Self-Improvement Omni

Tracks improvement signals (+1/-1) and directive hints for a named cron job, based on OpenClaw-RL research (arxiv.org/abs/2603.10165).

Scope (important)

• This skill only ever acts on the one job you name via --job.
• It is invoked explicitly — it does not scan, intercept, or evaluate every conversation. It never runs on its own from unrelated messages.
• It captures feedback only for a job where you pass --feedback.
• It never modifies your prompt, system config, or any file outside its own feedback store. All improvements are suggestions for you to approve.

What It Does

• Capture: Turn an explicit --feedback response into a score (+1/-1) and a hint
• Analyze: Find recurring hints for a job
• Suggest: Generate actionable improvement suggestions (pending your approval)

Security & Privacy

Read before use.

1. Opt-in only. Feedback is captured only when you run SKILL:self-improvement-omni --job <name> --feedback "..." for that specific job. There is no background/interactive listener.
2. Raw feedback is NOT stored. Your response text is discarded after evaluation. Only the derived score and a sanitised, secret-redacted hint are persisted to memory/learning/agent-feedback.json. Do not assume your exact words are kept — they are not.
3. Redaction. Things that look like tokens, API keys, private-key blocks, long hex/base64 blobs, or user:pass@ URIs are replaced with [REDACTED] before anything is stored or printed.
4. Retention. Stored data is bounded per job: the last 40 evaluations and last 20 suggestions are kept. Entries older than that are pruned.
5. No exfiltration. Aggregated stats and suggestions contain only scores and sanitised hints — never raw user text. Do not forward stored content outside this machine without a full review, because hints are derived from your responses and may preserve fragments of them.
6. No autonomous changes. The script never edits your prompt or behaviour. It only writes a suggestion list. Applying any suggestion is your decision.
7. Consent. Because your responses are retained (in reduced, sanitised form) and reused across time for suggestions, using this skill implies consent to that. If you prefer, stop using --feedback, or delete memory/learning/agent-feedback.json.

Usage

Capture Feedback (explicit)

SKILL:self-improvement-omni --job daily-report --feedback "Great!"
SKILL:self-improvement-omni --job daily-report --feedback "Add more stats"

Get Stats

SKILL:self-improvement-omni --job daily-report --stats

Generate Improvement Suggestions

SKILL:self-improvement-omni --job daily-report --improve

Arguments

Storage

sanitised shape in memory/learning/agent-feedback.json (no raw feedback):

{
  "jobs": {
    "daily-digest": {
      "evaluations": [
        { "date": "2026-03-14", "score": 1, "hint": null },
        { "date": "2026-03-13", "score": -1, "hint": "add weekly star comparison" }
      ],
      "improvements": [
        { "date": "2026-03-14", "suggestion": "Add weekly star delta",
          "from_hint": "add weekly star comparison", "implemented": false,
          "requires_approval": true }
      ]
    }
  }
}

Cron Integration

Daily: Ask for feedback (9:30 AM)

This step prompts the user for feedback about the job — it does not fabricate or auto-capture anything:

{
  "id": "feedback-ask",
  "schedule": "30 9 * * *",
  "message": "How was yesterday's daily-digest? Reply with feedback or 'no change'.",
  "deliver": "origin"
}

Only a reply that includes --feedback against the job is stored. A plain "no change" or unrelated reply must not be persisted.

Daily: Generate Suggestions (10 AM)

Suggestions go back to the origin chat for approval, never silently:

{
  "id": "improvement-suggestions",
  "schedule": "0 10 * * *",
  "message": "SKILL:self-improvement-omni --job daily-digest --improve",
  "deliver": "origin"
}

The output contains only scores and sanitised hints. Review it before acting on any suggestion.

Weekly: Summary (Saturday 9 AM)

{
  "id": "weekly-summary",
  "schedule": "0 9 * * 6",
  "message": "SKILL:self-improvement-omni --improve all --weekly",
  "deliver": "origin"
}

Example Workflow

1. User receives daily digest
2. User replies: "Good! But can you show star trends?"
3. Feedback captured (opt-in): score=1, hint="show star trends" — raw reply not stored
4. Next day, suggestions generated:
   • "Add star trend comparison (last 7 days)"
   • "User satisfaction: 75%"
5. Agent presents suggestions for approval — only you apply them; nothing is auto-updated

Improvement Suggestion Format

Improvement Suggestions - {job}

Stats: 8 evaluations, avg score: 0.75

Top Hints:
1. "add weekly star changes" (2x)
2. "use table format" (1x)

Suggested Actions (pending your approval):
- Add 7-day star delta to GitHub section
- Use table-image-generator for stats

Status: 1 suggestion pending

Manual Evaluation

If you prefer not to use automatic capture, you can evaluate manually:

Evaluate yesterday's output as: good/bad

Related Skills

• openclaw-continuous-learning: Internal session analysis (works alongside this skill)
• compound-engineering: Deep session analysis