agentware: Skill for Claude Code

Self-Improvement Skill — How agentware Grows Its Own Skills

When to invoke: at the end of any task or iteration, when you notice you just did something you'll likely do again — or when the post-phase assessment (run by the execution agent) extracts knowledge from a worklog. agentware treats incidental learnings as cheap (just write them) and skill promotion as a small, auditable negotiation with the user.

Why this skill exists

Without an explicit promotion procedure, knowledge dies in three places:

1. In conversation, when the agent says "ah, I figured out X" but never writes it down.
2. In assessment.md, when the post-phase assessment extracts a learning and nothing moves it into the live knowledge base.
3. In worklog.md, where worklogs grow long and the reusable bits get buried.

This skill closes the loop: every learning lands somewhere the next agent can find it, with the right level of authority. Learnings auto-promote, skills ask once, steering always asks.

The decision tree

After completing a task (or reviewing a worklog), classify any new knowledge:

Did I learn something?
├── No  → continue to next task
└── Yes → Is it project-specific?
         ├── Yes → write learning to the EXTERNAL dir (auto, no permission)
         └── No  → Is it a reusable procedure (≥2 steps, applies to many tasks)?
                  ├── Yes → write skill to the EXTERNAL dir (auto, just inform the user)
                  └── No  → Is it an always-true rule for the PACKAGE itself?
                           ├── Yes → package/steering change → EXPLICIT request + !! WARNING !!
                           └── No  → write learning to the EXTERNAL dir (auto)

Writing to the user's external dir is always safe and needs no permission — it's
their own space and never touches the orchestrator. Only changing the PACKAGE
(steering/skills/loop) requires an explicit request and the !! WARNING !!.

Examples

Discovery	Type	Where it goes	Permission
"This app's dev server runs on port 4173"	Project-specific fact	learnings/<project>-setup.md	None (auto)
"To debug a hydration mismatch in Next.js: do these 5 steps"	Reusable procedure	<knowledge-dir>/skills/debug-hydration/SKILL.md (external)	Ask user once
"Always run npm ci rather than npm install on CI"	Always-true rule	AGENTS.md (package — self-extension)	Explicit + !! WARNING !!
"Stuck Docker container? docker rm -f X then retry"	One-off fix	learnings/docker-gotchas.md	None (auto)