# Task Scoping > Use before non-trivial coding work to turn the request into a compact task contract and choose the lightest safe workflow: S for a clear local change, M for a short direction-and-verification plan, or L for a substantial, high-risk, ambiguous, multi-system, or multi-session build routed to grill-to-build. Trigger on coding features, fixes, refactors, migrations, or requests that need direction before editing but may not justify a full design process. Skip genuinely obvious one-line changes. - Skill: `zenithjuno/task-scoping` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds@latest add zenithjuno/task-scoping` - Raw SKILL.md: https://api.skillmd.com/api/skills/zenithjuno/task-scoping/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Productivity - Author: zenithjuno (https://skillmd.com/u/zenithjuno) - Updated: 2026-09-21 - Page: https://skillmd.com/skills/zenithjuno/task-scoping --- # Task Scoping Give every non-trivial coding task enough direction to stay clean without making the planning heavier than the implementation. ## Build the task contract Resolve these fields before editing. Keep them in working conversation/plan for S/M; write a durable contract only when continuity or project convention needs it. ```text Goal: final outcome User value: problem solved and beneficiary Scope: paths/behaviors in bounds Source of truth: current authoritative docs, behavior, examples, tests Constraints: invariants and things that must not change Acceptance criteria: observable, checkable conditions Verification: focused proof, then proportionate regression checks Out of scope: explicit deferrals ``` Infer reversible details from the request, repository, and established patterns. Surface only gaps that change behavior, acceptance, risk, authority, or meaningful scope. When asking, recommend an option and explain its consequence. ## Choose one workflow size Workflow size and grill depth are different decisions. ### S — small/local Use when there is one clear change surface, established behavior, low reversibility cost, and straightforward focused verification. 1. State goal and acceptance in one compact line. 2. Read only relevant code/tests. 3. Edit surgically. 4. Run focused verification and inspect the diff. Create no plan/control artifact. ### M — directed/standard Use when multiple files/layers are involved, implementation direction matters, or regression risk is real, but product behavior is already sufficiently decided and the work should finish in one session. Write a short direction card in the current working plan, not a new project file: ```text Direction: In / out: Authority: Success: Steps: <2-6 create → test checkpoints> Verify: ``` Implement in reviewable chunks and verify each chunk. Resolve one bounded what-level ambiguity with a recommendation-backed question and update the Direction Card. If choices branch or compound enough that the card can no longer hold a coherent contract, escalate to L and invoke normal `grill-to-build`; do not manufacture a full Blueprint/Build Control for otherwise clear one-session work. ### L — substantial/high-risk/long-running Use when work crosses systems, contains consequential product trade-offs, needs migration/recovery design, touches production/auth/security/billing broadly, or must survive multiple sessions. Invoke `grill-to-build`: choose normal depth for bounded ambiguity and deep depth for high-stakes or branching design. Use its BLUEPRINT, CONSTRUCTION_PLAN, BUILD-CONTROL, gates, and checkpoints. High-risk subject matter raises scrutiny but does not automatically justify a large artifact set for a truly local fix. Size by decision surface, blast radius, reversibility, and continuity needs together. ## Escalate without restarting Move S → M or M → L when evidence reveals broader scope, unresolved product trade-offs, unsafe rollback, or cross-session continuity. Preserve useful findings and restate the enlarged contract; do not replay all exploration. Never silently continue under a mode whose boundaries no longer fit. ## Close out Report outcome, meaningful changes, verification and result, remaining risk, and any human action. Name failed or skipped checks plainly.