality and team alignment.

The Bottom Line

AI is a powerful partner, but it's not a replacement for human judgment. Teams that build robust review and approval processes into their AI workflows produce better results, avoid costly mistakes, and maintain the collaborative culture that drives innovation.

The question is how to build review processes that are thorough enough to catch problems but efficient enough to maintain momentum. For AI coding workflows specifically, tracking what your AI agent changed is equally important. Get this balance right, and AI becomes a true force multiplier for your team's capabilities.

Appendix: Outline I gave AI

Why need to review and approve AI output

Hallucinations
Inaccuracies
Took it in direction you didn't want
Too much detail or too little
Understand what is being said
Whole team understand what has changed

Best individual workflow

(sometimes: Use AI to brainstorm and get you started. Then throw away most of this)
Write some hooks, context, bullets, an outline
Have AI generate based on that
Then review and edit that
Iterate

Best team workflow

Each team member brainstorm independently. AI can be another team member
Bring brainstorms together in one place and evaluate
Build an outline
Have AI flesh out outline
Team members edit and add
Team members review and AI reviews
Iterate

How to review and edit AI's output

Red/Green Diff
Undo
History

Managing AI Is Like Managing an Intern

Karl Wirth — Tue, 04 Aug 2026 17:00:00 +0000

I find it helpful to think of an AI agent like an enthusiastic intern. On the plus side, this AI intern is incredibly fast, has access to virtually all the world's information, will work 24x7, and is eager to please. This is why I am working with it. On the negative side, this AI intern needs guidance and frequent check-ins/feedback, sometimes fabricates information when it doesn't know the answer, and makes rookie mistakes.

If my smart, well-educated, motivated intern isn't producing great results, before firing them, I need to ask myself how I can improve my management skills. We've all had the boss whose approach to management is: "Guess what I am thinking and then perform magic". That is too often how we approach AI. In reality, the better work a good manager does, the better work his/her employees can do.

The analogy is just an analogy—AI isn't a person with feelings or career aspirations. But the management skills that work with interns translate remarkably well to managing AI.

Learn to manage earlier in your career

AI is accelerating the need for management skills earlier in careers. The idea of turning a rock into a statue captures this well -- the sculpting skill is what matters, not the raw material. Traditionally, professionals worked as individual contributors through their twenties, then transition to managing teams in their thirties. You structure and do your own work, then eventually learn to structure others' work.

But AI changes this timeline. College students and early-career professionals now need to learn management skills to effectively leverage AI agents. The ability to structure work, provide clear direction, and manage collaboration isn't just a senior-level skill anymore—it's essential for anyone who wants to multiply their impact with AI.

How to manage AI (and interns) well

Set Clear Expectations

Be specific about:

The goal of the project and how it connects to broader goals
What deliverables you expect and in what format
What success looks like for the specific task
Which sources or methods to prioritize

Poor expectation setting leads to generic outputs that require extensive revision. Clear expectations produce targeted results that accelerate your work. For practical tips on this, see our best practices for coding with agents.

Provide Structure

Break complex projects into manageable tasks with clear deadlines. How to do this well is a skill that you can practice.

Example Bad Structure: "Research our competitors and tell me what we should do about them."

Example Good Structure: "We're planning our Q1 product roadmap and need competitive intelligence on three specific rivals: [Company A], [Company B], and [Company C].

Phase 1 (by Friday): Create profiles for each competitor including their latest product features, pricing models, and recent funding/news.
Phase 2 (by next Tuesday): Analyze their go-to-market strategies - how are they positioning themselves, what channels are they using, who are their target customers?
Phase 3 (by Thursday): Identify 2-3 specific opportunities where we can differentiate or areas where we're falling behind that need immediate attention.
Final deliverable: A 2-page summary with actionable recommendations for our product and marketing teams, plus detailed research notes organized by competitor."

With AI today, you don't specify the timeframes.

Give (the right amount of) Context

You have a lot of context in your head and from your experience. The more background, relevant other projects and context you can share, the better job your intern can do.

