/ 6 min read / ChatView
The vibe coding startup checklist: from first prompt to first users
A practical six-step workflow for founders using AI coding agents: scope, prompt, review, test, deploy, and learn. Practise the decisions in the free Ship It game.
An agent can write code. You still own the decisions.
Vibe coding can shorten the distance between an idea and a working prototype. It does not remove the need to choose a user, understand the code you ship, or check that the product solves a problem. A polished preview is evidence that a page renders, not evidence that a startup works.
Use this checklist for one small release. Our free solo board game, Ship It, rehearses these tradeoffs through limited actions, runway tokens, and six original logic puzzles. Finishing a game is not a qualification in software engineering; you must still execute and verify the workflow on your real project.
1. Scope one useful workflow
Name the user and the job before naming the features. "Freelance designers need to save a client brief and retrieve it before a call" is a more useful starting point than "an AI productivity platform." Pick a success condition you can observe, such as three testers saving and retrieving a brief without help.
Write down what is explicitly outside the release: billing, team workspaces, imports, or recommendations. A smaller scope makes both human review and agent work easier. It also makes a disappointing result easier to interpret: you know which assumption you tested.
2. Prompt with boundaries and acceptance criteria
Give the agent the repository context, the small task, the existing conventions, and a way to check the result. Ask for a plan before a broad change. Work in reviewable increments instead of asking it to generate every feature at once.
Example: "Add a client-brief form using our existing form components. It needs a title and notes. Reject an empty title, save for the signed-in user, and show a clear success state. Add focused tests. Do not change billing or authentication configuration. Summarize the files changed and the checks run."
The prompt is a starting specification, not a warranty. An agent can misunderstand a constraint or produce a test that repeats its own mistake. Your next steps are where you collect evidence.
3. Review the diff, especially the boundaries
Read changed files and dependency additions. Ask why an unrelated file changed. For a saved brief, verify that one user cannot retrieve another user's record. Trace where secrets are read and make sure they never enter browser bundles, public environment variables, or logs.
Treat generated explanations as claims to verify. Request a walkthrough of unfamiliar code, then inspect the actual implementation. For sensitive workflows such as payments, access control, or personal data, involve someone qualified to review those risks.
4. Test the journey and its failure cases
Run the tests and read their output. For the brief example, cover a valid save and reload, an empty title, a failed network request, and an unauthorized read. A green test suite that never checks authorization cannot establish authorization correctness.
Then use the interface as a person would. Try a phone-sized viewport, a keyboard, and a slow connection. Confirm that a failed save is not presented as success and that a retry does not silently duplicate data.
5. Deploy with a way back
Use the deployment platform's server-side secret settings. A variable prefixed for browser exposure, such as NEXT_PUBLIC in Next.js, is not a secret store. Deploy the smallest useful release, retain the previous version, and document how to roll back.
Test the production URL after deployment. Check configuration, a real save-and-reload journey, and error logs. A successful build does not prove the live database, permissions, or environment values are correct.
6. Learn from a few real users
Observe whether users finish the intended job. Ask what they expected at the point they hesitated. Prefer concrete behavior over a general "looks great." Record the result, change one thing, and run the workflow again.
A useful build-in-public update contains the workflow shipped, an honest observation, and the next experiment. Never publish private customer records or credentials as proof of progress. ChatView can help you reach your coding agent from your phone; it does not replace your judgment about what is safe to approve.
Ship It: board rules
Start with ten runway tokens and two actions per week. Visit each workspace at most once per week: interview for evidence, build for progress and debt, review for quality and less debt, share for reach, or freelance for runway. Support adds evidence and quality; after release it also adds a retained customer.
Each puzzle attempt costs one action. Meet its listed resource requirements first. A correct answer spends only resources marked as spent; an incorrect answer costs the action but keeps your resources. Hints are free. Ending a week discards unused actions, costs one runway, refreshes workspaces, and applies the event shown on the board.
Clear all six checkpoints by the end of week twelve, before runway reaches zero. Release earns the first customer; support brings the other two needed for the final checkpoint. The journal records your decisions. Refreshing the page starts over.
This original worker-placement and logic-puzzle game is a simplified learning simulation. Real customer retention is not guaranteed by providing support, and real startups do not follow a fixed event calendar. The puzzles are original and are not affiliated with CS50 or adapted from its Puzzle Day materials.