| name | self-improvement |
| description | Review completed work, analyze mistakes, identify repeatable improvements, and turn validated lessons into memory or reusable skills. Use after substantial tasks, failures, or review-heavy work. |
| always | false |
| enabled | true |
Self-Improvement
Use this skill after meaningful work, not after trivial replies.
Primary Goals
- Review what was just done
- Identify concrete weaknesses or risks
- Extract reusable tactics
- Persist only the learnings that deserve to survive
When To Use
- after completing a non-trivial coding or debugging task
- after a failure, retry loop, or wrong assumption
- after a review that uncovered repeatable issues
- when a workflow feels reusable enough to become a skill
Skip it for one-off small answers with no reusable pattern.
How It Fits The Current Memory Framework
Use the current Horbot memory layout, not ad-hoc logs:
- durable facts and lasting rules go to
L2/MEMORY.md
- recent outcomes and concise summaries go to
L1/HISTORY.md
- reusable tactics and corrected assumptions go to
L1/REFLECTION.md
If the completed work forms a repeatable workflow, Horbot may also distill it into a user skill under the current agent's skills directory.
Recommended Workflow
1. Review
Use the templates in templates/ when helpful:
templates/code-review.md
templates/capability-assessment.md
templates/learning-plan.md
Focus on concrete evidence from the work that just happened.
2. Classify The Learning
- durable project truth ->
MEMORY.md
- recent progress or outcome ->
HISTORY.md
- reusable tactic / lesson learned ->
REFLECTION.md
- repeatable operating procedure -> candidate skill
3. Validate Before Persisting
Only persist learnings that are:
- specific
- evidence-backed
- reusable
- safe to keep
Do not preserve vague self-praise, unverified guesses, or secrets.
Relation To Automatic Skill Distillation
Horbot can run a background skill review after successful tool-backed work.
That mechanism works best when:
- the task used real tools
- the result contains a concrete procedure or checklist
- the final output clearly describes the steps that worked
When the workflow is reusable, the system may:
- create or update a skill
- record the reusable tactic in reflection/history
- make that workflow available for later sessions
Guardrails
- prefer small, high-signal learnings over long reports
- do not create a new skill for project-specific or one-off work
- avoid duplicating existing skills when an update is enough
- verify important improvements before treating them as durable knowledge
- keep backward compatibility and user intent in mind