But, you also don't want to overwhelm or distract. The intern can get distracted, not see the trees for the forest, and not know what is important or not. So, its on you as the manager to decide the right amount of context for them.

Check In Regularly

Establish a rhythm of iterative collaboration. Don't just dump a complex request and expect perfect results. Instead:

Review preliminary outputs and provide course corrections
Ask follow-up questions to dig deeper into promising areas
Clarify when the AI seems confused or off-track
Build on insights as they emerge

This iterative approach prevents you from wasting time on work that's heading in the wrong direction.

Check Work and Offer Feedback

Always verify AI outputs, especially for important decisions. Develop habits around:

Cross-checking factual claims with reliable sources
Testing logic and reasoning for soundness
Ensuring recommendations align with your business context
Refining communication for your audience

Treat initial AI outputs as first drafts that benefit from human refinement, not finished products.

Track Progress, Reflect, and Build

Document what works and what doesn't in your AI collaboration. Keep a simple record of:

Which prompts produce the best results for different types of tasks
Common areas where AI needs human oversight
Successful collaboration patterns you can replicate

For insights that will be useful to the AI on the next project, keep a list in a shared notebook that you and AI both reference

A simple system for managing AI and interns

With interns or employees, I tended to have two documents:

The detailed plan document for whatever project the intern is working on (this should be collaborative and reviewed by both of us)
A running note document that has the history of our interactions, learning, tasks, and a catalog of the detailed plan documents

This allows us to go deep on the project and have a meta-level running conversation about what we are doing and learning.

I am finding the same useful when managing and working with AI.

Management is a skill that takes work

Often, when I type something into an AI interface, I am tempted to be a bad boss. I want AI to read my mind and do magic work. Why? Because it takes work to manage well. It takes thought and effort. It is a skill that I need to hone. A good manager is skilled at the following and in every project with their team they do the following work:

Setting clear expectations and providing context
Breaking complex projects into manageable components
Giving constructive feedback and course corrections
Creating systems that improve performance over time

This is one of the ways in which I find that AI is making me smarter and better, not lazier and dumber. I am constantly honing my management skills as I work with AI. I am thinking harder about problems up front. I am structuring them better.

Management is how to grow your career and succeed

The professionals who thrive in an AI-augmented workplace will be those who learn to be great managers, whether they're managing people, AI agents, or both. The intern analogy gives us a practical framework to start building those skills today.

Your next AI interaction is an opportunity to practice management. And as you scale from one agent to many, you'll find that managing 6 agents at once requires the same skills, just at a faster pace. Approach it with the same intentionality you'd bring to developing a promising intern, and watch both your results and your leadership capabilities grow.

Claude Code: Superpowers for Product Managers

Karl Wirth — Tue, 04 Aug 2026 15:00:00 +0000

As a Product Manager, you've likely used ChatGPT or Claude to assist with technical concepts, research, user story drafting, or product idea exploration. Claude Code is a powerful, local, AI agent that enables you to leverage your actual codebase, build prototypes, make and run commands, and work with local files.

What Makes Claude Code Different:


Feature	ChatGPT & Claude (Web Chat)	Claude Code (Local Agent)
Where it runs	Browser window, cloud-based	Your computer with local access
Agent	Limited	Agent and sub-agents
Codebase access	No	Yes
File access	No	Sees and edits files and folders
Commands	Limited	Custom commands and skills (that you write with Claude in a markdown document)
Write and run scripts	No	Writes and executes scripts on your machine

Getting Local Access to Your Code: Read-Only Git Access is Enough

To fully utilize Claude Code's capabilities with your code, you need a local copy of your product code. This is typically achieved by cloning your product's code repository to your local machine using tools like GitHub Desktop. Read-only access to the repository is sufficient, allowing you to explore, ask questions, and learn without the risk of altering production code.

PM Use Cases That Claude Code + Git Access Can Unlock

Imagine having a personal junior developer to investigate the codebase and assist with your work. Claude Code enables powerful PM workflows:

Feature Planning: Bring Technical Context to your Product Plans and PRDs

Every PMs feature description or PRD gets better after a few iterations with an architect or the engineering team. They ask clarifying questions, help you think about interactions with current features, etc. Use Claude Code to advance your plans and PRDs based on its technical knowledge of your code base.

