DEV Community: Shahid Saleem

How to Use Claude to Extract Key Insights From a Dense PDF Report in Minutes

Shahid Saleem — Wed, 13 May 2026 20:34:16 +0000

Originally published at PickGearLab — practical AI tutorials for writers, freelancers, and small business owners.

Most people open a long PDF and start reading from page one. By page twelve they’ve lost the thread. By page twenty they’re skimming. By page forty they’re looking for a summary that doesn’t exist.

There’s a faster way. Claude — Anthropic’s AI assistant — can read a dense PDF alongside you and pull out exactly what you need in minutes, not hours. This is how I handle client reports, research papers, and technical documentation that I don’t have time to read cover to cover.

What Claude Can Actually Do With a PDF

Claude accepts PDF uploads directly in its interface (available on Claude.ai with a paid plan, or via the API). Once uploaded, you can ask it anything about the document:

Summarise the entire document in plain language
Extract every action item or recommendation
Find specific data points: numbers, dates, names, decisions
Compare two sections that contradict each other
Rewrite the executive summary in plain English for a non-technical audience
List every risk or caveat mentioned in the document

What makes Claude particularly useful for dense reports is that it reads the whole document before answering — it doesn’t just match keywords. Ask it “what are the three biggest risks mentioned in this report?” and it will synthesise across sections, not just quote the paragraph that contains the word “risk.”

Step 1: Prepare the PDF

Before uploading, a quick check saves time later.

Make sure the PDF is text-based, not scanned. Claude can read text embedded in a PDF; it cannot read a scanned image of a document (a photo of printed pages saved as PDF). To check: try selecting text in the PDF with your cursor. If you can highlight and copy individual words, it’s text-based. If your cursor turns into a crosshair and selects rectangles, it’s a scanned image — you’ll need to run it through OCR first (Adobe Acrobat has a free web tool; so does Smallpdf).

For very long documents (over 100 pages), consider splitting into sections if Claude hits its context limit. A 50-page financial report, a 60-page technical specification, or an 80-page research paper all typically fall within Claude’s context window without splitting.

Step 2: Upload and Set Context

In Claude.ai, click the paperclip icon in the chat input and upload your PDF. Then write your first message — and this is where most people underuse the tool.

Don’t just ask “summarise this.” Give Claude context about what you need and why:

Your role: “I’m a project manager reviewing this vendor proposal”
Your goal: “I need to brief my director in 10 minutes”
What to emphasise: “Focus on pricing, delivery timelines, and any conditions or exclusions”

A prompt like “I’m reviewing this 45-page vendor proposal as a procurement manager. Summarise the key commercial terms, highlight any conditions or exclusions buried in the document, and flag anything that looks non-standard or risky.” will produce a dramatically more useful output than a one-word instruction.

Step 3: Use the Right Prompts for the Right Output

Different documents need different extraction strategies. Here are the prompts I use most often:

For a research paper or academic report:
“What is the main argument of this paper? What evidence does it use to support it? What are the limitations the authors acknowledge? What are the practical implications for someone working in [your field]?”

For a financial or business report:
“List every specific number, metric, or KPI mentioned. Identify any year-over-year changes. Flag any risks, caveats, or disclaimers. What decisions or recommendations does the report make?”

For a legal or compliance document:
“What are my obligations under this document? What are the deadlines? What happens if I don’t comply? Highlight any clauses that are unusually broad or that I should get a second opinion on.”

For a technical specification:
“List the system requirements. Identify any dependencies. What are the stated limitations? Are there any contradictions between sections?”

For meeting minutes or a transcript:
“List every action item, who it’s assigned to, and the deadline. List every decision made. List any open questions that weren’t resolved.”

Step 4: Follow Up to Go Deeper

Claude remembers the document for the entire conversation. After the initial extraction, you can drill down:

“Can you find the exact page and quote where it mentions the penalty clause?”
“You mentioned three risks — expand on the second one. What specifically does the document say about it?”
“Rewrite the executive summary you gave me as a five-bullet briefing I can paste into a Slack message.”
“Is there anything in section 4 that contradicts what section 2 says about the timeline?”

This back-and-forth is where Claude becomes genuinely more useful than a static summary tool. You can interrogate the document the way you’d interrogate a colleague who had read it thoroughly.

