Scope

Scope

Diagnose why work isn't buildable yet, route to exactly one specialist skill, and open a ledger the moment routing happens. Never do the specialist's work yourself — a correct route with no other output is a complete helper result; return it to the caller, which continues its own authorized workflow. For an epic child, every specialist instead uses untracked session scratch outside the child branch, discarded after the final order; it creates no scope ledger or docs/scope ledger. /epic alone owns the parent proposal, design, and tasks.

Routing table

Classify the dominant uncertainty — the one that, if resolved, makes the others tractable — and route to exactly one of:

• Big and foggy, many interlocking decisions that block each other → epic. Mechanical bulk without an unsettled decision is not an epic route; hand-split it into serial build tickets.
• A concrete plan or design exists in someone's head, untested → interview mode (references/interview.md, in this skill).
• A written plan, work order, spec, or brief exists and needs stress-testing before anything is built → plan-review.
• Missing facts answerable from docs or sources → research.
• Only answerable by feeling it out in running code → prototype.
• The dominant uncertainty is what a user-facing surface should look like → ui-craft, lock phase.
• Mixed or unclear → ask the user one framing question, then route. One question at a time, always — never a menu of questions.
• Nothing genuinely uncertain — the plan is grounded, the decisions are already settled, and any open point has an obvious default → say so in one line, name the defaults being assumed, and return control without asking anything. No ledger. Do not manufacture questions to justify the invocation; a caller may route here unconditionally, and "nothing to scope" is a correct helper result.

If the request already names a specialist by trigger phrase (a grill-style trigger names interview mode; stress-test this plan names plan-review; an epic map names epic), route there directly without re-diagnosing.

Ledger — required

Outside an epic child, at the moment routing happens, open a ledger before invoking the specialist:

• Path: docs/scope/<slug>.md in the target project's repo when that path is writable; otherwise the session scratchpad. <slug> is a short kebab-case name for the work.
• Sections: Decisions, Open questions, Spawned tasks.
• Decisions get appended at the moment each one settles — never batched at session end. Each entry: the decision, a one-line why, and a disposition tag — → ADR, → issue, or inline.
• The ledger is the durable session state. A fresh agent resumes by reading it, not by re-deriving context from chat history.

For an epic child, the remaining ledger, risk-contract, evidence, and exit rules are applied only through the stamped work order and parent epic authority; session scratch is never made durable.

Risk contract

For interview-mode work, and for any other bounded plan being declared ready to build, settle a risk contract before admission. An epic map whose work is not yet bounded does not need one.

Keep one ### Risk contract block under the ledger's Decisions section:

• Must prevent: harmful outcomes the work may never produce.
• Must recover: concrete failures that require automatic recovery.
• Accepted failure: a concrete failure and the exact consequence the user accepts, such as a clear stop with manual recovery, skipped best-effort work, or degraded non-authoritative output.
• Unsupported: inputs or operating conditions the work does not promise to handle.
• Evidence owed: public-interface behaviors and named invariants that require tests or another explicit check.

Give the block one-line Why: and Disposition: fields rather than tagging every line separately. Default Must prevent to secret exposure, irreversible loss of authoritative data, and silent incorrect success; changing one of those defaults requires an explicit user decision. Default rare, recoverable failures to an Accepted failure with a clear stop and manual recovery, not automatic recovery.

Price the contract from concrete consequences, exposure, and recoverability. Audience size is evidence, not an assurance tier: one person's financial tool can be high-stakes, while a public toy can be disposable. Do not use a target test count. Evidence is owed only by supported behavior, a must-prevent outcome, an enforced invariant, or an observed regression.

At admission, copy the risk contract unchanged into the authoritative issue, plan, or brief that the implementation and reviews will use. The ledger remains the session record; the admitted artifact becomes the downstream authority. Do not declare bounded work ready while the contract is missing or still only in the ledger.

Exit protocol

The session is not done until every disposition is discharged:

• Every → ADR decision has a real ADR written.
• Every → issue decision has a filed GitHub issue.
• The ledger's remaining-dispositions list is empty.

Report which dispositions remain open if the session ends early; do not report the scoping work as finished while any disposition is outstanding.