Custom Commands: /plan, /create-prd, /prd-review

Understand What Your Team Actually Built (Without Interrupting Engineers)

Investigate how features work without interrupting your engineering team.

Examples:

"How does our user authentication flow work?"
"Where is customer payment data stored?"
"What API endpoints do we expose?"
“What is the schema for this”

Custom Commands: /explain-feature

Understand Status of Work

Why wait for a weekly status meeting to understand your team’s progress. Have Claude Code look at Github and summarize what it sees.

Custom Commands: /status-of-feature

Investigate User-Reported Issues Before Filing Bugs

Determine if an issue is a bug, UX problem, or user error before escalating.

A customer reports confusing behavior in the checkout flow. Instead of scheduling an engineering sync, you use Claude Code to trace through the checkout code, understand the logic, and come to your meeting with specific questions about the decision points you discovered.

Custom Commands: /triage-bug

Triage a Feature Request

PMs are overwhelmed by the number of feature requests. Claude Code can triage the feature request, do the background work, summarize the feature, make sure it is not a duplicate, categorize its tags, etc…

Custom Commands: /triage-request

Document Features for Customers, Sales, Marketing

PMs need to explain features for customers, sales, and marketing. Work with Claude Code to document new features automatically as they are built.

Custom Commands: /document-for-customers /document-for-gtm

Rapid Prototyping & MVP Development

Create functional, interactive prototypes of a feature without engineering resources. Transform wireframes, mockups, or hand-drawn sketches into working applications with real business logic. PMs can validate product ideas by creating functional prototypes that users can click through and interact with, moving from abstract concepts to concrete experiments.

PM Use Cases Using Claude as a Research Agent

Competitive Analysis & Market Research

Conduct comprehensive competitive analysis using Claude as an Agent. Generate SWOT analyses, comparison matrices, and market landscape reports.

PM Use Cases That Claude Code Local Files + Scripting Can Unlock

Product Metrics Analysis

There is gold in your usage data, but sometimes its hard to find it. Use Claude Code to generate SQL queries, create dashboards, analyze product metrics, and build data visualizations without requiring coding expertise.

Custom Commands: /update-metrics-weekly, /update-metrics-daily

Customer Feedback Analysis & Research Synthesis

Process and analyze thousands of customer feedback responses, user interviews, and survey data. Identify patterns, pain points, and actionable insights at scale.

Write scripts to analyze public data sources like Reddit comments, identify themes, categorize.

Custom Commands: /analyze-reddit

What Product Managers Don't Need to Know

Claude Code provides code access, but you are not now suddenly an engineer. Don’t expert to:

Write production code yourself
Debug complex technical issues
Review code for correctness or quality
Make architectural decisions

What you will do is:

Read code to understand how things work
Ask questions about your system's capabilities
Investigate issues before escalating
Make more informed product decisions

Think of it like a PM learning to read financial statements. You're not becoming a CFO, but understanding the numbers helps you make better business decisions.

Augmenting Claude Code with Nimbalyst

On your Desktop, Claude Code runs in the command line. It is powerful but its use is a bit technical. Nimbalyst augments your use of Claude Code by providing you with:

A UI to manage Claude Code
An editor to edit markdown files (the documents Claude uses for direction and for output)
A way to keep track of your Claude sessions, run multiple in parallel

Conclusion: Why PMs Need Direct Codebase Access

As a Product Manager, you're expected to make informed decisions about a technical product, but you're often working blind, relying entirely on engineers to explain how things work.

Claude Code changes that dynamic. It doesn't make you an engineer, but it gives you the ability to investigate, understand, and validate technical details yourself. This means:

Fewer interruptions for your engineering team
Faster answers to your questions about the product
Better-informed prioritization and planning decisions
More effective communication with technical stakeholders
Deeper understanding of technical constraints and opportunities

Ready to understand your product at a deeper level? Install Claude Code, point it at your codebase, and ask it to explain how your most important feature works. For more on the best AI tools for product managers, see our full guide. You'll immediately see why Product Managers who try it rarely go back to relying solely on engineers for technical context.