Real Example: A 50-Page Client Report

I recently used this on a 52-page market research report a client sent over. They needed a briefing for a board meeting in two hours.

First prompt: summarise the key findings and any strategic recommendations in plain language.
Second prompt: list every data point that supports or contradicts their current strategy.
Third prompt: rewrite the top three findings as a three-slide deck outline with one headline and two supporting bullets per slide.

Total time: 18 minutes. The briefing was more thorough than if I’d read the report myself under time pressure — because Claude doesn’t skim, doesn’t get tired, and doesn’t miss footnotes.

Limitations Worth Knowing

Claude is excellent at synthesis and explanation. It is less reliable for tasks that require absolute precision:

Exact quotes: Claude sometimes paraphrases slightly when asked to quote. If you need verbatim text for a contract or legal matter, always verify against the original.
Complex tables and charts: Data in image-based charts (bar graphs, pie charts) inside PDFs may not be read correctly. Text-based tables are fine.
Very long documents over 200 pages: You may hit context limits. Split the document into logical sections and process each separately.

For most professional use cases — reports, proposals, research, specifications — these limitations rarely matter. The value is in rapid synthesis and structured extraction, not in replacing the source document.

The Workflow in Summary

Upload the PDF. Tell Claude who you are and what you need. Ask for the extraction that matches your use case. Follow up to drill down. Use the output to brief, decide, or delegate.

A document that would take 90 minutes to read carefully takes 10 minutes to extract what you actually need. For anyone who regularly receives long reports, proposals, or research documents, that’s a real and repeatable time saving — not a one-off trick.

About the author

Shahid Saleem writes PickGearLab — a practical blog about AI tools, tutorials, and automation workflows for people who want real results, not another listicle. Certified in Microsoft AZ-900, CompTIA Security+, and AWS AI Practitioner, with 10+ years in enterprise IT.

→ Connect on LinkedIn · More about Shahid · Latest posts

Claude Pro vs ChatGPT Plus: A Decision Tree for Picking the Right $20

Shahid Saleem — Wed, 13 May 2026 20:33:59 +0000

Originally published at PickGearLab — practical AI tutorials for writers, freelancers, and small business owners.

Both cost $20 a month. Both are excellent. Both will tell you they’re the best. After eighteen months of paying for both at different points and twelve months of running them in parallel, here’s the decision framework I wish someone had handed me when I first hit the subscribe button.

This post is not a feature comparison. The feature gaps close every quarter, and a comparison written today will be outdated by August. Instead, this is a decision tree based on the kind of work you actually do. Walk through the questions. The branch you end up on is the subscription that fits your workflow.

TL;DR. If most of your work is short-form (emails, social posts, brainstorming, code snippets) — ChatGPT Plus. If most of your work is long-form thinking (writing, analysis, document review, research) — Claude Pro. If you do both seriously, the honest answer is to subscribe to both for one month, observe which tab you actually open, and cancel the other one. Almost no one needs both forever.

TL;DR

This post offers a decision tree to help users choose between Claude Pro and ChatGPT Plus based on their specific workflow needs, focusing on document length, writing style, feature requirements, coding, and hallucination tolerance.

Key takeaways

ChatGPT Plus excels at short-form tasks like emails, social posts, and quick code snippets.
Claude Pro is superior for long-form thinking, writing, analysis, and document review.
ChatGPT Plus uniquely offers voice mode and image generation capabilities.
Claude Pro provides cleaner, more architecturally sound code for complex problems.
Claude Pro is more cautious with facts, reducing hallucinations in technical or cited work.


ChatGPT Plus	Claude Pro
$20/month	$20/month
Short-form content, quick code, voice, image generation	Long-form writing, document analysis, complex coding, research
Smaller context window, silent trimming of long documents, higher hallucination tendency	Lacks voice mode and image generation, more cautious responses
Best for speed, short outputs, creative brainstorming, and multimedia needs.	Best for accuracy, deep analysis, structured writing, and large document processing.

Question 1: How long is the document you most often hand to AI?

This is the question that decides 70% of the cases. Both tools handle short prompts beautifully. The split happens at length.

