scbe-self-improvement-skill-management

# SCBE Self-Improvement & Skill Management

Use this skill to keep SCBE capabilities compounding through a disciplined loop:
build something -> document it -> extract patterns -> create/update skill -> reuse it in the next build.

## Core Rules

1. Enforce explicit approval before installing any new skill.
2. Keep `SKILL.md` concise (target under 500 lines); move depth into `references/`.
3. Prefer overtriggering to undertriggering in skill descriptions.
4. Run dimensional analysis before changing canonical math constants or formulas.
5. Keep stable naming: `scbe-{domain}-{function}`.

## Workflow Decision

Choose one path:

- New skill creation
- Existing skill update
- Knowledge propagation to another AI system

If the user intent is mixed, split output by path and make each path explicit.

## New Skill Creation

Create a new skill when one of these holds:

- Workflow repeated 3+ times across conversations
- New SCBE component stabilized (tests passing, spec written)
- User explicitly asks for a new skill
- Cross-service integration pattern validated

Procedure:

1. Name skill as `scbe-{domain}-{function}`.
2. Initialize with `init_skill.py`.
3. Write trigger-rich description (what it does + when to invoke).
4. Keep core workflow in `SKILL.md`; move deep material to `references/`.
5. Add templates/assets needed for reuse.
6. Validate with `quick_validate.py`.
7. Test with 2-3 realistic prompts.
8. Present approval gate summary and wait for explicit approval before install.

## Existing Skill Update

Update when:

- Canonical constant changes (rare; dimensional analysis required first)
- New layer/component added
- Workflow step changed
- Glossary expanded
- Instruction bug discovered

Procedure:

1. Read current skill fully.
2. Identify exact sections to change.
3. Draft minimal diff.
4. Run dimensional analysis if math/constants are touched.
5. Present diff for user approval.
6. Apply and validate.

## Approval Gate (Mandatory)

Never install a new or updated skill without explicit user approval.

Before installation, present:

1. What the skill does (one sentence)
2. Trigger phrases/contexts
3. Files included
4. New dependencies (if any)

Wait for: `yes`, `approved`, or equivalent.

## Knowledge Propagation

When user asks to teach/update another AI:

1. Identify target AI and role.
2. Scope relevant layers/components/constants.
3. Generate transfer docs using templates in `assets/`.
4. Keep constants exact unless user explicitly asks for revision.
5. Optionally prepare Hugging Face dataset packaging.

Use:

- `assets/context-template.md`
- `assets/scbe-knowledge-v4.yaml`
- `assets/action-summary.template.yaml`
- `references/knowledge-propagation.md`
- `references/external-identity-research.md` (for handle/name/repo disambiguation; treat as verification-first input)

## Skill Composition

Prefer reuse/composition over creating redundant skills.

Contract:
`Skill A output -> action_summary.yaml -> Skill B input`

Use `assets/action-summary.template.yaml` for the handoff schema.

## Memory Guidance

Persist only stable, validated information and only with user approval.

Persist:
- Canonical terms
- Workflow preferences
- Naming conventions
- Milestones

Do not persist:
- Temporary debugging state
- One-off experiments
- Unapproved drafts