AgentSkillsCN

plan-refinement

仅用于文档编写。仅对 AI 执行方案进行细化；输出仅为 Markdown 格式的方案文本，不包含任何代码。

SKILL.md
--- frontmatter
name: plan-refinement
description: Documentation-only. Refines plans for AI execution; output = plan markdown only. No code.
context: fork
agent: Plan

Plan Refinement

Plan → checklist → gaps → on any feedback: validate + verification loop → only when user confident to write → write new file.

Scope

  • •Output: Plan markdown in ./docs/plan/. No code, tests, or config.
  • •No code: No blocks/snippets; no source edits.
  • •Flow: Read → Review → Refine → Gaps → Iterate → on feedback: re-validate + "correct or more changes?" loop → "Confident to write?" loop → Write new file.

Before starting

Do not infer the plan. Always ask the user explicitly:

  • •Which plan to use? (path to existing plan in ./docs/plan/), or
  • •Start a new plan from scratch?

Do not read, guess, or infer from context. Wait for the user's answer before proceeding.

Feedback and verification (mandatory)

On every feedback (gap answers, task/plan changes):

  1. •Incorporate feedback into plan.
  2. •Re-run validation (Parameters to Review).
  3. •Present updated plan. Ask: "Is this correct or more changes?"
  4. •If more changes → go to 1. If user accepts → continue (or ask "Confident to go to writing?" if that step applies).

Do not write until user has accepted in this loop and (for final write) said confident to go to writing.

Iteration and writing

StepAction
1Present plan (full or summary).
2Gaps: Run Gap analysis (below). List in Gaps. Agnostic phrasing. Ask fill/confirm.
3"Accomplish your requirements? Fill/confirm Gaps first if any; then changes or confirm."
4On feedback (gaps or tasks): incorporate → re-validate → present → "Correct or more changes?" → loop until accepted.
5When accepted: "Confident to go to writing, or more changes?" If more → step 4. Loop until confident.
6Only when confident to write: Write. New file: ./docs/plan/{plan-name}-YYYY-MM-DD-HHmm.md with current date-time. If path exists → different timestamp. Never overwrite.

Workflow: Existing plan

  1. •Read plan.
  2. •Review vs Parameters to Review. Report: pass/fail; table Parameter | Status | Reason.
  3. •Refine plan to satisfy parameters. No code.
  4. •Gaps: Run Gap analysis. List in Gaps. Present; user fill/confirm.
  5. •On any feedback: Incorporate → re-run Parameters to Review → present → "Correct or more changes?" → loop until accepted.
  6. •"Confident to go to writing?" If no → step 5. Loop until yes.
  7. •Write. New file only. Current date-time; never overwrite. Front matter = allowed keys. Include Agent Execution Rules + Gaps (resolved).

Workflow: Create from zero

  1. •Intent (problem + solution); scope In/Out.
  2. •Codebase: Files to Reference grouped by folder (per folder: table File | Purpose; paths relative to folder).
  3. •Design: UI? Ask Figma/refs. External resources = bullets or "N/A".
  4. •Decisions: defaults, data, errors, dialogs. Bullets; infer or "TBD".
  5. •Draft: front matter (allowed keys); Title & Overview; Context (Files by folder, Technical Decisions, External resources); Implementation Plan (tasks - [ ]: Title, File, Action, Notes; AC Given/When/Then); Additional Context if needed; Summary.
  6. •Self-review: Parameters to Review; Writing Guidelines.
  7. •Gaps: Run Gap analysis. List in Gaps. Present; user fill/confirm.
  8. •On any feedback: Incorporate → re-validate → present → "Correct or more changes?" → loop until accepted.
  9. •"Confident to go to writing?" If no → step 8. Loop until yes.
  10. •Write. New file only. Same as Existing step 7. Create ./docs/plan/ if needed.

Parameters to Review

ParameterRequirement
Output fileNew file only. ./docs/plan/{plan-name}-YYYY-MM-DD-HHmm.md; current date-time at write. Never overwrite plan-you-read or existing path; if exists → different timestamp.
Front matterYAML ---. Only: title, slug, created, updated, status, stepsCompleted; optional epic, stories. No other keys.
Title & OverviewProblem (short), Solution (one para), Scope (In/Out). No code.
Files to ReferenceGroup by folder: one subsection per folder (base path). Per folder: table File | Purpose. File = path relative to that folder. No single flat table with full paths.
Technical DecisionsBullets: defaults, data, errors, dialogs.
External resourcesBullets. UI? Ask Figma/refs; none → inferred. No UI → "N/A".
Implementation PlanTasks - [ ]. Each: Title, File, Action, Notes. AC Given/When/Then.
Additional ContextOptional. If absent, ask. If yes: bullets (deps, testing, notes).
Agent Execution RulesPer task, mandatory order: (1) Dev context first. (2) TDD — no exceptions: add/extend tests → run (or scope) → implement until green → then mark complete. (3) Plan metadata. Skip = invalid.
SummaryEnd: what it delivers + work (scope, tasks, outcome).
GapsRun Gap analysis. List in Gaps; agnostic. User fill/confirm before write.

Gap analysis

Before listing Gaps, run; add each finding to Gaps (agnostic phrasing).

1. Designs (if plan has UI)
Figma/visuals in External resources? No → gap: request or confirm inferred. Yes → cover empty/error/loading for task-relevant states? No → gap.

2. Task list
Per task (Title, File, Action, Notes): what is unspecified? (error handling, loading, empty state, validation, navigation, dependency, edge case). One gap per undefined aspect.

3. Dimensions

  • •Architecture: Where does X live? Which provider/notifier? Data flow?
  • •Functionality: Validation, retry/offline, errors (where/recover?), loading (skeleton vs spinner).
  • •Product/UX: Empty/error state (copy, CTA, retry?). Navigation from/to. Confirmations. Copy/labels.

4. Dev / Additional Context
Plan missing Additional Context or dev refs (MCP, design, codebase) and tasks need it → gap: provide or confirm not needed.

Agent Execution Rules (in every written plan)

Place after Context, before Implementation Plan. Order is mandatory for every implementation task. Never skip 1 or 2.

Per implementation task (mandatory):

  1. •Dev context first - always. Fetch required context (MCP, Figma, codebase) before implementation.
  2. •TDD - mandatory, no exceptions. Add/extend tests → run (or scoped path) → implement until pass → then mark complete. Applies to every task (UI, refactors, bug fixes, all). Skipping is invalid.
  3. •Plan metadata - mandatory. status: implementing; append to stepsCompleted; on done → status: completed.

In every written plan, include this block in Agent Execution Rules:

Mandatory for every implementation task: 1) Dev context first. 2) TDD: tests → run → implement until green → mark complete. Never skip. Not optional.

Writing Guidelines

  • •One intent per item; one action per task. Explicit: file path, symbol. Verbs: Add, Replace, Move, Delete, Rename.
  • •Tables/bullets over prose. No implicit context; cite paths.
  • •Concrete: "Add X in path/file.dart" not "Add a method." Task = file + action + notes. AC = one-sentence testable.

Output format (report)

Summary (Parameters N/M, actions). Per parameter: ✅/❌ + comment. Optional: proposed changes. After present: Gaps → fill/confirm; "Accomplish requirements?" On any feedback: re-validate → present → "Correct or more changes?" → loop until accepted. Then "Confident to go to writing?" → loop until yes. Only then write new file (current date-time; never overwrite).

Project-specific params → Parameters to Review or linked ref.