Under 5 pages of input most of the time → ChatGPT Plus is fine. The context window gap doesn’t matter for your workflow.
Regularly pasting 10+ pages of source material (long PDFs, transcripts, research papers, codebases) → Claude Pro. Claude’s 200K-token context window holds an entire book; ChatGPT Plus tops out lower and starts trimming silently when you exceed it.

The “silently trimming” part is the real problem. ChatGPT will happily answer questions about a long document where it only read the first half — and you won’t know unless you check.

Question 2: Do you write for a living, or write occasionally?

Both tools write English. They write it differently.

ChatGPT writes faster and tighter for short, punchy outputs — taglines, social copy, email subject lines, bullet-point summaries. Claude writes better for long, thoughtful outputs — essays, blog drafts, analysis, anything where voice and structure matter more than speed.

“ChatGPT is a sprinter. Claude is a marathoner. Most people don’t need both at full price.”

If you publish blogs, newsletters, or any long-form content as a meaningful part of your work, Claude’s writing will save you more editing time than ChatGPT’s speed will save you drafting time. If you mostly write short stuff, the inverse.

Question 3: Do you use voice mode, image generation, or DALL-E?

This one is simple. ChatGPT has them. Claude does not.

Use voice for hands-free brainstorming, language practice, or interview prep → ChatGPT Plus. Voice mode is genuinely good and Claude has no equivalent.
Generate images for presentations, blog posts, or social → ChatGPT Plus. Built-in image generation removes a separate subscription.
Don’t care about either → this question doesn’t decide anything for you.

Question 4: Do you build with code, even casually?

For coding, the gap has narrowed but not closed. Claude Pro tends to produce cleaner, more architecturally sensible code for non-trivial problems — refactoring, multi-file changes, debugging. ChatGPT Plus produces code faster and with looser structure.

If you’re a serious coder or you use AI to ship anything more than glue scripts, Claude is the safer pick. If you mostly want quick “how do I do X in Python” answers, either works.

Question 5: How much do you care about hallucinations?

Both tools hallucinate. They hallucinate differently.


Type of error	ChatGPT tendency	Claude tendency
Confidently invents facts	Higher	Lower
Refuses harmless requests	Lower	Higher (more cautious)
Makes up citations	Higher	Lower
Adds disclaimers you didn’t ask for	Lower	Higher

If your work involves citing sources, quoting numbers, or summarizing technical documents — Claude is meaningfully better, because it more often says “I don’t know” or “this isn’t in the document” instead of inventing something plausible. If your work is creative or exploratory and you don’t mind editing for accuracy, ChatGPT’s looser style can actually be more useful.

The decision tree

The honest verdict on “both”

I get this question a lot: “Should I just subscribe to both?” My answer is no, with one exception.

The exception: if you genuinely use AI 4+ hours a day for work where the difference between a great answer and a good answer is worth real money, both makes sense. That’s maybe 5% of paying users.

For the other 95%, paying for both is paying $20/month for the comfort of “having access” to a tool you’ll open twice a month. The honest move: pick one based on the tree above, run it for 90 days, audit your usage, switch if your habits changed.

What about the free tiers?

Both have generous free tiers in 2026. If you’re doing maybe 30 prompts a month, the free tier of either tool is fine — and the free Claude is unusually generous (you get most of the same model, just with rate limits).

The Pro upgrade is for people who are hitting the rate limit, need the larger context window, or want priority access during peak hours. If you’re not in that category yet, pocket the $20.

The bottom line

The decision tree, condensed: long documents and long writing → Claude. Voice, images, and short tasks → ChatGPT. Both, regularly and at full intensity → maybe both, but verify before committing.

The worst case isn’t picking the “wrong” one. The worst case is paying $40 a month and using $5 worth. Pick one, use it for 90 days like you mean it, then re-evaluate. That’s a better workflow than perpetually wondering if the other tool would have been better.

Key takeaways:

Long documents and long-form writing → Claude Pro
Short tasks, voice mode, image generation → ChatGPT Plus
Both subscriptions only make sense if you use AI 4+ hours a day for paid work
Free tiers of either are sufficient for under 30 prompts/month
Re-audit every 90 days; the right pick changes as your work changes

About the author

→ Connect on LinkedIn · More about Shahid · Latest posts

How I Use Claude + Trello + Zapier to Automate Client Project Updates

Shahid Saleem — Wed, 13 May 2026 20:33:41 +0000

Originally published at PickGearLab — practical AI tutorials for writers, freelancers, and small business owners.

