Menú
Refine a specific agent skill using session evidence and feedback.
Instalar con Codex o Claude Copia este prompt, pégalo en Codex, Claude u otro asistente, y deja que revise la página de la skill y la instale por ti.
Relentless goal/design interview that sharpens framing and writes durable docs (brainstorm artifact, ADRs, glossary) as decisions crystallize. Use when refining a flywheel goal, pressure-testing scope, or "grill with docs".
Start or resume the full agentic coding flywheel. Drives the complete workflow: scan → discover → plan → implement → review.
Set up flywheel prerequisites for this project.
One-shot diagnostic of every flywheel dependency — MCP connectivity, Agent Mail liveness, br/bv/ntm/cm binaries, node version, git status, dist-drift, orphaned worktrees, and checkpoint validity. Use when debugging toolchain issues, before starting a new session, after /flywheel-cleanup, or as a CI gate.
Strategic gap analysis between vision (AGENTS.md / README.md / plan docs) and what's actually implemented. Converts gaps into beads, optionally launches a swarm. Use when "reality check", "where are we really", "gap analysis", "did we drift", or before declaring a long-running project done.
Launch a parallel swarm of agents to implement multiple beads simultaneously.
Basado en la clasificación ocupacional SOC
| name | flywheel-refine-skill |
| description | Refine a specific agent skill using session evidence and feedback. |
Refine a specific skill: $ARGUMENTS (skill name required)
1b. Check for recent proposal cache:
- Look for docs/skill-refine-<name>-proposed.md.
- If it exists and was modified within the last 7 days:
- Output: "Recent proposal found (modified ). Showing cached analysis:"
- Display the file contents.
- Ask: "Re-analyze with fresh evidence, or apply these changes?"
- If user says "apply", jump to step 6 (apply changes).
- If user says "re-analyze", continue to step 2.
- If not found or older than 7 days, continue to step 2.
Read the current skill from skills/<name>/SKILL.md.
Triage gate — check for evidence before spawning an agent:
a. Search agent-mail for feedback: search_messages with query: "<skill-name> feedback".
b. Search agent-mail for usage: search_messages with query: "<skill-name>".
c. Check if any .beads/ entries reference this skill name.
d. If ALL searches return empty AND no session evidence was provided in $ARGUMENTS:
<name> — skipping (nothing to improve)."$EVIDENCE_SUMMARY (max 2000 chars, newest first).Proceed to analysis only if triage gate passed.
Use Agent(subagent_type: "general-purpose", run_in_background: true, name: "skill-refine", team_name: "refine-skill-<name>") to analyze.
Agent prompt MUST include:
$EVIDENCE_SUMMARY collected in step 3 (verbatim feedback quotes, failure logs, success patterns)Agent prompt must also include Agent Mail bootstrap (macro_start_session) and instruction to write proposed changes to docs/skill-refine-<name>-proposed.md, then send the file path via send_message.
Save the task ID. Nudge if idle: SendMessage(to: "skill-refine", message: "Please send your proposed changes.").
Shutdown when done: SendMessage(to: "skill-refine", message: {"type": "shutdown_request", "reason": "Analysis complete."}).
Read the proposed changes file and show the user a diff:
BEFORE: <current text>
AFTER: <proposed text>
Ask: "Apply these changes?" If yes, write the updated SKILL.md.
Confirm: "Skill <name> updated."