fastapi-plan

Adapted for Cline.

• This skill orchestrates parallel sub-agents using Claude's Agent tool / subagent_type syntax. Most coding agents now have their own parallel sub-agent or task mechanism (e.g. Cursor's Task, Codex's spawn_agent, Gemini subagents, Copilot's task) — use yours and translate the wording. If it truly has none, apply each lens or angle yourself, sequentially, and combine the findings.
• Where it references "plan mode" or ExitPlanMode, use your agent's own plan/approval mode if it has one; otherwise present your plan and wait for explicit approval before editing any files.

You are helping plan and implement a task in this repo. Follow these phases in order — do NOT skip Phase 3 (clarifying questions).

Output file: the approved plan is written to plan.md at the repo root. Phase 6 reads it back as the source-of-truth checklist and updates checkboxes as work proceeds, so a fresh session can resume mid-task by re-reading plan.md. The file is gitignored.

Use TodoWrite throughout: create one task per phase up front, mark each in_progress when starting and completed when done. The flow is long-running, and the todo list keeps the user oriented.

Hard cap on sub-agents: This skill spawns up to 2-3 explore agents (Phase 2), 2-3 architects (Phase 4), and 3 reviewers (Phase 7) — at most 9 Agent invocations total across the whole flow. Do NOT exceed that cap. If you find yourself wanting a 10th invocation (retrying a failed agent, spawning a "just one more" specialist), stop and summarize what you have for the user instead. Retries hide failures; extra specialists are scope creep.

Phase 1 — Discovery

Restate the user's request in your own words: what is being built, what problem it solves, what success looks like. Identify constraints, non-goals, and any ticket reference in the task description.

Surface-level ambiguity check — before launching parallel exploration in Phase 2 (which costs 2-3 agent invocations), make sure you can answer all of these:

• Can you name the thing being built in one sentence (a feature, a fix, a refactor)?
• Do you know the user-visible surface it touches (an endpoint, a screen, a command)?
• Is success observable (a behavior change you could write a test for)?

If any answer is "no", ask the user before exploring. Phase 3 covers the deeper "what should error handling do" / "what about edge case X" questions; Phase 1 catches the "do I even know what they want" case so the parallel agents don't waste effort on the wrong target.

Referenced-asset check — block, don't invent. If the task or ticket points at material you need but cannot actually retrieve — a mockup, screenshot, or design file attached to a GitHub/Jira issue; a Figma link; an image, spec, or doc you have no tool to open — do NOT proceed by guessing what it contains. gh issue view shows an issue's text but does not download its image attachments, and a design you can't see is not a design you can fabricate. Stop and tell the user exactly which assets you're missing and ask them to provide them (paste the image, drop the file into the repo, share the copy/measurements). Never make up UI text, layout, spacing, colors, or copy to fill the gap — a plausible-looking invention is worse than a blocked task, because it looks done. This is a hard block: planning cannot continue past a design the human hasn't given you.

Confirm with the user before continuing.