By Shahid Saleem

|

Founder & Editor, PickGearLab

|

5 min read

Client project management has a specific failure mode: things fall through the gap between conversations. A deliverable gets approved on a call, someone forgets to update the project board, no one sends the status email, and three days later the client is chasing you wondering what’s happening.

I spent two years solving this with manual discipline — religiously updating Trello after every call, drafting and sending status emails, writing end-of-week summaries. It worked, but it was friction-heavy and the first thing to slip when things got busy.

The stack I use now handles most of that automatically: Trello for project tracking, Zapier for automation triggers, and Claude for drafting the communications. Here’s how it’s wired together.

How the Three Tools Divide the Work

Trello is the source of truth. Every client project lives on a Trello board with cards representing deliverables. Cards move through lists: Backlog → In Progress → Review → Done. Everything that matters about a project — status, notes, attachments, due dates — lives in Trello.

Zapier watches Trello for events and triggers actions when things change. When a card moves to a new list, Zapier can send a notification, create a task, or fire a webhook. It connects Trello’s state changes to the rest of the workflow without any manual intervention.

Claude handles the communication drafting. Rather than sending automated robotic status emails, I use Claude to turn structured project data into natural-sounding client updates. The output goes into a draft — I review and send — not directly to the client.

Step 1: Set Up the Trello Board Structure

Every client gets a board with these standard lists:

Backlog — agreed deliverables not yet started
In Progress — active work this week
Awaiting Client — delivered, waiting for feedback or approval
Review — client feedback received, in revision
Done — completed and signed off
Blocked — stalled, needs action from client or third party

Each card has a due date, a description of the deliverable, and a checklist of sub-tasks. This structure is what Zapier reads to generate status data.

Step 2: Set Up the Zapier Automation

The core automation: when a card moves to “Awaiting Client” in Trello, Zapier fires a trigger. The trigger sends the card’s details — title, description, due date, any checklist notes — to a Google Sheet row (or a Notion database if you prefer). This creates a log of every deliverable submitted for client review.

A second automation runs on a weekly schedule (Friday at 4pm): Zapier pulls all cards that moved during the week — what was completed, what is awaiting feedback, what is in progress — and compiles a structured summary. This summary gets sent to a specific email address or Slack channel where I review it before sending anything to the client.

Setting up these two Zaps takes about 30 minutes. The Trello-to-Google-Sheets Zap is straightforward — Zapier has a template for it. The weekly summary Zap requires a search step to filter cards by date modified, which takes a few more steps but is well-documented in Zapier’s interface.

Get the next tutorial in your inbox

One AI tutorial or comparison per week. No filler, no listicles.

Subscribe free

Step 3: Use Claude to Draft the Client Update

Every Friday I copy the weekly summary from the Google Sheet — the structured list of what moved, what’s waiting, what’s coming next — and paste it into Claude with this prompt:

“Write a brief client project update email based on this week’s activity. Tone: professional but direct, not overly formal. Structure: one paragraph on what was completed this week, one paragraph on what is awaiting their feedback and what I need from them, one paragraph on what is planned for next week. Keep it under 200 words. Don’t start with ‘I hope this email finds you well.'”

Claude produces a draft in about 15 seconds. I read it, adjust anything that doesn’t sound like me or that needs client-specific context, and send. The whole process — reviewing the summary, drafting, editing, sending — takes five to eight minutes.

The quality difference over a manually written update is mostly in consistency. When I write these manually after a long week, they’re often thin or slightly off in tone. Claude’s drafts are consistently well-structured, even when I’m tired.

What Gets Automated vs What Stays Manual

Automated: tracking card movements, compiling the weekly data, generating a structured summary log.

Manual (intentionally): reviewing the draft, editing for context, sending. Client communication is not something I want fully automated. The five-minute review exists because sometimes things happened during the week that the Trello board doesn’t capture — a call where expectations shifted, a scope change discussed informally, a client concern that needs addressing. The automated draft is a starting point, not a finished output.

