The expensive mistake nobody budgets for
You already review your AI agent's output. You catch the wrong tone, the missing section, the fact it invented a customer who doesn't exist. But by then the tokens are spent, the time is gone, and the rework costs real money.
What almost nobody reviews is the plan — the step-by-step assumptions the agent makes before it starts working. Wrong assumptions cost nothing to fix at the plan stage. They cost plenty once the agent has already acted on them.
A free, open-source tool called Plannotator fixes this by turning the agent's plan into a document you can mark up in your browser before the agent runs a single step.
What the checkpoint actually catches
Plannotator hooks into the plan-approval step in Claude Code (and integrates with Codex, OpenCode, and Pi). When your agent proposes a plan, the tool opens it in a browser workspace instead of leaving it in the terminal. You select text, mark deletions, add comments, or rewrite sections. Your annotations flow back to the agent as structured feedback. The agent revises the plan. You approve. Only then does execution start.
The tool is local by default — no telemetry, no analytics, no cloud dependency. It runs on your machine, reads your agent's plan file, and sends your feedback back through the same session.
But the point isn't the tool. The point is the checkpoint.
The SMB version: three questions before any expensive AI job
You don't need Claude Code or a coding agent to use this discipline. Any multi-step AI task — a report, a proposal, a customer journey, a content batch — benefits from a two-minute plan review before execution.
Write this prompt before you hand off anything that costs more than a few cents to run:
"Show me your plan first: what you will do, in what order, and what you are assuming about the inputs. I will review and approve before you execute."
Then ask three questions:
What could go wrong? If the agent assumes your CRM data is clean and it isn't, the whole output is garbage. If it assumes a file format that changed last month, you'll get an error at step seven with six steps already billed.
What am I assuming? Your own unstated assumptions are the hidden risk. "Sort customers by value" assumes you and the agent agree on what "value" means — total revenue, profit margin, lifetime value, or something else. Spell it out.
What's missing? Agents tend to scope to what you asked, not what you needed. A plan to "write follow-up emails" may omit the step that checks whether the customer already replied. A plan to "analyze Q3 sales" may miss the returns data that changes the answer.
The two-minute discipline
The actual review takes about two minutes. The agent writes the plan in its own words. You read for wrong assumptions while they are still free to fix. One comment — "check for returns before summing" — prevents a full rerun.
This is especially important for jobs that chain multiple steps. Each step builds on the previous one. An error in step one compounds through steps two, three, and four. Catching it at the plan stage is the difference between a two-minute comment and a multi-dollar rebuild.
What this is not
This is not a replacement for output review. You still need to check what actually shipped. The plan checkpoint catches wrong assumptions — it does not catch correct assumptions applied to bad data, unexpected edge cases, or things you didn't think to ask about.
It is also not a guarantee. A plan can look right and still fail in execution. The checkpoint raises your hit rate; it does not eliminate misses.
How to start today (with or without the tool)
With Plannotator: Install via the project's installer, add the Claude Code plugin, and enter plan mode normally. The tool intercepts ExitPlanMode automatically. Review, annotate, send feedback, approve. The project's docs at plannotator.ai cover Claude Code, Codex, and VS Code integration. The repo is open-source under backnotprop/plannotator on GitHub.
Without the tool: Add one line to your prompt template: "List your plan before executing. Wait for my approval." Read the plan for the three questions above. Reply with corrections. This costs nothing and works in any chat interface.
The habit matters more than the tooling. Agents are fast. The temptation is to skip the review and let them run. The two-minute pause is the cheapest insurance you can buy.
About the AI Clarity Sprint
If your business is using AI but you're not sure what's actually running, what's shadow usage, or what to automate first, the AI Clarity Sprint is a two-week fixed-fee engagement ($1,495 CAD) that delivers a complete inventory of your AI touchpoints, an own-vs-use analysis, three ranked quick wins, and a first-project plan. It ends with a plan you own — not a platform you rent.
Top comments (0)