How PMs Track Dev Progress with Claude Code

Karl Wirth — Tue, 04 Aug 2026 13:00:00 +0000

Most Product Managers struggle to get quick, accurate answers about development progress. Asking engineers "What's the status?" interrupts their flow. Checking project management tools shows planned work, not actual code changes. There's a better way: understanding git status through AI assistance.

This guide shows you how to use git and Claude Code to track development progress in real-time, making you a more effective PM without requiring deep technical knowledge. If you want to go deeper into understanding how features are built, see how PMs can master feature understanding with Claude Code and Git. We'll explore general techniques that work anywhere, then show how Nimbalyst streamlines this workflow into a seamless experience.

Why Git Status Matters for Product Managers

Git is where the actual work happens. While your project management tools track intentions and estimates, git repositories contain the truth: what code has actually been written, modified, and merged.

Understanding git status gives you:

Real-time visibility into what's being worked on right now
Evidence of progress between standups or status meetings
Early warning signs when work diverges from plans
Context for better technical conversations with your team

The challenge? Git commands and output weren't designed for non-developers. That's where AI assistance transforms everything.

Using Claude Code to Decode Git Status

Claude Code acts as your technical translator, turning cryptic git information into clear product insights. Here's how to leverage it effectively.

Getting a Plain-English Status Update

Instead of running git commands yourself, ask Claude Code: "What development work is currently in progress?" Claude Code will examine the git repository and tell you:

Which branches exist and what they're for
What files have been modified but not committed
What changes are staged and ready to commit
Recent commits and their purposes

This gives you an instant snapshot without interrupting your developers.

Understanding What Changed Since Last Sprint

Ask Claude Code: "What code changes have been made since last Friday?"

You'll get a narrative summary of commits, not a wall of technical diff output. Claude Code can explain:

Which features had code written for them
What bugs were addressed
Whether changes align with sprint commitments
If unexpected technical work emerged

Tracking Progress on Specific Features

When you need to know about a particular feature: "Show me the git history related to user authentication work."

Claude Code will search commit messages, changed files, and code content to surface relevant development activity. This helps you:

Validate that priority work is actually happening
Understand if a feature is 90% done or 10% done
See if additional scope was discovered during implementation
Identify potential integration points with other work

Better Standup Prep in Seconds

Before your daily standup, ask Claude Code:"What development progress happened yesterday across all active plans?"

You'll get a consolidated update showing:

Which plans had code commits
What changed in each area
Potential blockers or divergences from plans
Developers who contributed to each plan area

You enter the standup informed and can ask better questions.

Streamlined Sprint Reviews

When preparing for sprint reviews or demos, Nimbalyst helps you quickly assess completeness: "Compare git changes to the acceptance criteria for the checkout-flow plan."

Claude Code will:

Review what code was actually written
Cross-reference against your documented acceptance criteria
Identify gaps between completed work and original requirements
Suggest areas to focus testing or discussion

This makes sprint reviews more productive and keeps everyone honest about what "done" means.

Pre-Meeting Prep

Before any meeting where you'll discuss technical progress: "Summarize git activity relevant to [specific feature or area]."

Walking into meetings informed changes the conversation from status reporting to problem-solving.

Retrospective Input

During sprint retrospectives, git history provides objective data: "What does our commit history reveal about how this sprint actually went?"

This grounds retrospective discussions in evidence rather than subjective memory.

Streamlining with Custom Commands

One powerful way to make git status checking fast and consistent is creating custom Claude Code commands. These are reusable shortcuts that execute specific workflows with a simple slash command.

Creating a Status Command

You can create a custom command like /dev-status that automatically:

Checks current branch and recent commits
Lists all modified and staged files
Identifies any unmerged branches
Summarizes recent development activity
Highlights potential issues or blockers

To set this up, create a markdown file in your .claude/commands/ directory.

