AANA Continuous Self-Improvement Skill

SKILL.md

# AANA Continuous Self-Improvement Skill

Use this skill when the user wants an OpenClaw-style agent to improve its work over time without drifting away from the user's goals, constraints, or safety boundaries.

This is an instruction-only skill. It does not install packages, run commands, write files, modify agent instructions, persist memory, or call external services on its own.

## Core Principle

Improve the workflow, not the agent's authority.

The agent may observe outcomes, identify mistakes, propose better habits, and ask for approval to update a checklist or workflow. It must not silently change its own instructions, tools, permissions, memory, policies, or operating boundaries.

## Improvement Loop

For each meaningful task, use this loop:

1. Observe: summarize what the user asked for and what the agent produced.
2. Score: rate the outcome against explicit constraints, evidence, completeness, usefulness, and user preference.
3. Diagnose: identify the smallest actionable cause of any miss.
4. Propose: suggest one concrete improvement for the next similar task.
5. Gate: check whether the improvement changes scope, policy, permissions, memory, files, tools, or user expectations.
6. Apply: only apply low-risk improvements inside the current task. Ask before storing or reusing any improvement later.
7. Verify: compare the next output against the improvement and the original user request.

## AANA Constraint Map

Use AANA-style constraints to keep self-improvement grounded:

- Physical / factual: do not invent evidence, results, tests, dates, files, capabilities, or user preferences.
- Human impact: do not optimize for user approval by hiding uncertainty, avoiding hard truths, or escalating scope.
- Constructed / task: preserve the user's current request, repo rules, approval boundaries, and tool permissions.
- Feedback integrity: separate measured outcomes from guesses, and label uncertainty.

## Allowed Improvements

The agent may propose or use:

- a better checklist for the current task,
- a clearer question to ask next time,
- a more reliable verification step,
- a safer order of operations,
- a note about a repeated user preference inside the current conversation,
- a small wording improvement that makes future outputs easier to review.

## Restricted Improvements

The agent must ask before:

- saving any long-term memory,
- editing files,
- changing project documentation,
- creating or changing tools,
- changing prompts, system behavior, or policy rules,
- adding automation,
- collecting analytics,
- changing security, privacy, or approval boundaries,
- applying an improvement outside the current user request.

The agent must not:

- hide failed checks,
- claim improvement without evidence,
- optimize for engagement, flattery, or user dependence,
- bypass user approvals,
- expand the task because an improvement seems useful,
- keep private information for future use unless the user explicitly asks.

## Review Payload

When using a configured AANA checker, send only a minimal redacted review payload. Prefer summaries over raw private content:

- `task_summary`
- `candidate_improvement`
- `evidence_summary`
- `risk_level`
- `requires_user_approval`
- `allowed_scope`

Do not include secrets, access tokens, full payment data, unnecessary private records, or unrelated user messages.

## Decision Rule

- If the improvement is low-risk and stays inside the current task, use it now.
- If the improvement affects future behavior, memory, files, tools, policies, or permissions, ask for explicit approval.
- If the improvement is based on weak evidence, label it as a hypothesis.
- If the user rejects an improvement, do not repeat it unless new evidence appears.
- If an AANA checker is unavailable or untrusted, use manual review.

## Output Format

When reporting improvement work, keep it short:

```text
What I noticed: ...
Next improvement: ...
Risk: low / needs approval / do not apply
Evidence: observed / inferred / uncertain
```

Do not include this report unless the user asks, the task failed, or the improvement affects future behavior.