gagharutyunyan1993/methodology-toolkit

# methodology-toolkit [](https://github.com/gagharutyunyan1993/methodology-toolkit/actions/workflows/validate-manifests.yml) [](#install) [](LICENSE) A Claude Code plugin that makes Claude reason through real analytical frameworks instead of intuition. It bundles three interlocking pieces over one shared index of **29 methodologies** (Cynefin, First Principles, OODA, ACH, Pre-mortem, Red Team, BATNA, JTBD, Theory of Constraints, Minto, SWOT/TOWS, and more). **Created by [Gagik Harutyunyan](mailto:gagik.harutyunyan17@gmail.com).** If the ideas here are useful to you, keep the attribution — that's the whole deal (MIT). See [`CITATION.cff`](CITATION.cff). For the longer rationale, read: [Claude Does Not Need More Prompts. It Needs Reasoning Discipline.](docs/why-methodology-toolkit.md) ## What's inside | Component | Type | What it does | | --- | --- | --- | | `methodology-driven-thinking` | **skill** | Auto-triggers on non-trivial tasks. Classifies the problem (Cynefin), selects 1–3 fitting methods, applies each by its steps, synthesizes an actionable answer. | | `/methodology-toolkit:method` | **command** | Manual trigger. `/...:method ` runs the full protocol; `/...:method ACH+pre-mortem ` forces specific methods. | | `red-team-critic` | **subagent** | Independent adversarial pass. Returns critique only — load-bearing assumptions, failure modes, exploits, disconfirming evidence — verified against primary sources. | | `methodology.md` | shared index | The authoritative library of all 29 methods (`use_when` / `avoid_when` / `steps` / `output`). All three read from it. | ### How they fit together Non-trivial task ├─ auto → methodology-driven-thinking skill ─┐ └─ manual → /methodology-toolkit:method ───────┤──→ read methodology.md │ double-pass (stress-test the first answer): /methodology-toolkit:method recheck (inline, by yourself) or red-team-critic (subagent, independent context) ## Install /plugin marketplace add gagharutyunyan1993/methodology-toolkit /plugin install methodology-toolkit@methodology-toolkit To try it locally before publishing, point `--plugin-dir` at the **plugin subdirectory** (not the marketplace root): claude --plugin-dir /path/to/methodology-toolkit/plugins/methodology-toolkit ## Usage /methodology-toolkit:method how should we prioritize the Q3 backlog? /methodology-toolkit:method ACH+pre-mortem should we migrate polling to WebSocket? /methodology-toolkit:method red-team Or just ask a hard question — the skill triggers itself. For an independent teardown of a plan, invoke the `red-team-critic` subagent. ## Examples - [Polling vs SSE vs WebSocket](examples/websocket-decision.md) — architecture decision under uncertainty. - [Codebase diagnosis without guessing](examples/codebase-diagnosis.md) — bug-fix workflow grounded in primary evidence. For external launch copy, see [Launch Posts](docs/launch-posts.md). ## Design philosophy Two failure modes this plugin is built to prevent: 1. **Methodology theater** — naming a method without walking through its steps. Every component refuses to do this: the output is the *reasoning the method produces*, never its label. 2. **Confident wrongness** — asserting load-bearing claims from memory. The rule is *verify before you assert*: read the file, run the grep, check the test — primary sources outrank docs and comments. ## Why this is not just a prompt collection The goal is not to make Claude _sound_ more strategic. The goal is to make it harder for Claude to skip the parts of analysis where mistakes usually hide: - **Method selection is gated by problem type.** Cynefin runs first so a simple syntax question does not get buried under frameworks, while uncertain or adversarial decisions get methods that match the situation. - **Every invoked method must be applied step by step.** If the answer says "using ACH" or "using OODA", the output must show the ACH matrix or the Observe/Orient/Decide/Act pass. Labels alone are treated as failure. - **Load-bearing facts must be verified before conclusion.** The toolkit pushes Claude toward primary evidence: code read, commands run, tests checked, git history inspected. - **The critique path is separate.** `red-team-critic` runs as an adversarial reviewer with its own context, so the stress test is less likely to rationalize the first answer. ### Example: architecture decision **Without the toolkit:** Claude may recommend "migrate polling to WebSocket" because it sounds technically cleaner, then list generic pros and cons. **With the toolkit:** Claude classifies the decision as complicated/complex, applies First Principles to identify what actually needs real-time delivery, uses ACH to compare polling, SSE, and WebSocket against disconfirming evidence, then runs a pre-mortem. The likely outcome is a narrower recommendation: WebSocket only if bidirectional low-latency updates are required; otherwise SSE or tuned polling may be cheaper and safer. ### Example: codebase diagnosis **Without the toolkit:** Claude may infer the root cause from one symptom and patch the nearest visible file. ## The 29 methods Cynefin · OODA · PDCA · First Principles · 5 Whys · 5 Forces (Porter) · ADKAR · Kotter · Eisenhower · GTD · SMEAC · Schwerpunkt · Mental Models · System 1/2 · JTBD · Double Diamond · Theory of Constraints · OKR · Minto Pyramid · SCQA · BATNA · Thomas-Kilmann · ACH · Red Team · Pre-mortem · PMESII · SWOT/TOWS · SAT · Quality of Information Check. Full definitions in [`plugins/methodology-toolkit/skills/methodology-driven-thinking/methodology.md`](plugins/methodology-toolkit/skills/methodology-driven-thinking/methodology.md). ## License [MIT](LICENSE) © 2026 Gagik Harutyunyan. Use it, fork it, build on it — just keep the attribution.

标签：自定义脚本