Example Claude Code Custom Command for Git Status

For example, create dev-status.md:

# Development Status Check

Provide a comprehensive development status update:

1. Current branch and recent commits (last 5)
2. All modified, staged, and untracked files
3. List of all branches and their last commit dates
4. Any long-running branches that haven't been merged
5. Summary of development activity in the last 24 hours
6. Potential risks or blockers based on git patterns

Format the response in a clear, scannable way suitable for a Product Manager.
Place the response in the DevStatus folder in this directory

Now typing /dev-status instantly triggers a comprehensive analysis without you needing to remember what to ask.

Feature-Specific Commands

Create targeted commands for specific areas of your product:

/auth-progress.md: Show all git activity related to authentication features in the last week, including commits, branches, and file changes.

/sprint-summary.md: Analyze all commits since the start of the current sprint and organize them by feature area with progress assessments.

These custom commands become your personal PM toolkit, giving you instant access to exactly the information you need most often.

Making It Even Easier: Nimbalyst and Claude Code

While Claude Code makes git accessible to Product Managers, Nimbalyst takes it further by providing a UI for working with Claude Code, a markdown editor to review its output and edit it iteratively with Claude Code, and a link between your planning environment and your git via Claude Code.

UI for Managing Claude Code and Edit Markdown

Nimbalyst provides a UI for Managing Claude Code where you can use Slash commands, see the output Claude Code comes back with, edit it, and iterate on it with Claude Code.

Unified Planning and Development Visibility

Nimbalyst combines your markdown-based plans with live git status. When you open a plan document for a feature, you can immediately ask:

"What's the development status for this plan?" and Nimbalyst's integrated Claude Code:

Checks git for branches and commits related to the plan
Matches code changes against your acceptance criteria
Updates you on progress without leaving your planning document
Highlights discrepancies between plans and actual implementation

Your planning documents and development reality stay connected.

Seamless Status Updates Without Context Switching

As a PM using Nimbalyst, you work in markdown files that contain your plans, specs, and acceptance criteria. You don't need to switch to a terminal or separate git tool. Simply ask questions in context: "Has the user authentication plan been started?"

Nimbalyst checks git status and tells you:

Whether a feature branch exists
What files have been modified
Recent commit activity
How the work aligns with your documented plan

Intelligent Progress Tracking

Nimbalyst understands the structure of your plans because they're written in markdown with YAML frontmatter. It can:

Automatically suggest status updates based on git activity
Warn you when code is being written for plans marked as "draft"
Identify when completed code exists but plans are still "in-development"
Help you keep plan status synchronized with actual progress

This closed-loop system ensures your plans reflect reality.

Building a Git-Aware PM Workflow

Integrating git status into your regular product management routine doesn't require becoming a developer. Here's a practical approach.

Daily Practice

Start each day with a quick status check. In Claude Code or Nimbalyst, ask: "What code changes happened since I last checked yesterday?"

This five-minute habit keeps you continuously informed without scheduled status meetings.

Weekly Deep Dive

Once per week, do a more thorough analysis: "Show me all development activity from the past week organized by feature area."

Use this to:

Validate that sprint commitments are being addressed
Identify work that wasn't planned but emerged
Spot patterns in where time is actually being spent
Prepare for planning meetings with concrete data

Conclusion

Product Managers don't need to become git experts, but understanding development status through git gives you superpowers: real-time visibility, objective progress data, and the ability to spot risks early.

Claude Code makes git accessible by translating technical information into product insights. You can track progress, understand velocity, and make better decisions without interrupting your engineering team. For a broader look at what Claude Code unlocks for PMs, see Claude Code superpowers for product managers.

Nimbalyst takes this further by integrating AI-assisted git analysis directly into your markdown-based planning workflow. Your plans and actual development stay connected, status updates happen in context, and you maintain a clear picture of progress without leaving your planning documents.

The result? You're a more effective Product Manager who makes decisions based on what's actually happening in the codebase, not just what's reflected in project management tools.

Ready to transform how you track development progress? Start by asking Claude Code about your git status today. If you want the seamless, integrated experience, explore how Nimbalyst brings planning and development visibility together in one place.