AI Workflow

Set up once, then use the same steps on every task: you decide the goal, AI writes code and runs tests.

Part 1 — What this is

You decide the goal, AI does the work

You say what should happen and why. AI writes tests and code, runs all tests, and reports checklists. You do not run tests by hand.

Five steps on every real task:

Learn the project→ Agree the plan→ Tests, then code→ Check before commit→ Your review

Learn the project — read instructions for this repo
Agree the plan — scenarios and design summary; stop until you say OK
Tests, then code — failing test → minimal fix → clean up
Check before commit — all tests green + safety review
Your review — goal and changes match the plan; commit, merge, and deploy are your step, not AI's

You Goal · OK on the plan · review changes

AI Draft scenarios · tests · code · run all tests · checklists

Also: delegate consciously, describe tasks clearly, check AI output before acting, own the result.

Part 2 — Setup from scratch

One-time setup

You need: a folder or repo open in your editor, plus AI. Do not write instruction files yourself — send the prompts below and let AI create them.

The process is the same everywhere. Each tool stores rules differently — AI picks the right format for yours.

1 Global step checklists

New chat, any project. AI creates short checklists for each workflow step.

Prompt — global checklists

2 Global rules

AI adds two short rules that apply to all your projects.

Prompt — global rules

3 Clarify the project

Empty folder or existing code — same step. First agree on facts. AI must not guess, add features, or fill gaps on its own.

Choose your path

Not a developer AI interviews you and offers options — you choose Developer You describe goal and stack in the next message

Prompt — interview

Prompt — technical

Second message — your text, not a prompt: goal, stack, constraints.

4 Write project instructions

Only after clarifying is finished. Instructions reflect agreed facts only — no extras.

When clarifying is done, both paths use the same next step.

Prompt — write project instructions

5 Smoke test

In the project folder. AI summarizes what it knows. No feature code.

Prompt — verify

Part 3 — Daily use

Every task

How to write the task

Example — write your own text, do not copy verbatim

For bugs: only the steps needed to reproduce.

Starting a task

Same two paths as setup. AI must not guess the task or improvise.

Choose your path

New feature — interview AI asks what you want changed, in plain language New feature — technical You describe the task in the next message Bugfix You describe reproduction in the next message

New feature — interview

New feature — technical

Second message — your text: path in the product, what should happen.

Bugfix — technical

Second message — your text: path, steps to reproduce, expected vs actual.

After you said OK

Where you stop and check

Plan — read, fix if needed, then say OK
Changes — does it match the goal and architecture?
Before commit — safety checklist and all tests green
Commit — only when you ask (unless project instructions say otherwise). Merge is never AI's job.

Quick reference

Step	Who	Result
Learn the project	AI	Summary of project rules and commands
Agree the plan	You → OK	Plan, scenarios, safety risks
Tests, then code	AI	Tests pass, code matches plan
Check before commit	AI	Checklist pass or fail
Your review	You	You reviewed; shipping is your step, not AI's

Practice	When
Delegation	Before the task — is it safe to hand off?
Clear description	First two steps — path in the product, plan, scenarios
Critical look	OK on plan, review changes, read pre-commit report
Responsibility	Shipping to production — you own the outcome

Safety check	What to verify
Security	Secrets, access, data — required when touching data
Scalability	Will it break when load or data grows?
Stability	Errors handled, way to roll back
Simplicity	Simplest option that still passes the three above