mental_models v1.0.0

name: "mental_models" description: "Apply mental models to real decisions: inversion, second-order thinking, base rates, opportunity cost, and 20+ more, with a router that picks the 2-3 models that create signal for the decision at hand, worked examples, and notes on when each model lies. Use when making, auditing, or stress-testing a decision, plan, proposal, or strategy." metadata: { "includeInPrompt": true }

Mental Models

Purpose

Give an agent a thinking toolkit for decisions that matter. The skill does three things competitors' model lists don't: it routes the decision to the 2-3 models most likely to flip the conclusion, it shows each model applied to a worked example, and it says plainly when each model lies. A catalog of 28 models is a file drop. This is applied judgment.

Workflow

Step 1: Restate the decision

Strip the jargon. Reduce it to "should we do A or B (or C)?" If you cannot, ask the user to clarify before touching any model.

Step 2: Pick the lenses (2-3, not all)

Scan the router below and take the 2-3 models most likely to change the conclusion. Never force all models onto one decision. Models not in the catalog are out of scope.

| Situation | Reach for | |---|---| | Big irreversible commitment (hire, stack choice, launch) | Margin of safety, Asymmetric payoffs, Path dependence | | Multi-party system (partners, markets, org) | Incentives, Second-order thinking, Hanlon's razor | | Crowded space / common idea | Base rates, Opportunity cost | | Stuck on positive framing | Inversion | | Plan keeps adding complexity | Via negativa | | Outside your familiar domain | Circle of competence | | Metrics, OKRs, incentive design | Goodhart's law, Incentives | | "Why did X happen / why did they do that?" | Hanlon's razor, Incentives | | Rewrite, migration, or replacement proposal | Path dependence, Opportunity cost | | Everyone agrees, no friction | Confirmation bias, Inversion | | Estimating a timeline or cost | Base rates, Hofstadter's law | | Deciding what NOT to do | Opportunity cost, Via negativa |

Full model cards (what the model says, when it lies, worked example): references/model-cards.md. Graded decision walkthroughs (simple to hard): references/worked-examples.md.

Step 3: Apply each lens

For each chosen model, write one short block:

• The model: what it says, in one sentence.
• What it reveals here: the specific fact about this decision the lens exposes.
• When it lies: the known failure mode of the model itself (see model cards).

If a model reveals nothing new, drop it. A lens that finds nothing is not evidence, it's a discarded lens.

Step 4: Converge

Collapse the findings into concrete output (see Output Contract). Flag which model flipped or reinforced the initial leaning, and name one piece of evidence that would change your conclusion.

Output Contract

Every run returns:

1. Decision restated — the A/B/C reduction from Step 1.
2. Lenses used — the 2-3 models chosen and why these, in one line each.
3. Findings — per model: what it revealed, and the model's known failure mode checked against this case.
4. Recommendation — the call, with the single biggest risk named.
5. Disconfirming evidence — one observable fact that would flip the call.

Operating Rules

1. Never apply more than 3 models unless the user asks. Depth beats coverage.
2. Inversion is the default first move when the user is already sold on an idea. Second-order thinking is the default when the decision has ripple effects. Base rates come before any estimate, not after.
3. If a model contradicts your recommendation, say so explicitly. Hiding the tension is a failure.
4. No model is a proof. "Inversion says X" is weaker than "inversion reveals cost Y nobody priced." Always attach a concrete fact.
5. When models disagree with each other, the disagreement IS the finding. Report which lens won and why.
6. Check the model's own failure mode (Step 3) before finalizing. A lens that lies here gets dropped, not footnoted.
7. Never invent numbers, dates, quotes, or events to make an example land. Unverifiable = cut it or frame as opinion.