{"id":"8df0e6bd-e07a-4fdf-9d79-817de8eb4233","slug":"thinking-toolkit","kind":"skill","name":"thinking-toolkit","summary":"30 thinking models as one portable agent skill — Claude Code, Codex CLI, OpenClaw, and any SKILL.md runtime","owner":"ponomr","repoUrl":"https://github.com/ponomr/thinking-toolkit","license":"MIT","licenseTier":"permissive","trust":"community","stars":14,"match":null,"fewStars":false,"format":"skill","createdAt":null,"addedAt":"2026-10-14T06:08:43.646Z","logoUrl":"https://avatars.githubusercontent.com/u/220024070?v=4","linkOnly":false,"collection":null,"skillPath":null,"categories":["developer-tools"],"tags":["claude-code","codex","decision-making","mental-models","openclaw","thinking-models"],"runtime":"binary","local":true,"cloudCapable":false,"cloudCapability":"unknown","cloudReadyOnPlatform":false,"needsApiKey":false,"reviewed":false,"syncedAt":"2026-10-14T07:12:14.132Z","sample":false,"url":"https://conduits.dev/tools/thinking-toolkit","language":"Python","transports":["stdio"],"install":{},"requirements":{"env":[],"upstreamMayCharge":false},"tools":null,"readme":"![Thinking Toolkit — best battle-tested thinking models, one field manual for agents](assets/banner.png)\n\n# 🧠 Thinking Toolkit\n\n**English** · [Русский](README.ru.md) · [中文](README.zh.md)\n\n**31 battle-tested thinking models, packaged as one portable skill for any AI agent.**\n\nThinking Toolkit gives an agent explicit, reusable thinking procedures. Invoke\nthe toolkit, `/logic`, or a named model when you want a structured method; it\nthen applies the smallest useful model set and returns a concrete artifact such\nas a weighted matrix, causal map, premortem table, backward plan, or draft.\n\nIt is plain Markdown with zero runtime dependencies. If your agent can read\nfiles, it can use this skill.\n\n## Why\n\nLeft to itself, an LLM agent reasons in prose that sounds convincing and skips\nsteps. Meanwhile engineering, strategy, intelligence analysis, and\norganizational learning have spent decades turning the situations agents face\ndaily — comparing options, finding root causes, estimating without data,\nstress-testing plans — into named, tested procedures. This skill turns those\nprocedures into instructions an agent can execute, with the pitfalls\npractitioners documented along the way.\n\nThree commitments shape the design:\n\n1. **Selection.** Knowing what a decision matrix is costs nothing; the\n   expensive mistake is reaching for one at the wrong moment. The catalog\n   encodes selection cues, contrast rules (\"Five Whys for one causal chain,\n   Ishikawa for breadth, Connection Circles for feedback\"), and anti-patterns\n   for every model.\n2. **Procedure.** Each model ships as a full operational card: inputs,\n   numbered steps, guiding questions, output format, a worked example, and\n   documented pitfalls. The agent follows the card instead of improvising\n   around a label.\n3. **Honesty.** The core contract separates observed facts from assumptions,\n   marks unknowns instead of inventing evidence, runs sensitivity checks on\n   consequential results, and ends every artifact with an action and a review\n   trigger.\n\n## How It Works\n\nWhen a model is named explicitly, the agent loads that card and follows it.\nWhen the toolkit is explicitly requested but the method is left open, it\nidentifies the job, routes through the catalog's selection cues and contrast\nrules, and picks one primary model with at most two complements. Ordinary\nplanning, analysis, problem-solving, and writing requests do not activate the\ntoolkit automatically. Once invoked, it frames the situation, applies the\ncard's procedure at an appropriate depth, tests material assumptions, and\nreturns the requested artifact or action in a format suited to the request.\n\nThe skill uses **progressive disclosure** to stay context-efficient: the agent\nfirst reads the concise `SKILL.md`, consults the routing catalog when selection\nis authorized or an explicit alias must be resolved, and loads *only* the model\ncards it actually needs.\nThirty-one models cost nothing until one is used.\n\n## The Catalog\n\n### Decision making — 12 models\n\n| Model | Use when |\n|---|---|\n| Six Thinking Hats | A choice needs balanced perspectives instead of debate |\n| Eisenhower Matrix | Tasks differ by urgency and importance |\n| Second-Order Thinking | Immediate benefits may hide later effects |\n| Decision Matrix | Several options must be compared across weighted criteria |\n| Impact-Effort Matrix | A backlog needs rough portfolio prioritization |\n| Ladder of Inference | A conclusion may have outrun its evidence |\n| Hard Choice Model | The right amount of decision effort is unclear |\n| OODA Loop | Conditions change while action is underway |\n| Cynefin Framework | The response must fit the situation's causal nature |\n| Confidence → Speed vs. Quality | Product work must trade speed against polish |\n| Pareto Analysis | A few contributors may drive most of a measured effect |\n| Backcasting | A distant goal needs a path