# Skill: Convert a one-sentence idea into a step-by-step project blueprint

> Apply this skill before writing any code for a multi-session or multi-feature project -- it transforms a vague request into a verifiable, dependency-aware plan saved in a file so work can resume after context resets.

Source: sodigi·learn -- vibecoding/ot-idei-k-planu · https://sodigi.io/learn/vibecoding/ot-idei-k-planu

## When to use
- The project spans more than one session or more than a few files.
- The task has dependencies (step B cannot start before step A is done).
- The user says something like "build me an app" or "make a shop" without prior planning.
- You want to run some work in parallel (two unrelated features at the same time).

## Core rules
- Plan first; write no code until the plan is approved. Fixing a drawing takes minutes; rebuilding wrong code takes hours.
- Each step must be small enough to verify independently (one clear chunk of work, not "build the whole store").
- Every step must include a self-contained context brief: what to do, how to check it's done, what depends on it -- so a fresh agent with no prior context can execute it cold.
- Explicitly label which steps are sequential (B depends on A) and which are parallel (can run simultaneously).
- Save the plan to a file (e.g., PLAN.md) immediately -- never keep it only in the chat thread.
- After drafting, review the plan with a critical second pass (or second agent) before starting the build.
- When the plan changes mid-build, update the file first, then continue. Do not silently go off-plan.

## Procedure
1. Survey the project: check existing files, notes, and rules before drafting anything.
2. Decompose the idea into 3-12 steps, each the size of one verifiable chunk of work.
3. For each step, write: (a) what to do, (b) what it depends on, (c) how to confirm it is done.
4. Mark which steps can run in parallel and which must be sequential.
5. Write a critical review pass: are there gaps, wrong order, or missing dependencies?
6. Save the complete plan to PLAN.md (or a named file the user specifies).
7. Show the plan to the user and wait for approval before executing any step.

## Ready-to-use prompt
```
Do not write any code yet. First, put together a step-by-step plan for this
project.

Requirements for the plan:
1. Break everything into small steps, each the size of one clear chunk of work.
2. For each step write: what to do, what depends on what, and how to check
   that the step is done.
3. State which steps can be done in parallel and which must go strictly
   in sequence.
4. Save the plan to a file called PLAN.md so I can come back to it later.
5. Show me the plan and stop -- wait for my "ok," and only then start
   the first step.
```

## Pitfalls
- Asking for the entire project in one sentence and expecting it to be coherent -- the agent will choke or produce a tangled mess.
- Skipping the plan for a big project because "it feels faster" -- redoing badly structured code always costs more time than writing the plan.
- Making steps too large ("build the whole store") -- such steps are impossible to verify and fail invisibly.
- Keeping the plan only in the chat thread -- after a session restart it is gone forever.
- Treating the plan as immovable -- it should change deliberately when reality changes, but always update the file first.
- Ignoring the dependency order -- building payments before authentication leads to expensive rewrites.