── ── Open source
How We Made 163 Mental Models Executable for AI Agents
July 8, 2026 · 7 min read
Mental models like first principles and inversion usually live in books as prose an agent can't run. We turned 163 of them into open-source Agent Skills — each with explicit trigger conditions, a step-by-step process with hard gates, and worked historical case studies cited to primary sources. MIT-licensed, 10K+ installs on ClawHub.
Every founder has read about first-principles thinking. Almost no one — human or AI — actually runs it. The method lives in books and blog posts as prose: inspiring to read, impossible to execute. Ask an LLM to "use first principles" and you get vibes — a paragraph that name-checks Musk and then does whatever it was going to do anyway.
We wanted our agents to genuinely run these methods before acting — tear assumptions to bedrock before pricing a product, invert before shipping a launch plan, update like a Bayesian when evidence lands. So we rebuilt 163 mental models as Agent Skills: self-contained SKILL.md files an agent loads and executes. Then we open-sourced all of them under MIT. Here's what "executable" actually took.
A mental model becomes executable when it can fail
The difference between a description and a process is that a process can halt. Each skill has three load-bearing parts most prompt libraries skip:
- Explicit triggers — every skill's description states "Activate when: …" with concrete user phrasings, and most add "Do NOT activate when: …". This is what lets an agent load the right method at the right moment instead of applying its favorite hammer to everything.
- A numbered process with hard gates — steps like "Gate: no quantified baseline → stop." If the input doesn't meet the bar, the skill refuses to proceed rather than producing confident filler.
- A Common Rationalizations table — the fake moves people (and models) use to dodge the method, each paired with the reality check. "Everyone feels more aligned" is not a velocity change. "Value exists but is hard to quantify" is the standard defense of programs that don't work.
That last one matters more than it looks. LLMs are excellent at producing text that sounds like the method. The rationalization table is a list of exactly those moves, so the agent can catch itself mid-dodge.
Every skill ships a real case, cited to a primary source
Abstract methods don't transfer. So every skill includes worked historical examples — real cases walked through the skill's own steps, with the primary source cited at the bottom. Not invented "imagine a startup…" scenarios. These:
| Skill | Case | The move |
|---|---|---|
| checklist | Van Halen's brown-M&M rider clause (1980s) | A trivial, instantly verifiable item embedded in a safety checklist proves whether the counterparty read the rest |
| black-swan | Fukushima vs. Onagawa tsunami design basis (2011) | The 869 AD paleotsunami record was in the literature; one plant's design basis treated the tail as real, one didn't |
| inversion | Abraham Wald's bomber armor analysis (1943) | Armor where returning planes show no damage — survivor data is silent exactly where hits are fatal |
| second-order-thinking | The Hanoi rat bounty (1902) | Pay per severed rat tail and you don't get fewer rats — you get tail-less rats released to breed, and rat farms |
| bayesian-reasoning | The search for Air France 447 (2009–2011) | Discounting negative search evidence by the probability both acoustic beacons failed kept the posterior near the crash point — found in a week |
| first-principles | SpaceX's rocket cost teardown (2002) | Decompose launch prices into commodity bill-of-materials; the citable industry number was inherited belief, not bedrock |
40 of the 163 skills now ship two cases from different domains, and we're extending that to the whole library. The sourcing rule is hard: cite only documents that exist, paraphrase rather than fabricate quotes, and when a number is a reconstruction, label it as one.
The numbers
deciqAI Knowledge Skills, today
164 skills · MIT licensed · 10K+ installs on ClawHub · worked examples cited to primary sources · the same skills wired into every deciqAI agent. Install: npx skills add deciqAI/knowledge-skills
The library covers reasoning methods (first principles, inversion, Occam's razor, Bayesian updating), cognitive biases (anchoring, sunk cost, confirmation bias, hindsight), strategy frameworks (Porter's five forces, blue ocean, disruption), and operating tools (checklists, OKRs, OODA, pre-mortems). Each is a folder: SKILL.md plus examples/ plus references/ — copy one into your agent's skills directory, or install all 164 with one command.
Why we open-sourced the thinking layer
These are the same skills every deciqAI agent runs before it acts — when it drafts your pricing, plans an outreach campaign, or tears down a business plan. Open-sourcing them costs us nothing competitive: the moat isn't the method, it's the execution loop around it — the data, the agents, and the company operating system they run inside. Publishing the thinking layer just means anyone can check our agents' reasoning against the source, and any builder can give their own agent the same rigor.
Contributions are open — the repo ships a skill template and quality bars (real triggers, hard gates, primary-sourced cases, no placeholders). The easiest first contribution is an [O] entry: a real failure pattern you observed using a skill, pasted into its rationalization table.
FAQ
What is an Agent Skill?
A self-contained SKILL.md file (an open format introduced by Anthropic and now supported across Claude Code, Codex, Cursor, Gemini CLI and others) that an AI agent loads to execute a specific capability — here, a thinking method with triggers, steps, gates, and worked examples.
How is this different from a prompt library?
Prompts describe; skills execute and can halt. Each skill has explicit activation conditions, numbered steps with hard gates that stop the process when inputs don't meet the bar, and a table of the fake moves that imitate the method — so the agent can catch itself producing method-flavored filler.
Are the case studies real?
Yes — every worked example is a documented historical case walked through the skill's own steps, with the primary source cited (e.g., Wald's survivability memoranda, the NAIIC Fukushima report, Stone et al.'s AF 447 Bayesian search paper). No invented scenarios.
Can I use them outside deciqAI?
That's the point — they're MIT-licensed and work in any agent that supports the SKILL.md format. Run npx skills add deciqAI/knowledge-skills, install from ClawHub, or copy a single folder from GitHub.