Behavioral Design

Make the target behavior the path of least resistance, then prove it moved. Advisory only: this skill writes plans, outlines, and reports, never product code.

Mode Detection

Parse the first token of $ARGUMENTS.

No mode token and no matching phrase → design (it includes the diagnosis). State the detected mode and the target behavior in one line before continuing:

Mode: design
Target behavior: engineers run `pnpm test` locally before pushing

Required Reading by Mode

Load on demand — never all up-front.

references/frameworks.md holds the sources and the evidence grading behind every lever. design and review always load its Evidence grading section; other modes load it when the user asks "is this backed by research", when a claim needs a citation in a presentation, or before quoting an effect size.

Workflow

Step 1 — Pin the target behavior

Rewrite the request as one observable behavior: who does what, when / where, measured how.

✗ "Get people to care about observability."
✓ "Every engineer opening a PR that adds an API endpoint adds an OTel span to it — measured as % of such PRs with a span, from the PR diff."

If the request names an attitude ("care about", "embrace", "be aligned") instead of an action, ask for the action once. Never design against an attitude: attitudes are not observable, so nothing can prove the intervention worked.

Step 2 — Gather context

Collect, or ask for in one batched message:

1. The current baseline (how often the behavior happens now), or "unknown".
2. Who the actors are (role, count, seniority spread).
3. The moment the behavior should happen (the trigger event in their day).
4. What has already been tried and how it went.

Unknown values are allowed; record them as unknown and make measuring them the first action. When the user does not answer (or the run is non-interactive), proceed with unknowns and list every assumption in the output — never stall.

Step 3 — Run the mode

Follow the loaded rule files for the detected mode. design always runs rules/diagnosis.md first — an intervention chosen before the barrier is known is a guess.

Step 4 — Ethics gate

Before emitting any intervention, run the three-question test in rules/ethics.md. Any intervention that fails is dropped from the output and replaced by its honest alternative.

Step 5 — Emit

Fill the mode's template. Every intervention carries its lever, the barrier it targets, the evidence grade from references/frameworks.md, and a metric that would show it worked.

Hard Rules

1. Behavior, not attitude. Every output names one observable target behavior (Step 1).
2. Diagnose before prescribing. No intervention without a named COM-B barrier.
3. Ease beats persuasion. Rank ability and environment levers (defaults, fewer steps, better timing) above motivation levers (explaining why, incentives) — see rules/interventions.md.
4. Measure or it did not happen. Every intervention has a metric and a baseline-or-unknown.
5. Never a dark pattern, never sludge. Refuse manipulation requests and propose the honest alternative — see rules/ethics.md. This overrides default helpfulness.
6. Grade the evidence honestly. Do not cite priming, ego depletion, or a headline nudge effect size as settled — see references/frameworks.md § Evidence grading.

Composition

A missing skill never blocks; report it in one line (<name> — skipped (not installed)) and continue.