The Result in Practice

Before this setup: status updates happened when I remembered them, which wasn’t always weekly. Clients sometimes chased me. The mental overhead of “I should send an update” was constant low-level friction.

After: weekly updates go out every Friday without fail. Clients have stopped chasing for status because they trust the cadence. I’ve had two clients specifically mention the weekly update as something that made working with me feel more organised than other freelancers they’d used.

The time cost is five to eight minutes per client per week. The relationship cost of not doing it is higher than that.

About the author

→ Connect on LinkedIn · More about Shahid · Latest posts

How to Combine Claude and Zapier to Automate Your Gmail Inbox

Shahid Saleem — Wed, 13 May 2026 20:32:41 +0000

Originally published at PickGearLab - practical AI tutorials for writers, freelancers, and small business owners. Subscribe free at pickgearlab.com/subscribe.

If your inbox looks like a newsletter graveyard with the occasional real email buried underneath, you are not alone. Traditional filters are too brittle, and reading every email to triage it eats up 30-45 minutes a day. This tutorial shows you how to build a smarter inbox using Claude (Anthropic’s AI) and Zapier (the no-code automation platform) working together. The result is a Gmail that reads, classifies, summarizes, and even drafts replies for you. Setup takes about 45 minutes. After that, your email triage drops to under 10 minutes a day.

TL;DR

This tutorial shows how to automate Gmail inbox triage using Claude AI and Zapier, reducing daily email processing from 30-45 minutes to under 10 minutes.

Key takeaways

Combine Claude AI and Zapier to automatically read, classify, summarize, and draft replies for emails.
Initial setup requires about 45 minutes, then daily email triage drops significantly.
Use Gmail labels like AI/Action Required and AI/Newsletter for Claude’s classification.
Configure Zapier to trigger on unread emails, send content to Claude, and apply labels based on AI output.
Claude Sonnet 4.5 offers a good balance of quality and cost for this automation.

What You’ll Need

Here is the starter stack. Everything has a free tier good enough for the basic setup; you will only need paid plans once volume grows.

Gmail account - free. Works with both personal and Google Workspace accounts.
Zapier account - free plan includes 100 tasks/month across 2 Zaps. The Professional plan ($19.99/month) raises limits if you process more than 100 emails/month.
Anthropic Claude API key - pay as you go. Claude Sonnet 4.5 is currently about $3 per million input tokens; you will spend pennies on a personal inbox. Get a key at console.anthropic.com.
Basic Gmail labels you already use (or are willing to create): Action Required, Waiting, Reference, Newsletter, Spam-like. Simple is better.

Step 1 - Create the Gmail Labels Claude Will Use

Automation needs a clear taxonomy. Open Gmail, go to Settings, click the Labels tab, and create these five labels: AI/Action Required, AI/Waiting, AI/Reference, AI/Newsletter, AI/Low Priority. The AI/ prefix makes it obvious which labels are machine-applied versus your own.

You also want a master view. Create a Gmail filter that matches any email with an AI/ label and apply a star. This gives you a single starred view showing every AI-classified email across all categories. Test the labels by manually applying one to an existing email before moving on - this confirms Zapier will be able to see and use them.

Step 2 - Set Up the Core Zap: Gmail to Claude

Log into Zapier and click Create Zap. Choose Gmail as the trigger app and New Email Matching Search as the trigger event. In the search string, type is:unread -label:AI. This runs Claude only on unread emails you have not already processed.

For the action step, add Anthropic (Claude) as the app. Pick Send Prompt as the action. Connect your API key. In the prompt field, paste this:

Prompt: “You are my inbox triage assistant. Read the email below and return JSON with two fields: category (one of: action, waiting, reference, newsletter, low) and summary (a one-sentence plain-English description of what the email is about or needs from me). Email subject: {{subject}}. From: {{from}}. Body: {{body_plain}}.”

Set the Claude model to Claude Sonnet 4.5 (good balance of quality and price). Test the step - Zapier will run the Zap on a sample email and show Claude’s JSON response. Confirm it parses correctly before moving on.

Step 3 - Use Claude’s Output to Label the Email in Gmail

Add a second action step: Gmail ? Add Label to Email. For the label, use a Zapier Formatter step to map Claude’s category output to the actual label name. The simplest way is a Lookup Table in the Formatter: action ? AI/Action Required, waiting ? AI/Waiting, reference ? AI/Reference, newsletter ? AI/Newsletter, low ? AI/Low Priority.

