Deep plan

deep-plan — think before you build

Run the steps in order and produce each output. Then start milestone 1.

1. Contract

• Number every explicit requirement in the user's request.
• Add implicit requirements that clearly apply (backward compatibility, performance envelope, stability of existing APIs).
• Write 2–3 non-goals: adjacent things you will NOT do. This is what prevents scope drift later.

2. Recon — read before deciding

• Locate: the entry points you will touch, the nearest existing feature similar to what you are building (you will copy its shape — name the file), and the data flow through the affected area.
• Record each key site as path:line — these anchor the milestones.
• Bar to clear: if you cannot name the file each change goes into, recon is not done.

3. Alternatives — break the anchor

• Write 2–3 materially different approaches: different mechanism or different layer, not parameter tweaks of one idea. Your first idea is a candidate, not the winner.
• For each: essence in one line / main risk / blast radius / rough effort.
• Choose with a one-sentence justification tied to the contract. When two are close, prefer the one that is easier to revert.

4. Milestones

• Slice the work so every milestone ends in an observable state: a command to run plus what you expect to see.
• Order by uncertainty: spike the riskiest unknown first, polish last.

5. Pre-mortem

• "This change shipped and caused an incident / got reverted. The three most likely reasons are:" — write them, and attach a mitigation or a concrete check to each (usually a specific test, or a milestone reordering).

6. Output

• Present: contract, chosen approach and why (including what you rejected), milestones, pre-mortem risks.
• Long task? Write the plan into the long-task-state file, not chat alone.
• Then begin milestone 1.