ワンクリックで
flywheel-research
Deep research on an external GitHub repository to extract implementation insights.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
メニュー
Deep research on an external GitHub repository to extract implementation insights.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
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.
| name | flywheel-research |
| description | Deep research on an external GitHub repository to extract implementation insights. |
Research an external GitHub repository: $ARGUMENTS
Run a 7-phase research pipeline to extract implementation insights.
Parse: Extract the GitHub URL from $ARGUMENTS.
Clone safety (MANDATORY): the URL MUST be https:// on an allowed Git host (github.com, gitlab.com, bitbucket.org, codeberg.org, git.sr.ht). Reject anything else unless the user explicitly confirms a private / self-hosted host. After cloning, pin the HEAD SHA via git -C <dir> rev-parse HEAD and surface it in the final research doc as Source: <url> @ <short-sha> so readers can reproduce the exact commit you studied. Refuse http:// / git:// / ssh:// unless the user has set FLYWHEEL_ALLOW_INSECURE_CLONE=1.
Setup:
Bootstrap Agent Mail: call macro_start_session(human_key: cwd, program: "claude-code", model: your-model, task_description: "Research: <repo-name>").
Create a team: TeamCreate(team_name: "research-<repo-slug>").
Phase 1 — Investigate: Use Agent(subagent_type: "Explore", name: "research-explore", team_name: "research-<repo-slug>", run_in_background: true) to analyze the repository:
Save the task ID. If agent goes idle without reporting, nudge: SendMessage(to: "research-explore", message: "Please send your findings.").
Phase 2 — Deepen: Use Agent(subagent_type: "general-purpose", name: "research-deep", team_name: "research-<repo-slug>", run_in_background: true) to explore 3 most interesting areas in depth. Agent prompt must include Agent Mail bootstrap (macro_start_session) and instruction to send findings via send_message.
Phase 3 — Inversion: Use Agent(subagent_type: "general-purpose", name: "research-invert", team_name: "research-<repo-slug>", run_in_background: true) to ask: "What does this repo do badly or unconventionally that we should avoid?" Agent prompt must include Agent Mail bootstrap.
Phase 4 — Blunder hunt: Use Agent(subagent_type: "general-purpose", name: "research-blunders", team_name: "research-<repo-slug>", run_in_background: true) to look for known pitfalls, anti-patterns, or design regrets in the codebase. Agent prompt must include Agent Mail bootstrap.
Phases 2–4 can run in parallel after Phase 1 completes. Nudge idle agents individually. Shutdown each individually (NOT broadcast):
SendMessage(to: "research-deep", message: {"type": "shutdown_request", "reason": "Research phase complete."})
SendMessage(to: "research-invert", message: {"type": "shutdown_request", "reason": "Research phase complete."})
SendMessage(to: "research-blunders",message: {"type": "shutdown_request", "reason": "Research phase complete."})
Phase 5 — User review: Present findings to the user. Ask: "Which insights are most relevant to your project? Any areas to explore further?"
Phase 6 — Multi-model synthesis: Spawn 2 parallel agents with run_in_background: true:
Agent(subagent_type: "Plan", model: "opus", name: "research-synth-a", team_name: "research-<repo-slug>", run_in_background: true, prompt: "... Agent Mail bootstrap ... What can we learn from this repo and apply to our codebase? Write findings to docs/research/<repo>-apply.md and send path via send_message.")Agent(subagent_type: "Plan", model: "sonnet", name: "research-synth-b", team_name: "research-<repo-slug>", run_in_background: true, prompt: "... Agent Mail bootstrap ... What ideas from this repo would improve developer ergonomics in our project? Write findings to docs/research/<repo>-ergonomics.md and send path via send_message.")Shutdown each individually after collecting results.
Phase 7 — Synthesis: Combine all findings into a structured research proposal.
Write the proposal to disk first: docs/research-<repo-name>-<date>.md. The proposal's header MUST include a Source: line with the repo URL and the short HEAD SHA recorded in Phase 1 (e.g. Source: https://github.com/foo/bar @ abc1234). Then present key takeaways.
If the user's goal is to integrate a major feature inspired by the researched repo (not just extract insights), extend with these post-research phases:
Phase 8 — Integration proposal: Write a dedicated integration document at docs/research-<repo>-integration.md:
Phase 9 — Iterative deepening: Push past conservative initial suggestions. Spawn an agent to deepen the proposal:
Agent(model: "opus", name: "deepen-integration", prompt: "Read docs/research-<repo>-integration.md. The initial proposal is too conservative. Push further: what ambitious integration would create the most value? What non-obvious synergies exist? Expand the proposal and write back to the same file.")
Phase 10 — 5x blunder hunt: Run 5 sequential blunder-hunt passes on the integration proposal. Each pass reads the proposal and asks: "What will go wrong if we build this? What are we missing? What assumptions are wrong?"
Phase 11 — Cross-model feedback: Make the proposal self-contained (no external references needed), then get feedback from 2-3 different models via parallel agents. Each writes feedback to docs/research-<repo>-feedback-<model>.md.
Phase 12 — Final synthesis: Blend all feedback using the Best-of-All-Worlds approach. Present the final integration proposal to the user with a recommended action:
/start