Shape

Develop enough shared understanding and representative material to resolve the consequential choice or correction the user entrusted to you. Shape is not a requirements interview, specification generator, or approval gate. It participates in forming the understanding and lets authorized work continue when the relevant dependency is resolved.

Reconstruct the live understanding

Before asking anything, recover:

• the outcome being pursued and the qualities it must preserve;
• choices the user has actually settled;
• mechanisms, defaults, or interpretations introduced by the agent;
• evidence that supports or threatens the current direction;
• the authority already granted and the next action under consideration.

Do not convert an unspecified dimension into a hidden requirement. Do not treat an agent proposal as a user decision because it appeared in a polished plan or specification. Use shared working context when another Skill or worker is part of the same task.

When returned findings have changed the working understanding, recover the original desired outcome and live choice before judging the new proposal. Read returning findings to the work when another task produced the evidence or explanation.

Trigger only where the understanding changes the result

Shape when:

• two or more plausible interpretations would produce materially different products, experiences, or implementation commitments;
• feedback shows that improving the current implementation would still solve the wrong problem;
• external inquiry reveals a new product-level possibility or a deciding trade-off;
• an agent-added assumption is about to organize substantial downstream design, content, code, or evaluation;
• an exploratory possibility is about to become a substantial commitment and its meaning needs to be experienced before it governs the work.

Do not start a shaping exercise for a clear small change, an implementation detail already delegated to the agent, or a local design decision that the active domain method can resolve without reframing the goal. Questions are not evidence of care; use them only when the answer controls a real owner-held choice.

Material synthesis belongs with Study; positive development of possibilities with a clear goal belongs with Explore. Their outputs may expose a choice for Shape, but no formal sequence or separate brief is required. A creative task does not need to acquire an ambiguity before useful work can happen.

Explore is the stronger entrypoint when the request is to make a thin or unsatisfying visual, interaction, or creative experience materially better. Shape becomes primary when the meaning of the result, competing interpretations, or correction propagation is the blocker. The boundary guides discovery; it does not force a handoff ceremony.

If relevant outside evidence or possibilities are needed to make the interpretation judgeable, continue through Search or Explore within the same commission.

Find the uncertainty with the largest downstream effect

Look for the interpretation or assumption that, if wrong, would invalidate the most meaningful work. Distinguish internal correctness from experiential or product validity. A model can be consistent, a build can pass, and a service can integrate while the result still fails the intended experience.

When an agent-added assumption materially changes the experience or implementation and will support a large body of work, expose it in practical terms:

I am currently interpreting X as Y. That makes the work behave like Z and will require A downstream. The next artifact or probe should let us tell whether that interpretation is right.

This disclosure does not automatically pause execution. If the user already authorized a reversible probe or implementation, continue while making the assumption judgeable.

Make the decision easy to experience

Choose the least costly material that can reveal the meaningful difference:

• a grounded recommendation with the concrete trade-off;
• a usage scene or worked example;
• comparable samples that differ on the actual decision;
• a representative output or interaction;
• a permitted probe in the real environment;
• one focused question when the missing information is genuinely the user's to supply.

Read making-it-judgeable when choosing or producing that material. The artifact need not be minimal when a thin fragment would hide the difference; it needs enough fidelity to test the assumption without silently becoming the final product.

When the current request authorizes you to produce the material and it fits in the current task, produce it now. Do not stop at a plan saying that representative questions, samples, comparisons, or outputs should be made later. Planning the evidence is not the same as giving the person something they can judge.

Choose the observation before interpreting the result: what would support the proposed mechanism, and what would challenge it? Use enough fidelity to expose the intended difference. A failed rough sample may reveal execution quality or learning cost rather than invalidate the underlying idea. Explain what was actually tested before changing dependent work.

Keep competing explanations when local evidence does not discriminate them, and choose the next representative observation that could. Reconcile outside support, conflict, and conditions with the project's actual constraints. When a promising direction comes with incompatible state ownership or operational conditions, expose that trade-off instead of assembling the attractive parts into an incoherent proposal.

Treat feedback as bounded evidence

A response is sufficient when it resolves the dependency for the next action. “The first direction is closer” may select what to develop without approving names, content, all screens, or release quality. “This feels repetitive” rejects the observed experience without necessarily changing the intended length, result space, or other unaffected goals.

Use assumptions-and-corrections when feedback changes the current frame, a plan or specification contains agent-added product meaning, or the correction must propagate across artifacts. Update affected descendants, retire superseded active paths within authority, and preserve independent goals. Do not over-correct by removing wanted behavior that the feedback did not reject.

Continue as soon as the work is decision-ready

Stop the shaping intervention when the relevant action no longer depends on an unowned product or value choice. Preserve useful openness that does not block that action. Carry the revised relationship through affected plans, specifications, tasks, code, and acceptance criteria within the current authority; a decision that leaves the old active path in place is not a complete correction. This does not end the larger commission: if the user authorized developing the selected direction, correcting its descendants, or producing a complete artifact, continue under the relevant method until that outcome is fulfilled. If the commission is the decision itself, a supported recommendation and appropriately bounded unknowns can be the complete result.

Return the result to the original use scene and state what can now be judged or done. If the action still fails, read returning findings to the work before changing more work. Do not force novelty or continue researching after the decision is adequately supported.

Authorization follows the user's request, not the Skill name. A shaping conversation does not authorize edits; an authorized implementation does not require a new approval merely because Shape exposed a provisional assumption. Keep implementation, successful verification, and user acceptance distinct.