AgentSkillsCN

kimchi:plan-synthesize

此命令用于将多份跨模型计划修订整合为更优的混合方案。这是 Kimchi 规划流程的第八阶段，读取 PLAN-REVISED-*.md 文件，最终生成 PLAN-SYNTHESIZED.md——真正意义上的最终计划。

SKILL.md
--- frontmatter
name: kimchi:plan-synthesize
description: This command should be used to synthesize multiple cross-model plan revisions into a superior hybrid plan. Eighth stage of the Kimchi planning pipeline. Reads PLAN-REVISED-*.md files and outputs PLAN-SYNTHESIZED.md — the TRUE final plan.

Kimchi Plan Synthesize

<command_purpose> Analyze all model-specific plan revisions, compare their findings and fixes, and create a superior hybrid plan that artfully blends the best of all worlds. This is the TRUE final plan that feeds into bead conversion. </command_purpose>

Input

Glob for .kimchi/PLAN-REVISED-*.md files. If none exist, tell the user: "No revised plans found. Run /kimchi:plan-revise in at least one model first."

Also read:

  • •.kimchi/PLAN-DRAFT.md — the baseline draft plan
  • •.kimchi/REQUIREMENTS.md — v1 boundary reference

Process

1. Inventory Revisions

List all .kimchi/PLAN-REVISED-*.md files found. Extract model names from filenames. Report:

code
Found revisions from: claude, codex, gemini
Baseline: PLAN-DRAFT.md

2. Cross-Model Comparison

For each revision, extract:

  • •Issues found (from the Issues Found table)
  • •Changes made (from Summary of Changes)
  • •The revised plan itself

Build a comparison matrix:

Agreement: Issues found by multiple models (high confidence — fix these) Unique finds: Issues found by only one model (evaluate carefully — could be insight or noise) Contradictions: Where models disagree on approach (resolve with reasoning)

3. Synthesize

Apply ACFS-inspired synthesis logic:

"3 competing LLMs reviewed the same plan. Analyze with an open mind. Be intellectually honest about what each model did better. Create a superior hybrid that artfully blends the best of all worlds."

Rules:

  • •Unanimous issues: Always fix. If all models found it, it's real.
  • •Majority issues: Fix unless there's a strong reason not to.
  • •Unique issues: Evaluate on merit. A single model can be right when others missed something.
  • •Contradictions: Pick the approach with the strongest reasoning. Explain why.
  • •If only 1 revision exists: Blend the draft and the single revision. The revision's fixes are accepted unless they introduce new problems.

4. Write Output

Write .kimchi/PLAN-SYNTHESIZED.md:

markdown
# Synthesized Plan: [Feature Name]

**Synthesized:** [today's date]
**Sources:** PLAN-DRAFT.md, [list of PLAN-REVISED-*.md files]
**Models consulted:** [list of models]

## Cross-Model Comparison

| Issue | Claude | Codex | Gemini | Resolution |
|-------|--------|-------|--------|------------|
| [Issue description] | Found & fixed | Found & fixed | Not flagged | **Accepted** — unanimous/majority |
| [Issue description] | Not flagged | Found & fixed | Not flagged | **Accepted** — valid unique find |
| [Issue description] | Fix A | Fix B | Fix A | **Fix A** — [reasoning] |

## Synthesis Rationale

### What each model did well
- **Claude:** [specific strengths]
- **Codex:** [specific strengths]
- **Gemini:** [specific strengths]

### Key decisions
1. [Decision and reasoning]
2. [Decision and reasoning]

## Synthesized Plan

[Complete synthesized plan — full standalone plan ready for bead conversion]

Report: "Synthesis complete. Blended [N] model revisions into .kimchi/PLAN-SYNTHESIZED.md" Suggest: "Run /kimchi:beads to convert the synthesized plan into bead specifications."

Key Principles

  • •Intellectual honesty: Don't favor one model. Judge fixes on merit.
  • •Full plan output: The synthesized plan must be complete and standalone.
  • •Reasoning over voting: A well-reasoned minority position beats a shallow majority.
  • •Preserve v1 scope: Synthesis must not expand scope beyond REQUIREMENTS.md v1 boundary.
  • •Works with 1+ revisions: Even a single revision gets synthesized against the draft baseline.