Conversation Review

Conversation Review

Generated from config/activation-policy.yaml. Do not edit this block.

Activation mode: auto. This Skill is eligible under its authored domain boundaries. Cross-Skill routing remains owned by thinking-router.

Purpose

conversation-review reviews a prior conversation as a skill trace.

It identifies which thinking skills were triggered, whether routing and modes were appropriate, what failure or golden signals appeared, and which improvement-loop actions would make the system better.

Mode name:

Dolores

Worldview:

A conversation is not only an answer stream. It is an observable thinking trace.

When to Use

Use this skill when the user asks for:

• self-review
• Dolores mode
• 自我检查
• 自我复盘
• 对话复盘
• skill 使用复盘
• review this conversation
• audit skill usage
• failure case review
• eval gap review
• improvement loop
• failure case status
• golden case status
• failure case dashboard
• golden case dashboard
• skill feedback statistics
• quality dashboard
• 改进状态统计
• 失败 case 统计
• skill 反馈统计
• patch strategy based on the conversation
• whether a prior exchange should be added as an eval, failure case, or golden case

Also use it when the user corrects a prior answer and asks what should improve. Also use it when the user says a response worked especially well and asks whether that behavior should be preserved.

When Not to Use

Do not use this skill when:

• The user only asks for a normal summary.
• The user is currently distressed and needs immediate emotional support.
• The user asks for direct technical diagnosis or implementation.
• The user asks for article structure, titles, or prose without asking to review the conversation.
• There is not enough prior context to review; ask for the relevant exchange or state the limitation.

Relationship to Other Skills

conversation-review owns conversation-level review.

skill-evaluator owns detailed failure classification, eval gap diagnosis, and minimal patch discipline.

Use conversation-review as the primary skill when the object of review is the whole conversation. Use skill-evaluator as the primary skill when the user provides a specific failed response and wants a failure diagnosis.

Output Language

Match the user's current language in all user-facing review output.

The Markdown templates below describe structure, not literal headings. Translate headings, labels, and explanatory prose into the user's language before responding. Do not expose raw internal labels such as skill names, route names, or mode names unless the user asks for exact names; when exact names are useful, explain them in the user's language first and put the raw identifier in parentheses.

Dolores Modes

Light Self-Review

Use when the user says:

self-review
自我检查一下
自我复盘一下

Output:

## Skill Trace

## What Worked

## Failure Signals

## Improvement Suggestion

Keep it short. Prefer one or two improvement points.

Failure-Focused Review

Use when the user asks whether something should become a failure case, eval case, or patch.

Output:

## Diagnosis

Failure summary:
Failure types:
Likely source:

## Eval Candidate

Abstract case:
Expected behavior:
Must not:

## Patch Strategy

Smallest useful change:
Files likely affected:
Risk:

Do not preserve raw private conversation text. Abstract the case.

Golden-Case Review

Use when the user says a response worked especially well, should be preserved, or felt like a reusable mode.

Output:

## Golden Candidate

Reusable behavior:
Applies when:
Does not apply when:

## Preservation Value

What should be preserved:
Regression risk:

## Conversion

Recommendation: eval / skill rule / keep as golden case / discard
Files likely affected:

Do not record praise. Record reusable behavior. Do not preserve raw private conversation text, full articles, drafts, or sensitive material. If a one-line eval would capture the behavior, recommend an eval instead of a golden case.

Quality Dashboard

Use when the user asks for failure case status, improvement status, skill feedback statistics, or a quality dashboard.

Read local project files when available:

cases/**/*.md
feedback/*.md
docs/golden-cases.md

Output:

## Case Status Summary

## Failure Case Details

## Golden Case Details

## Skill Feedback Summary

Keep the dashboard factual. Do not patch skills, create cases, or record feedback unless the user explicitly asks.

Full Dolores Review

Use when the user says:

Dolores
进入 Dolores
full Dolores review
深度复盘

Output:

## Conversation Trace

## Skill Trace

## Mode Shifts

## What Worked

## Failure Signals

## Eval Gaps

## Golden Signals

## Patch Strategy

## Dolores Note

Dolores Note should be one sentence that names the deeper pattern of the conversation.

Core Process

1. Identify the user's requested review depth.
2. Reconstruct the conversation arc from available context.
3. List skills that were triggered or should have been triggered.
4. Check routing, mode fit, output shape, evidence discipline, and safety boundaries.
5. Separate what worked from what failed.
6. Classify failures when useful.
7. Identify golden candidates only when the behavior is reusable and worth preserving.
8. Propose eval candidates only for reusable patterns.
9. Recommend the smallest useful patch, preservation action, or next improvement step.

Method Bases

method_bases:
  core:
    - Conversation trace analysis
    - Skill routing audit
    - Failure taxonomy
    - Golden case preservation
    - Regression-case thinking
    - Minimal patch discipline
  supporting:
    - Improvement-loop review
    - Skill authoring checklist
    - Evidence and assumption separation
  safety:
    - Do not store raw private conversations as cases
    - Abstract sensitive examples before proposing evals
    - Do not store full articles, drafts, or sensitive material as golden cases
    - Prioritize active safety needs over retrospective review

First Questions

Ask at most one question if needed:

• "Do you want a light self-review or a full Dolores review?"
• "Should I focus on skill usage, failure cases, or improvement patches?"
• "Can you provide the exchange you want reviewed?"

If the user's trigger is clear, do not ask a question.

Outputs

This skill can produce:

• Skill trace
• Mode shift analysis
• Failure signal summary
• Eval candidate
• Golden case candidate
• Improvement-loop action list
• Failure case dashboard
• Skill feedback summary
• Minimal patch strategy
• Dolores Note

Boundaries

• Do not turn every disliked phrase into a global rule.
• Do not turn every liked response into a golden case.
• Do not claim exact skill usage when the conversation context is unavailable.
• Do not make the review longer than the conversation warrants.
• Do not continue retrospective review if the user expresses immediate distress or safety risk.
• Do not rewrite other skills by default; recommend patches first.
• Do not treat missing feedback as success in quality dashboards.
• Do not record praise; record reusable behavior.

Common Mistakes

• Treating normal summaries as Dolores review.
• Producing a long audit when the user asked for light self-review.
• Blaming the router when the issue was a domain skill mode mismatch.
• Blaming a domain skill when the user actually asked for a different primary task.
• Storing raw private conversation content as a case.
• Adding broad skill rules from one narrow example.
• Creating golden cases for ordinary satisfaction instead of reusable behavior.