Connect the Gmail action to the mapped label, and map the email ID from the trigger step so Zapier knows which email to label. Test the action. If you go back to Gmail you should see the test email now carries its new AI/ label. Turn the Zap on. Congratulations - you now have an AI inbox triager running 24/7.

Step 4 - Build the Summary Digest

Triage is useful but digest is where you get time back. Create a second Zap with a Schedule by Zapier trigger running every morning at 7 AM your timezone. Add a Gmail action: Find Emails with Search. Search for label:AI/Action Required newer_than:1d. This fetches everything Claude flagged as needing your action in the last 24 hours.

Add a Looping by Zapier step to iterate. For each email, send the subject and summary to a new Anthropic Claude step with this prompt:

Prompt: “Compress these email summaries into a 5-bullet morning briefing. Each bullet: the sender, what they want, and the single fastest action I can take. Emails: {{looped_summaries}}.”

Finally, add a Gmail step to email yourself that briefing. You now wake up to a one-email morning digest of everything that needs your attention - instead of scrolling through 80 new emails.

Step 5 - Add a Draft Reply Zap for the Easy Ones

The last step saves another 15 minutes a day. Create a third Zap with a Gmail trigger that only fires when an email is labeled AI/Action Required AND contains common short-reply phrases like “let me know,” “can you confirm,” “is this still on.” (Use Zapier’s Filter step to require both conditions.)

Add a Claude action with this prompt:

Prompt: “Write a short, polite reply to this email. Two options: (A) a 1-2 sentence ‘yes’ version, (B) a 1-2 sentence ‘no or delay’ version. Keep both under 40 words. Sign as ‘Thanks, [your name].’”

Send Claude’s output to a Gmail action: Create Draft. The reply sits as a draft in your inbox; you just review, pick A or B, and hit send. Zero blank-page friction on routine replies.

Tips to Get Better Results

Start with triage only, add digest and drafts later. Running all three Zaps on day one makes it hard to debug when something misclassifies. Nail triage for a week first.
Keep the classifier prompt short. Claude responds better to tight prompts than to a two-paragraph manifesto. Every extra instruction is a chance for drift.
Never auto-send replies. Drafts only (and if you need to process audio messages landing in your inbox, see our voice memo to notes workflow). The moment Claude sends a reply without you seeing it, one bad hallucination costs you a relationship.
Monitor Claude usage weekly. In the Anthropic Console check your spend. For a personal inbox it should be under $2/month. If it climbs, tighten the trigger search to reduce processed emails.
Use labels, not replies to folders. Labels let you stack classifications and keep the email in the inbox; folders hide emails you often still want to see.

Common Mistakes to Avoid

Three traps kill most inbox automations. First: over-classifying. Stick to five labels. Ten is already too many for Claude to apply consistently. Second: running Zaps on your entire inbox history. Always use newer_than:1d or similar in Gmail searches, or you will burn through Zapier tasks and API credits in minutes. Third: skipping the test phase. Test every action with a real sample email before turning the Zap on. A misconfigured Zap can apply the wrong label to hundreds of emails in an hour.

Conclusion + Next Steps

You now have a smart Gmail that reads everything, tags it, summarizes the important items, and drafts replies for the routine ones. Give it 14 days. Somewhere around day 10 you will notice your inbox no longer feels like a chore - it feels like a filtered feed of only the things worth your attention.

Once this works, try two upgrades. First: add a follow-up tracker Zap that searches for label:AI/Waiting older_than:3d each Monday and asks Claude to draft polite “just checking in” nudges. Second: pipe Claude’s triage output into Google Sheets so you have a searchable log of every classification. Both add serious leverage to the base system and take about 20 minutes each.

About the author

Shahid Saleem writes PickGearLab - a practical blog about AI tools, tutorials, and automation workflows for people who want real results, not another listicle. Certified in Microsoft AZ-900, CompTIA Security+, and AWS AI Practitioner, with 10+ years in enterprise IT.

? Connect on LinkedIn · More about Shahid · Latest posts

I'm Not a Developer. I Built a Working App in One Weekend With Claude Code.

