Composing skills

A single skill is one durable cognitive move: restate the problem, sort evidence from inference, imagine the plan has already failed. Real work usually needs a few moves in sequence. Composing is how you string them together without losing the thread.

How recipes work

A recipe is a fixed chain of skills aimed at one recurring job. It runs the steps in order and carries forward only a compressed artifact at each handoff, never the full transcript. The artifact is named, so you know exactly what crosses the boundary.

The flagship is stress test decision, which pressure-tests a consequential decision in four moves:

1. Decision Option Review yields the recommended option and what would flip it.
2. What Would Have to Be True yields the killer conditions.
3. Premortem yields the top risks with tripwires and kill criteria.
4. Reference Class Forecasting yields the outside-view estimate range.

The output is a decision brief: the recommended option, the conditions it depends on, its top risks with pre-decided responses, and an honest base-rate estimate. The other recipes (reframe, expand options, audit reasoning) follow the same pattern with different moves. Browse them all at the recipes index.

Composing your own

You do not need a pre-built recipe. The pattern is the same by hand: run one skill, take its compressed output, hand that as the input to the next. The discipline that makes it work is the handoff. Pass forward the flagged items (the chosen frame, the load-bearing unknowns, the recommended option), not the raw working notes. Each skill’s output format already tells you what is worth carrying.

A natural sequence is reframe, then expand, then decide, then stress-test: get the right problem, widen the options, pick one, then try to break it. Stop as soon as the artifact is good enough to act on.

Picking depth on a page

Each framework page is layered so you can stop at any depth, from a quick-facts card to the full evidence dossier. When you compose, you rarely need the whole page for every step: read the card and procedure for skills you know, open the deep dive and grounding only where the stakes or your doubt justify it. See how to read a page for the four layers and when each is worth your time.

A note on token budgets

For an agent, every skill loaded and every artifact carried costs context. Running fewer, well-chosen skills beats running many. That is why recipes carry a compressed artifact rather than full transcripts, and why the chains stay short. If two skills would surface the same insight, run the one that fits the situation and skip the other. Depth where it pays, not breadth for its own sake.