# Shape Work > Clarify or stress-test an idea into an implementable outcome, capture a useful specification, or directly reconcile domain meaning, invariants, context relationships, and settled decisions. Use for shaping work or requested domain-record updates; exclude routine vocabulary lookup, straightforward implementation, and delivery-ticket decomposition. - Skill: `stevennitesh/shape-work` (Agent Skill, multi-file: 5 files) - Install (CLI): `npx skillmds@latest add stevennitesh/shape-work` - Raw SKILL.md: https://api.skillmd.com/api/skills/stevennitesh/shape-work/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: stevennitesh (https://skillmd.com/u/stevennitesh) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/stevennitesh/shape-work --- # Shape work Make consequential decisions clear enough to build the intended outcome, using conversation, evidence, and a specification only when useful. Preserve the requested scope; shaping alone does not authorize implementation. For direct domain-modeling or settled-decision capture, follow [Domain modeling](references/domain-modeling.md) without a feature interview. ## 1. Establish the outcome and known decisions Identify the problem, intended beneficiary, desired outcome, and decisions already made. Use the supplied conversation, brief, issue, or audit findings; inspect decision-bearing references and relevant repository contracts. Distinguish settled requirements, current implementation facts, recommendations, and assumptions. Capture settled input without restarting an interview or seeking approval again. Revisit only choices affected by new evidence. Assumptions and deferrals are not accepted requirements. Look up accessible facts instead of asking the user to retrieve them. When a proposed solution depends on an existing capability, verify enough of its actual behavior to avoid promising something it cannot do. Correct incidental factual errors; surface contradictions that would change an accepted commitment. ## 2. Resolve the decisions that affect the outcome When the mechanism is still open, check whether an existing capability or smaller change achieves the intended outcome. Explain any meaningful loss or ceiling; do not silently simplify accepted requirements. Separate the required outcome from a proposed mechanism. When the choice matters, compare genuinely different approaches, including no software change when credible; investigate the assumption most likely to invalidate the choice before refining it. Ask only about unresolved choices that materially affect behavior, scope, constraints, acceptance, or tradeoffs, one question or a small independent group at a time. Wait for answers before dependent questions or decisions; continue useful independent work when available. Incorporate each answer without asking permission to proceed or repeating unchanged questions. Once the necessary choices are settled, continue to acceptance and capture within the requested scope. Use concrete scenarios to challenge vague agreement. Explain a recommendation and its decisive tradeoff when the evidence supports one, while eliciting participant-held facts neutrally. Technical decisions within settled requirements are yours to recommend and resolve within authorization; product priorities and accepted meaning belong to their decision owner. Do not invent their answer. When a term, invariant, or relationship changes or conflicts with current domain meaning, read [Domain modeling](references/domain-modeling.md) and resolve the collision before building dependent decisions. Clarifying meaning does not require writing a record after every answer. For an unresolved architecture choice, use `$codebase-design` when available and useful. For an empirical question, use `$prototype` or available evidence-gathering tools within scope. These are methods, not installation prerequisites. Keep a conclusion conditional when the necessary evidence or owner is unavailable. Record what remains unknown, its impact, and what would resolve it. Summarize a complex synthesis for review when it could conceal a material misunderstanding; this does not reopen choices already made. ## 3. Make behavior and acceptance precise Describe a representative caller journey from initiating input to observable outcome. Specify materially different rejection, partial-success, or state behavior when it affects the outcome. State what must remain true and what evidence could distinguish success from a plausible wrong result. Avoid implementation recipes unless a particular mechanism is itself a settled constraint. Could two reasonable implementations satisfy the wording but produce materially different outcomes? Resolve unintended ambiguity or identify the remaining decision; preserve deliberate implementation freedom within accepted behavior. For composed results, durable state, conflicting rules, or measured claims, read [Acceptance meaning](references/acceptance-meaning.md). Preserve consequential semantics without enumerating irrelevant cases. Keep unresolved choices visible; do not fill gaps with plausible prose merely to make a specification look complete. ## 4. Capture only what future work needs Keep a small outcome in the conversation when that is sufficient. Produce or update a durable specification when requested or needed for multiple sessions, delivery slices, or owners. Read [Durable decisions](references/durable-decisions.md) when writing a spec or updating domain records. Follow the repository's configured destination; a spec need not be a tracker parent. External publication requires authority for that effect; a local draft can still make progress. A useful result contains the problem and outcome, scope and consequential exclusions, settled behavior and constraints, acceptance, and material uncertainty. Add source pointers and rationale that help a fresh agent preserve the decisions. Omit empty sections and exhaustive user-story catalogs. Reuse the existing owner of a decision rather than creating competing current documents. ## 5. Return or continue within scope Do not label work ready when implementation would need to invent consequential policy. If blocked, return the exact decisions or evidence still needed. Once shaping is settled, continue already-authorized work. For one coherent outcome, implementation can proceed directly. Use `$to-tickets` when the user requests ticketing, including a single ticket. When durable delivery boundaries or coordination would benefit from tickets, suggest that explicit next step; do not start it merely because it could help. Neither a spec nor tickets are a prerequisite for ordinary coding. Creating plans does not itself authorize implementation, publication, or delegation.