Scope

Scope

Turn an idea into decided, testable scope under project constraints.

Frame

Define the decision, users and done. Search specs and registers. Emit the minimum config/risk-policy.json tier plus allowed source/artifact paths, prohibited actions, external effects, expiry and approver. Ordinary authorised workspace content may cross approved provider families; credentials, secrets, platform rules and narrower task/project bans control. Only a user may downgrade risk. Judgement-bearing AI, ranking or heuristics require evaluate.

Preserve decision context: intake/revision, goals, constraints, alternatives, evidence and decided/parked branches. Revise it; never fork competing scope. Record disclosure in the authority envelope, never private memory.

Grill

Load grill-me only when the user explicitly asks to be grilled or dependent owner decisions remain materially unresolved. Then work purpose -> users -> constraints -> edge cases -> failure modes -> success -> exclusions, one question per round. Otherwise present a compact decision packet with 2–3 choices, a recommendation and parked owner calls.

Agents decide engineering calls. Business, legal or financial owner calls stay parked as named open-decision rows; never guess. Put every unresolved branch in the spec.

Resolve uncertainty

Research surviving questions; use orchestrate for fan-out and source retained claims. Use prototype when a timeboxed throwaway answers feasibility; harvest its result and delete/quarantine only manifest-owned scratch under its authority. Neither lane exists for curiosity.

For viable options, compare cost, reversibility, risk and prior decision fit through correctness/cost/operations lenses. Use independent reviewers when available; recommend one. Put costly-to-reverse choices and rejected alternatives in an ADR, reversible detail in the spec/story. Paired-primary mode has one chair ask while the peer audits evidence; record authorship for later independence.

Land outputs

First resolve artifact authority and canonical owners. In advisory/read-only mode, return proposed scope and named open decisions in chat; do not change project files. In project-write mode, land only approved artifacts:

| Output | Owner | |---|---| | Spec, stories, acceptance criteria | project docs via engineering-docs | | One-way decisions | project ADR process | | User gates | existing register or docs/OPEN_DECISIONS.md | | Work items | project tracker | | Durable context | project context/state owners |

Write clear, observable, verifiable acceptance criteria. Given/When/Then helps behavioural cases but is not mandatory for research, documents or operations. Preserve project Markdown/YAML/JSON schema. Pin only decision-critical external interfaces through project-native locks/constraints. Link or cache permitted authoritative material with source, version/date and digest; never vendor without licence and redistribution authority.

Before handoff confirm: branches decided/parked; explicit exclusions/failure modes; machine-readable authority/risk; anchored evidence; user approval of spec and one-way doors. Execution handoff is digest-bound to exact approved scope, decisions and authority; change creates a new revision and gate.

Adapter-absent path

Without optional Console, Herdr or GitHub, use canonical project artifacts and emit the skill-owned kind in [portable-workflow.v1.json](portable-workflow.v1.json). It records scope evidence; never supplies user approval.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: mblauberg
• Source: mblauberg/provenant
• License: MIT

Install and usage instructions live in the source repository linked above.