RonanCodes/skills/capability-retro

Capability Retro

Close the self-improvement loop: each session that hand-rolled something painful is a signal a tool should exist. This skill spots those signals and turns the good ones into new skills/agents (or edits to existing ones) before context clears.

The split with /ro:skill-retro: skill-retro sharpens skills you used this session. capability-retro spots capabilities you lacked (create-new) and is the step that feeds the handoff template's "Skills / subagents this session wished it had".

When it runs

• As a near-final step inside /ro:handoff (before writing the file) and a check inside /ro:close-session — so the loop closes every time you wrap a session.
• On demand whenever the user asks "should this be a skill/agent?".

Process

1. Gather signals from THIS session (the main-context advantage)

Only the live thread can see these — a fresh subagent can't. Skim the conversation + this session's git diffs for:

• Manual multi-step dances — a sequence you executed by hand (especially if repeated, or laborious) that has a stable shape → candidate skill.
• "I wish I had…" moments — anywhere you or the user said a tool should exist, or worked around its absence.
• Heavy context-filling work — large reads/searches/authoring the main thread did inline that bloated context → candidate agent (isolated context).
• Repeatable workflows the user described — "if I ever X, then do Y", "whenever I get a Z" → candidate skill (+ usually a memory trigger).
• Friction on existing skills/agents — a missing step, wrong default, ambiguous flag → an edit (hand to /ro:skill-retro).

2. Dedup against the live registry (never propose what exists)

ls ~/Dev/ronan-skills/skills ~/Dev/ronan-private-skills/skills 2>/dev/null   # skills
# agents: project .claude/agents + ~/.claude/agents + the agent registry
grep -ril "<keyword>" ~/Dev/ronan-skills/skills 2>/dev/null | head

For each candidate, find the closest existing skill/agent. If one already covers it, the outcome is already-exists (point at it) or edit-existing, not create-new.

3. Decide skill vs agent (for genuine create-new candidates)

• Skill — a procedure invoked by the main agent or the user via /. The work needs the live context, or it's a reusable workflow/checklist. Most candidates.
• Agent — an isolated-context worker for heavy read/search/authoring that should stay OUT of the main context (e.g. a wiki-curator, an explorer). Pick this when the value is "do a lot of reading/writing and return only the conclusion".
• A skill may dispatch agents; the two compose.

4. Triage + present

Classify each candidate: create-skill / create-agent / edit-existing / already-exists / not-worth-it (one-tool-use things aren't worth a skill). Then AskUserQuestion with the create/edit candidates — let the user pick which to act on now. Show the proposed name + one-line description + why for each.

5. Execute the approved ones

• New skill → /ro:create-skill (or dispatch a general-purpose Agent to author the SKILL.md against the create-skill template). Keep authoring in the agent so the main context stays lean.
• New agent → write the agent definition (.claude/agents/<name>.md or ~/.claude/agents/), following existing agent conventions.
• Edit existing → /ro:skill-retro --skill <name> --apply or a direct Edit.
• Commit per the repo's flow (feature branch → PR → squash-merge; skills repos push so the plugin cache re-syncs). Honour weekday commit-timestamp rules.

6. Record what you didn't build

Anything approved-but-deferred or not-worth-it-now → the handoff's "Skills / subagents this session wished it had" section, or a ready-for-agent issue on ronan-skills. So the gap survives even if it wasn't built this session.

Output (when run standalone)

CAPABILITY RETRO
  signals found:    <N>
  → create skill:   <name> — <why>
  → create agent:   <name> — <why>
  → edit existing:  <skill> — <what>
  → already covered: <candidate> → use <existing>
  → not worth it:   <candidate> (one-off)
  acted on:         <N>  |  deferred → <issue/handoff>

Flags

• --apply — also create/edit the approved candidates (otherwise just propose).
• --no-agents — only consider skills, skip agent candidates.

Related

• [[skill-retro]] — sharpen skills you USED (this is the create-new counterpart).
• [[create-skill]] — the authoring step for new skills.
• [[handoff]] / [[close-session]] — the session-end skills that invoke this.