from the endpoint to today |\n\n### Problem solving — 12 models\n\n| Model | Use when |\n|---|---|\n| Ishikawa Diagram | A defined effect has many possible causes |\n| Five Whys | One incident needs its causal chain traced to a process fix |\n| After Action Review | A completed episode should become a lesson and a next action |\n| Abstraction Laddering | The problem statement may be too narrow or vague |\n| Conflict Resolution Diagram | Opposing demands appear mutually exclusive |\n| Zwicky Box | A solution can be assembled from independent dimensions |\n| Productive Thinking Model | A defined problem needs a complete creative process |\n| Inversion | Failure modes reveal the path to success |\n| Red Teaming | A confident plan needs an independent adversarial challenge |\n| Issue Trees | A large problem needs non-overlapping decomposition |\n| First Principles | Conventions and analogies constrain solution quality |\n| Fermi Estimation | A quantity must be estimated without direct data |\n\n### Systems thinking — 5 models\n\n| Model | Use when |\n|---|---|\n| Iceberg Model | Repeated events point to deeper structures and beliefs |\n| Connection Circles | Variables and feedback relationships need mapping |\n| Concept Map | Concepts and their semantic relationships need clarity |\n| Balancing Feedback Loop | A system resists change or seeks a target |\n| Reinforcing Feedback Loop | Growth or decline feeds further growth or decline |\n\n### Communication — 2 models\n\n| Model | Use when |\n|---|---|\n| Situation-Behavior-Impact | Feedback must be specific and non-judgmental |\n| Minto Pyramid | A busy audience needs the conclusion first |\n\nBeyond single models, the catalog ships **combination recipes** — proven\nsequences such as *Consequential choice* (Hard Choice Model → Decision Matrix →\nRed Teaming) or *Root-cause investigation* (Pareto → Iceberg → Ishikawa → Five\nWhys → OODA) — plus guardrails against the classic failure of stacking models\nto appear thorough.\n\n## Logic analysis (`/logic`)\n\nThe 31 cards help you *choose how to think*. `/logic` does the opposite job: it\n*audits reasoning that already exists* — a claim, an argument, or a draft — and\nreturns a verdict on its validity.\n\nIt ships **discipline, not a textbook**. The model already knows what modus\ntollens is; what `/logic` adds is a fixed procedure (reconstruct → test validity\n→ separate form from truth → verdict), a consistent fallacy taxonomy with stable\nEnglish + Latin names, and a repeatable verdict format. Three modes, routed by\nphrasing, in any language:\n\n- **review** (default) — diagnose and deliver a verdict; no rewrite.\n- **fix** — repair the reasoning with minimal intervention, preserving voice.\n- **solve** — work a textbook task (validate a syllogism, build a truth table,\n  apply a Mill's method, reconstruct an enthymeme).\n\n> *\"Check the logic: we shipped the redesign and signups rose, so the redesign\n> worked.\"*\n> → **1 critical error.** *Post hoc ergo propter hoc* — a marketing campaign\n> launched the same week is an unruled-out cause. Formally invalid; the premise\n> may still be true, but the argument doesn't establish it.\n\nFour lean reference files back it: [overview](logic/overview.md) (procedure +\nverdict format), [fallacy taxonomy](logic/fallacies.md),\n[formal validity](logic/formal-validity.md) (syllogism + propositional), and\n[induction](logic/induction.md) (Mill's methods, analogy, hypothesis).\n\n## Installation\n\nThe skill follows the [Agent Skills](https://simonwillison.net/2025/Dec/12/openai-skills/)\nconvention (`SKILL.md` + resource files), supported by Claude Code, OpenAI\nCodex CLI, OpenClaw, and other SKILL.md runtimes. Stable installation uses a\ncopied payload from a fixed, immutable release.\n\n### Recommended: verified release\n\n```bash\ngh release download v1.1.1 --repo ponomr/thinking-toolkit \\\n  --pattern 'thinking-toolkit-v1.1.1.tar.gz' --pattern SHA256SUMS\ngh release verify v1.1.1 --repo ponomr/thinking-toolkit\ngh release verify-asset v1.1.1 thinking-toolkit-v1.1.1.tar.gz \\\n  --repo ponomr/thinking-toolkit\ntar -xzf thinking-toolkit-v1.1.1.tar.gz\n./thinking-toolkit/install.sh\n```\n\nThis requires [GitHub CLI](https://cli.github.com/). It verifies that the\ndownloaded bytes belong to the published release before running the installer.\nThe installer copies the payload into every detected agent, prints each path,\nand preserves any replaced installation as a sibling backup.\n\nOptions:\n\n- Pick a host: `./install.sh claude`, `codex`, `openclaw`, or a custom path.\n- `--force` replaces an existing install without asking; it still keeps a backup.\n\n### Updating\n\nUpdates never run inside an agent session. Start them explicitly from an\ninstalled copy:\n\n```bash\npython3 ~/.claude/skills/thinking-toolkit/update.py --dry-run\npython3 ~/.claude/skills/thinking-toolkit/update.py\n```\n\nThe updater verifies the immutable release and its archive, displays every\nadded, removed, and changed file, and asks before replacing anything. It prints\nthe path to the preserved backup and the exact rollback command. Use the\ncorresponding Codex or OpenClaw path if that is the copy you want to update.