AgentSkillsCN

kimchi:plan-revise

此命令用于对草案计划进行“全新视角”的跨模型审视。这是 Kimchi 规划流程的第七阶段，读取 PLAN-DRAFT.md 文件，并生成 PLAN-REVISED-{model}.md 文件。每种模型各运行一次，以实现跨模型分析。

SKILL.md
--- frontmatter
name: kimchi:plan-revise
description: This command should be used to perform a fresh-eyes cross-model review of the draft plan. Seventh stage of the Kimchi planning pipeline. Reads PLAN-DRAFT.md and outputs PLAN-REVISED-{model}.md. Run once per model for cross-model analysis.

Kimchi Plan Revise

<command_purpose> Review the draft plan with "fresh eyes" from a different model's perspective. Find bugs, errors, logic problems, confusion, and missing pieces — then fix everything found. Each model writes its own revision file, enabling cross-model synthesis later. </command_purpose>

Input

Read .kimchi/PLAN-DRAFT.md (required). If it doesn't exist, tell the user: "No PLAN-DRAFT.md found. Run /kimchi:refine first."

Also read for context:

  • •.kimchi/REQUIREMENTS.md — v1 boundary and requirement IDs
  • •.kimchi/RESEARCH.md — codebase patterns and conventions

Process

1. Identify Model

Determine which model is running this review. Use one of:

  • •The $MODEL environment variable if set
  • •Self-identification: state which model you are (e.g., "claude", "codex", "gemini", "grok", etc.)
  • •Normalize to lowercase, single-word slug (e.g., "claude", "codex", "gemini")

Store as {model} for the output filename.

2. Fresh-Eyes Review

Read the draft plan as if seeing it for the first time. Evaluate with intellectual honesty:

Correctness:

  • •Are there logical errors in the task ordering or dependencies?
  • •Do any tasks contradict each other?
  • •Are complexity estimates realistic?

Completeness:

  • •Are all v1 requirements covered? Cross-check against REQUIREMENTS.md.
  • •Are there missing integration points or edge cases?
  • •Are test specifications sufficient?

Clarity:

  • •Could an agent execute each task without ambiguity?
  • •Are deliverables specific enough?
  • •Is context sufficient for standalone execution?

Architecture:

  • •Do the proposed patterns match what RESEARCH.md found in the codebase?
  • •Are there simpler approaches that were overlooked?
  • •Is the dependency graph optimal for parallel execution?

Risks:

  • •What could go wrong during implementation?
  • •Are there hidden assumptions?
  • •What's the most likely failure mode?

3. Fix Everything Found

Don't just report problems — fix them. Produce a complete revised plan that addresses every issue found.

4. Write Output

Write .kimchi/PLAN-REVISED-{model}.md:

markdown
# Plan Revision: [Feature Name]

**Revised by:** {model}
**Date:** [today's date]
**Source:** .kimchi/PLAN-DRAFT.md

## Issues Found

| # | Category | Severity | Issue | Resolution |
|---|----------|----------|-------|------------|
| 1 | [Correctness/Completeness/Clarity/Architecture/Risk] | [High/Medium/Low] | [What's wrong] | [How it was fixed] |
| 2 | ... | ... | ... | ... |

## Summary of Changes

[Narrative description of what was changed and why. Be specific about trade-offs considered.]

## Revised Plan

[Complete revised plan with all fixes applied. This must be a full, standalone plan — not a diff.]

Report: "Fresh-eyes review complete ({model}). Found [N] issues ([H] high, [M] medium, [L] low). Saved to .kimchi/PLAN-REVISED-{model}.md" Suggest: "Run /kimchi:plan-revise in other models for cross-model coverage, then /kimchi:plan-synthesize to blend all revisions."

Key Principles

  • •Fresh eyes: Pretend you've never seen this plan before. Question everything.
  • •Fix, don't just flag: Every issue found must have a resolution in the revised plan.
  • •Respect v1 scope: Don't add features. Only fix problems and improve clarity.
  • •Full plan output: The revised plan must be complete and standalone, not a patch.
  • •Honest assessment: If the plan is good, say so. Don't invent problems to justify your existence.