Step Isolation

A steering lever for the failure where an agent doesn't do enough "leg work" on an intermediate step — it under-invests in exploring, or in asking clarifying questions, and moves on too soon. The source's diagnosis: the agent can see the ultimate goal and rushes toward it. The canonical case is plan mode, whose two steps are "ask clarifying questions" then "create a plan" — the agent, seeing that the goal is a plan, does a token amount of clarifying and then eagerly produces the plan.

The fix is step isolation: split the multi-step skill so the agent sees only one step at a time, hiding the future goal (and future steps) it would otherwise sprint toward. The example: instead of a single plan-mode skill, a separate grill-with-docs skill is the clarifying- questions phase — the agent only sees that part — and only after it completes does control pass to /to-prd for planning. Same two logical steps, but the agent experiences them one at a time, which reliably forces more leg work on the current one. It isn't always necessary to split a skill into individual steps, but where you specifically need an extra chunk of effort, "there's no technique like it."

Claims

Agents under-invest leg work in intermediate steps because they can see the ultimate goal and rush to it — plan mode's clarifying-questions phase is the classic case. observation — a behavioral pattern the source reports finding "everywhere it exists."
Splitting a multi-step skill so the agent sees one step at a time — hiding the future goal — reliably increases leg work on the current step. best practice — context: a step that genuinely needs deeper effort (exploration, clarifying questions); not every skill should be split, since the split adds handoff structure — reach for it when a specific step is under-served.
The mechanism is hiding the future, not adding instruction — removing the visible downstream goal is what changes the agent's allocation of effort. principle — durable: it's a context-visibility intervention; you change behavior by changing what the agent can see, not by telling it to try harder.
"Plan, don't do": keep the planning stage isolated from execution so the agent charts the route instead of charging at the destination — and keep refactoring out of the red-green loop so it doesn't overload the implementation. best practice — context: a second source (skills v1.1) applies stage separation two ways — a Wayfinder planning skill that only plans, and a TDD loop that drops refactoring into the code-review stage. The mechanism is the same one-thing-at-a-time isolation as within a skill, lifted to the pipeline boundary; contingent on the stages being genuinely separable (some tight work is cheaper interleaved).

Leading Words — the paired steering lever in the talk; when a leading word alone can't get enough effort, restructure with step isolation.
Minimal Skill Surface — splitting into single-step skills also shrinks each skill's surface, though the primary motive here is leg work, not size.
Skill-Driven Loop Development — chaining single-step skills (grill-with-docs → to-prd) is a hand-run version of the orchestration-over-execution-skills split.
Skill Authoring Checklist — pillar 3 (Steering), the leg-work lever.
Map-First Planning — "plan, don't do" at the pipeline scale: chart the route as a map before any execution, so the agent doesn't sprint at the destination.
Distillate: Building Great Agent Skills: The Missing Manual
Distillate: Skills v1.1: Wayfinder, the SDLC flow, and naming the artifact right — stage isolation applied to the plan/do boundary (Wayfinder) and the TDD/refactor boundary.
Context-Independent Review — the same hide-what-would-bias-you move across the write/review boundary: the reviewer must not have seen the authoring, for the same reason the griller must not see the plan it is heading toward.
Distillate: The whole flow, end-to-end: the smart zone is the unit of work — grill-with-docs shown in practice: a deliberately vague opening idea, ~6 questions (the author's sessions "usually about 20"), run in auto mode rather than plan mode, ending at shared understanding before any plan is laid out.

Step Isolation

Claims

Related

Linked from