goatcitadel-native-safe-self-improvement

Files (20)

• SKILL.md
• INSTALL.md · docs/
• REVIEW_CHECKLIST.md · docs/
• SECURITY_POLICY.md · docs/
• TARGET_SURFACES.md · docs/
• WORKFLOW.md · docs/
• proposal-agents-rule.example.md · examples/
• proposal-playbook.example.md · examples/
• proposal-soul-rule.example.md · examples/
• proposal-tools-rule.example.md · examples/
• HOT_MEMORY.md · templates/
• corrections.md · templates/
• index.md · templates/
• ERRORS.md · templates/logs/
• EVAL_IDEAS.md · templates/logs/
• FEATURE_REQUESTS.md · templates/logs/
• LEARNINGS.md · templates/logs/
• ROUTING_GAPS.md · templates/logs/
• reflections.md · templates/
• target-map.md · templates/

SKILL.md (first lines)

# GoatCitadel Native Safe Improvement

Help GoatCitadel get better over time **without** granting autonomous self-rewrite. All changes go through human review.

This skill combines two patterns:

1. **Bounded reflection memory** - store corrections, preferences, and post-task reflections inside `.goatcitadel/self-improvement/`
2. **Structured operational learnings** - log errors, gaps, and wins, then convert recurring patterns into **reviewable proposals**

## Primary Goal

Create a GoatCitadel-native improvement loop that can:

- remember explicit user corrections safely
- avoid repeating known workflow mistakes
- spot routing, tool, prompt, and playbook gaps
- suggest durable improvements to GoatCitadel guidance files
- keep a roadmap of missing capabilities
- **never** silently mutate core agent files, project code, provider configs, or secrets

## Workspace Scope

All skill-owned data must live under:

```text
.goatcitadel/self-improvement/
```

If the folder does not exist, create it from this bundle's templates.

## GoatCitadel Target Model

This skill is aware of common GoatCitadel guidance surfaces, but it may only **propose** changes to them.

### Proposal Targets
- `AGENTS.md` for durable operating rules and coordination rules
- `TOOLS.md` for tool-use constraints, routing notes, and edge-case gotchas
- `SOUL.md` for stable assistant behavior, tone, and identity guardrails
- `PLAYBOOKS/*.md` for repeatable workflows and runbooks
- `MEMORY.md` for explicit durable preferences only
- `ROUTING.md` for model and capability routing logic
- `EVALS.md` for test cases and regression ideas

Current repo guarantee in this pass:

- `AGENTS.md`
- `docs/*` style destinations

If a target surface does not exist yet, keep it as a conceptual proposal target. Draft the proposal under `.goatcitadel/self-improvement/proposals/pending/` and do not create new root governance files automatically.

### Hard No-Touch Files
This skill must never directly edit:
- `SKILL.md`
- `AGENTS.md`
- `TOOLS.md`
- `SOUL.md`
- `MEMORY.md`
- `ROUTING.md`
- `EVALS.md`
- `CLAUDE.md`
- `.env`
- provider configs
- secret stores
- Docker, Traefik, or infrastructure files
- application source code
- tests
- CI workflows

It may only draft proposals for those surfaces.

## What This Skill May Do

This skill may only:
- read files inside `.goatcitadel/self-improvement/`
- write files inside `.goatcitadel/self-improvement/`
- append structured entries to local logs
- summarize recurrence patterns
- draft proposal files for human review
- generate eval ideas and playbook suggestions as proposals
- propose concise memory candidates for explicit durable preferences

## What This Skill Must Never Do

This skill must never:
- make network requests
- access browser, email, calendar, contacts, or external accounts
- read files outside `.goatcitadel/self-improvement/`
- infer preferences from silence or weak hints
- self-edit this skill
- directly apply any proposal
- modify source code, tests, prompts, configs, or infrastructure
- escalate its own permissions
- request secrets unless the user is explicitly configuring something in the current task

## When To Use

Use this skill when one or more of these happens:

- the user explicitly corrects GoatCitadel
- a tool call fails or causes friction
- a model routing choice was weak
- a workflow succeeded unusually well and should be repeatable
- a missing capability becomes obvious
- a playbook step was missing or unclear
- GoatCitadel finishes a substantial task and should self-review
- a preference is stated clearly and appears durable
- an eval or regression test case should be added later

## Folder Layout

```text
.goatcitadel/self-improvement/
├── HOT_MEMORY.md
├── corrections.md
├── reflections.md
├── index.md
├── target-map.md
├── logs/
│   ├── LEARNINGS.md

From goatcitadel/GoatCitadel, Apache-2.0.

More from goatcitadel/GoatCitadel