mcpField
S
self-review

self-review

github.com/arthurzakirovAgent workflows0159 installsUpdated 5d ago
Repository

About self-review

self-review examines the current conversation or a chosen number of days of session history to identify assistant mistakes and propose concrete improvements. It is for users who want to refine prompts, skills, or other controls using conversation evidence.

Key features

  1. 01Analyze errors in the current conversation
  2. 02Mine and review historical sessions
  3. 03Map problems to suitable improvement mechanisms
  4. 04Propose edits backed by transcript examples

Use cases

  1. 01Review missing information in an answer
  2. 02Find recurring mistakes across past sessions
  3. 03Choose whether to change a skill, rule, or hook

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/arthurzakirov/systemsmith --skill self-review

After installing, type/self-reviewin Claude Code to confirm it loaded.

Details

From the project README

Self Review

If $ARGUMENTS is empty or non-numeric, analyze the current conversation. If it is a number, treat it as a request to mine that many days of historical sessions.

Current Conversation Mode

For each mistake, capture:

  1. What the assistant concluded
  2. What the actual answer turned out to be
  3. What information was missing
  4. Whether that information was discoverable
  5. What concrete change would prevent the same mistake

Output shape:

### Mistake: <what went wrong>
**Assistant said:** "<quoted claim>"
**Actual answer:** "<what turned out to be true>"
**Missing info:** <missing context>
**Findable?** YES / NO
**Fix:** <skill, prompt, hook, data-source, or workflow change>

After listing mistakes, propose concrete edits and ask the user which to implement.

Historical Mode

Step 1: Pre-Filter

Use a mining script or equivalent search only as a pre-filter to find candidate sessions.

Skip frozen or retry-only sessions caused by infrastructure issues rather than assistant behavior.

Step 2: Deep Read

Read flagged sessions message by message. For each user intervention, ask whether the assistant could have already done that without being told.

Step 3: Choose The Right Fix Mechanism

Map the failure to the right control surface:

MechanismBest for
CLAUDE.md or equivalent global instructionUniversal behavior rules
Path-scoped rulesFile-type-specific behavior
SkillsSituational knowledge
MemoryPersistent decisions or preferences
HooksAutomated guardrails that must always run
AgentsIsolated delegated task types
CommandsExplicitly user-invoked prompts
SettingsPermissions, env vars, model config

Decision heuristic:

  • "Assistant keeps forgetting X" -> Hook or always-loaded rule
  • "Assistant should do X when editing Y" -> Path-scoped rule
  • "Assistant lacks knowledge about X during Y" -> Skill
  • "Assistant should delegate X" -> Agent plus routing rule
  • "I want to run X on demand" -> Command
  • "Remember this decision" -> Memory
Step 4: Present Findings With Examples

Every pattern must include concrete examples from transcripts. Do not present regex counts as findings.

For each pattern include:

  • What happens
  • Two or three concrete examples
  • Proposed fix: bucket, file, specific edit, and why it helps
Step 5: Validate Against User Experience

Check with the user that the proposed pattern matches real experience before editing prompts or skills.

Comments

Markdown supported