\n\n### Manual install\n\n| Agent | Personal skills directory | Project-scoped |\n|---|---|---|\n| Claude Code | `~/.claude/skills/thinking-toolkit/` | `.claude/skills/thinking-toolkit/` |\n| OpenAI Codex CLI | `~/.codex/skills/thinking-toolkit/` | `.codex/skills/thinking-toolkit/` |\n| OpenClaw | `~/.openclaw/skills/thinking-toolkit/` | `.openclaw/skills/thinking-toolkit/` |\n| Anything else | Wherever your runtime discovers `SKILL.md` | — |\n\nFrom the unpacked release, copy the `thinking-toolkit/` directory to the chosen\nskills directory. The repository's `scripts/` and `tests/` directories are\ndevelopment tooling and are deliberately absent from the release archive.\nUsing the skill requires no API keys, network access, or code execution;\nnetwork access is used only when you explicitly install or update it.\n\nFor self-hosted gateways and custom harnesses: inject `SKILL.md` into the\nsystem context and make `references/` readable; the skill assumes nothing\nbeyond file reading.\n\n### Development checkout\n\nContributors can clone `main` and run `./install.sh`, but `main` is a development\nline rather than an audited release. Stable users should install a versioned\nrelease. Symlink installation is intentionally unsupported because it would\ncouple every agent to changes in a mutable working tree.\n\n## Security model\n\nThinking Toolkit is instruction code. Treat a change to its Markdown as you\nwould treat a software update. Published releases are immutable, CI validates\nthe skill and reproducibility of its archive, and updates always require a\nuser action and confirmation. See [SECURITY.md](SECURITY.md) for the trust\nboundary, verification commands, and vulnerability reporting. Release and\nrepository maintenance is documented in [MAINTAINING.md](MAINTAINING.md).\n\n## Usage\n\nAsk for a model by name:\n\n> *\"Run a premortem on this launch plan.\"*\n> *\"Compare these three vendors with a weighted matrix.\"*\n> *\"Backcast from where we want the newsletter to be in three years.\"*\n\nOr invoke Thinking Toolkit and describe the situation so it can choose:\n\n> *\"Support tickets keep growing and I don't know where to start.\"*\n> → Pareto Analysis to find the vital few, then Five Whys on the top category.\n\n> *\"We keep arguing about whether to rewrite the billing service.\"*\n> → Six Thinking Hats to structure the debate, Decision Matrix if options\n> stabilize, Red Teaming before committing.\n\nThe agent states which model it chose and why, applies the card's procedure,\nand returns the artifact with its assumptions and a review trigger.\n\n## Repository Layout\n\n```\nthinking-toolkit/\n├── SKILL.md                  # entry point: contract, routing, workflow\n├── VERSION                   # installed and release version\n├── references/\n│   ├── catalog.md            # index, aliases, selection cues, recipes\n│   └── <model>.md            # 31 operational cards, one per model\n├── logic/                    # /logic procedure and focused references\n├── agents/openai.yaml        # optional host discovery metadata\n├── scripts/                  # validation and deterministic release build\n├── tests/                    # regression and supply-chain tests\n├── MAINTAINING.md            # release process and repository safeguards\n├── install.sh                # copy payload into agent skill dirs\n└── update.py                 # verified updates and rollback\n```\n\n### Development\n\n```bash\npython3 scripts/validate_skill.py .        # structure + content invariants\npython3 -m unittest discover -s tests -v   # validator regression tests\npython3 scripts/build_release.py           # reproducible payload + SHA256SUMS\n```\n\nThe validator enforces the invariants that make the skill reliable: every card\ncarries all ten operational sections, every link resolves internally, the skill\npayload contains no external URLs (it must work fully offline), and the catalog\nmatches the card set exactly.\n\n## Design Principles\n\n- **Provider-neutral by construction.** No tool calls, no browsing, no code\n  execution, no memory assumptions. Any capable LLM with file access qualifies.\n- **The smallest useful model set.** One model by default, three at most, and\n  explicit permission to use none: a communication model is chosen only after\n  the underlying analysis is adequate.\n- **User agency preserved.** Model outputs support a decision that stays with\n  the user. The skill never manufactures certainty the evidence does not\n  contain.\n- **Depth matched to stakes.** Quick pass for reversible low-stakes calls; deep\n  pass — with sensitivity checks and disconfirming evidence — for consequential\n  or irreversible ones.\n- **Deterministic quality gates.** A stdlib-only validator keeps all 31 cards\n  structurally complete and the routing tables in sync with the card set.\n\n## License\n\nMIT — see [LICENSE](LICENSE).\n"}