| name | crow-pet-self-improvement |
| description | Convert corrections, approvals, failures, and task patterns into memory, examples, rubrics, and policy proposals. |
crow.pet Self-Improvement
Use this skill to make crow.pet need less human interaction over time.
Workflow
- Read the task record and evaluation.
- Identify correction, pattern, discovery, or approval trend.
- Read execution trace evidence when present.
- Reflect on one reusable lesson.
- Curate the smallest context delta, then propose a delta update only when a file should change.
- Keep large context in raw sources, providers, or capsules.
- Propose skill, rubric, contract, playbook, memory, agent profile, or policy changes when evidence exists.
- Detect recurring ad-hoc task shapes in Ledger evidence. At three or more recurrences, propose compiling the shape into a workflow definition.
- Propose retiring a compiled workflow at two failures in the last five runs. Staleness also triggers retirement: no run inside the 30-day trust decay window.
- Route compilation and retirement proposals through Loop curation as delta updates; never auto-apply them.
- Ask before lowering approval levels, applying higher-risk deltas, or promoting persistent agent profiles.
- Record before and after behavior notes.
Progressive Disclosure
Load workflows/self-improve.md when processing a completed task, correction, failure, or /learn note.
Load workflows/playbook-evolution.md when a task produces a reusable lesson or proposed delta update.
Load only the matching file under memory/ for the learning type.
Load ../../playbooks/ only for the target domain named by the delta.
Load ../../templates/delta-update.yaml when proposing memory, playbook, skill, rubric, contract, or policy changes.
Load ../../templates/trust-state.yaml, ../../templates/digest-report.md, and ../../templates/workflow-definition.md when proposing workflow compilation, retirement, or trust changes.
Load ../../templates/execution-trace.yaml, ../../templates/context-delta.yaml, ../../templates/reflection-note.md, and ../../templates/curation-decision.md when using Loop trace, delta, reflection, and curation artifacts.
Load ../../templates/memory-item.yaml, ../../adapters/memory/ledger.md, and ../../adapters/patterns/loop.md when Ledger evidence may become a Memory Garden note.
Load ../../templates/agent-profile.md only when proposing a persistent role-agent promotion.
Load templates/evaluation-report.md only when writing or checking an evaluation report.
Load contract.md and rubric.md only for acceptance or evaluation.
Dependency Edges
promotes_to: memory item, context delta, playbook delta, skill update, rubric update, contract update, policy proposal, workflow compilation proposal, workflow retirement proposal, agent profile, or Memory Garden note.loads_context: context capsule when task history or owner feedback affects the learning.evaluates_with: contract.md, rubric.md.
Rules
- Do not remove safety gates.
- Do not rewrite history.
- Do not rewrite whole files by default.
- Keep memory specific.
- Keep deltas small and evidence-backed.
- Keep always-on memory bounded.
- Link improvements to task IDs.
- Promote agent profiles only when continuity or repeated evidence justifies persistence.
- Retirement proposals are separate from automatic demotion; both may fire on the same trigger.