Shahid Saleem — Wed, 13 May 2026 20:32:38 +0000

Originally published at PickGearLab - practical AI tutorials for writers, freelancers, and small business owners. Subscribe free at pickgearlab.com/subscribe.

I am not a developer. I have never shipped production code in my life. I have made HTML buttons that don’t do anything in 2014, and that’s about the extent of it. Last weekend I sat down at my kitchen table on a Saturday morning with a vague idea, opened Claude Code, and by Sunday evening I had a working web app deployed online that I now actually use every day.

This post is the honest weekend journal - what I tried, what broke, what worked, the actual prompts I used, and what the finished app does. If you’ve been watching the “vibe coding” trend and wondering whether a non-developer can really ship something useful, this is one data point. Spoiler: yes, but with caveats I’ll cover at the end.

TL;DR. Saturday 9am to Sunday 6pm. Built and deployed a personal expense splitter for my family group chat. Total Claude Code prompts: 47. Total hours hands-on: about 11. Total dollars spent: $0 (used Claude Pro and free Vercel hosting). I’d do it again. I would not have shipped this without AI - not in a weekend, not in a month.

TL;DR

A non-developer built and deployed a working web app in one weekend using Claude Code, demonstrating that AI can enable rapid development for novices with specific caveats.

Key takeaways

Picking a small, useful idea is harder than coding itself for rapid AI-assisted builds.
Claude’s clarifying questions before coding save significant rework time.
When Claude’s code is wrong, ask it to explain its reasoning for a concrete example.
Deployment with AI assistance can be surprisingly simple, even for non-developers.
AI is crucial for understanding unfamiliar technical concepts and overcoming roadblocks.

Saturday 9:00am - picking the idea

The hardest part was not the code. It was picking something small enough to actually finish in two days but useful enough to want to use after.

I rejected three ideas in the first hour:

A budgeting app - too broad, the database design alone would eat the weekend
A meal planner - I don’t cook, the dogfood test would fail
A habit tracker - the world has a million of these and I wouldn’t open it past day three

What I picked: a tiny web app where my family group chat (six people, four cities, monthly shared expenses) can log who paid for what and see who owes whom at the end of the month. Splitwise exists. I didn’t want Splitwise. I wanted a single page, no login, no account, that I could share by URL.

Saturday 10:00am - the first prompt

I opened Claude Code in the terminal and typed the dumbest possible version of what I wanted:

I want to build a small web app where my family can log shared expenses and see who owes whom. No login. Six users with names hardcoded. Anyone with the URL can add an expense (who paid, amount, what for). The page shows running totals and a "settle up" view that says who pays whom. Use whatever stack is easiest to deploy free. Walk me through the build step by step.

Claude responded with a stack recommendation (Next.js + a simple JSON file or SQLite for storage, deployed to Vercel), a project structure, and the first file to create. It also asked three clarifying questions before writing any code - what currency, how should the settle-up math work when three people split unevenly, do I want history or just current month.

Those clarifying questions saved me an hour of rework. The lesson: if Claude doesn’t ask clarifying questions, you haven’t given it enough context.

Saturday 12:30pm - first working version

Three hours in, I had a local app running on my laptop that could add expenses and show totals. Claude wrote the React components, I copy-pasted them into files, ran the commands it told me to run, and watched things appear in the browser.

“The first time the localhost page actually loaded with a working form, I felt like I had hacked the universe. I had not. I had pasted what Claude told me to paste.”

The hard part was the gap between Claude saying “now run npm install” and me knowing what npm was. The first hour, every command was a question: “what does this do, and what happens if I run it.” Claude was patient. By hour four, I stopped asking and just ran them.

Saturday 3:00pm - the first wall

I tried to add a feature: when an expense is “split unevenly” (Alice paid $100 but only owes $20), the math should reflect that. Claude wrote the code. The numbers came out wrong.

I told Claude the numbers were wrong. It rewrote the code. Still wrong. We went around three times. Eventually I asked it to walk me through the math by hand for a specific example. It found its own bug - it was double-counting the payer’s share.

Lesson: when Claude is wrong, telling it “this is wrong” once is fine. After two failed fixes, switch tactics. Ask it to explain its reasoning for a concrete example. The bug will fall out.

Saturday 7:00pm - Saturday wraps

By dinner, the app worked locally for all the basic cases. It did not handle edge cases (zero-amount expenses, negative numbers, the user typing emojis into the amount field). It was not deployed. I had not made it look like anything. But the math worked.

I closed the laptop and watched a movie. Treating it as a finite project with a hard stop on Sunday night was the only thing keeping it from sprawling.

Sunday 9:00am - deployment

This is where I expected to fail. Deploying anything has historically been the part where I give up.

I asked Claude: “I want to deploy this. The cheapest, simplest way that doesn’t require me to learn anything new.” It walked me through a Vercel deployment in 12 steps. Most of the steps were “click this button on this page.” Two were “paste this command into your terminal.” Twenty minutes later the app was live at a vercel.app URL I could send to my family.

The first message in the family group chat after I shared the link was “is this safe.” Fair question. I had not added a password.

Sunday 11:00am - the password problem

Adding password protection without a real auth system was the trickiest part of the weekend. Claude suggested a single shared password stored in an environment variable, checked client-side. Not real security - but enough that the URL couldn’t be casually shared with strangers.

I said: “explain to me what ‘environment variable’ means, and what ‘client-side’ means.” Claude explained. I made it work. The lesson here is that I would have been completely stuck without AI - not because the concept was hard, but because I wouldn’t have known what question to type into Google.

Sunday 2:00pm - making it look like something

The app worked but looked like a 1998 form. I asked Claude to “make it look clean and modern, similar to Notion’s aesthetic.” It pulled in Tailwind CSS and rewrote the components. The before-and-after was the most satisfying moment of the weekend. Twenty minutes for visual upgrade that would have taken me a week alone.

Sunday 6:00pm - shipped

I sent the family the link. They started using it that night. As of this morning, six expenses logged across four people. The math is right. Nobody has complained.

What I’d tell someone trying this


Lesson	What it means in practice
Pick something small	If you can’t describe the whole app in two sentences, it’s too big for a weekend
Use an idea you’ll actually use	Building for “users” you don’t have means no feedback, no motivation
Let Claude pick the stack	Don’t argue with the technology choice unless you have a real reason
Run every command Claude gives you	You don’t need to understand the command before running it. Understand later.
When stuck, ask for the math by hand	“Walk me through this with a specific example” finds bugs faster than “this is wrong”
Treat it as a finite project	Sunday at 6pm means Sunday at 6pm. Stop adding features.

The honest caveats

This post is going to make some people angry, so let me address the obvious objections honestly:

This is not “real” software engineering. True. Real engineering involves architecture decisions, tradeoffs, security review, testing, operations, and a hundred things I did not do. What I built has bugs I don’t know about. It would not survive contact with the public. It works for six family members.
I didn’t really “build” it - Claude did. Sort of. Claude wrote the lines. I made every decision about what to build, picked when something was wrong, debugged the math, and was the one whose family used it Sunday night. The line between “built” and “directed the building of” is blurry, and I think that blur is the actual story of where software is going.
It would have been easier to use Splitwise. Yes. The point wasn’t to solve the splitting-expenses problem. It was to find out whether someone with my skill level could ship anything at all in a weekend. The answer is now yes.

The bottom line

If you’ve been waiting for permission to try building something with AI as a non-developer: this is your permission. Pick something small, give yourself 48 hours, and use Claude Code (or Cursor, or any of the equivalents). The worst case is you spend a weekend learning what npm is. The best case is you ship something your family actually uses.

I’m not going to pretend this turns me into a software engineer. It doesn’t. But it does mean that the gap between “I have an idea for a small tool” and “the small tool exists” used to be roughly infinite, and is now roughly two days. That’s not nothing.

Key takeaways:

Non-developers can ship working web apps in a weekend with Claude Code
Pick something small enough to finish (one page, hardcoded users, no login)
Let the AI choose the stack - don’t argue without reason
When debugging, ask Claude to walk through the math with a specific example
Set a hard deadline (Sunday 6pm) and respect it - scope creep is the real enemy
Free hosting on Vercel + Claude Pro at $20/month is the entire toolchain

About the author

? Connect on LinkedIn · More about Shahid · Latest posts