sciman-top Grill with Docs
Run a focused design interview using the grilling and domain-modeling
skills when they are available. Ask one high-leverage question at a time and
keep the current decision, alternatives, constraints, and unresolved questions
visible.
Native-child dispatch invariant
This skill's multi_turn_user_decision contract is executed by the native
design-griller, never by the parent. Before producing a question, conclusion,
or delegation claim, call the host-native spawn_agent tool with the exact
user request, available evidence, and constraints. Continue only after it
returns a non-empty child task or thread identifier; bind that same identifier
to every later wait and follow-up call.
Follow-up answers may be either attributable human verbatim input or an
explicitly authorized authorized_ai_delegate_answer. The parent must preserve
the attribution kind, authorization evidence, and answer SHA-256 in the
successor admission. An AI delegate answer can advance the interview, but it is
never a human answer, human acceptance, or host_specific_live_accepted
evidence; unclassified parent summaries and automatic continuation remain
invalid.
If native spawn is unavailable, fails, or returns no identifier, stop with
native_bridge_unavailable. Do not call a bare wait, simulate a child,
continue the interview in the parent, or replace the interview with a one-shot
analysis, report, or summary. Evidence gathering informs each child question;
it never substitutes for the child.
Treat a request to brainstorm, implement, or refactor as out of scope unless the user also asks for design grilling.
Only update CONTEXT.md, a glossary, or an ADR after the user has confirmed a
durable decision or explicitly asked for that record. If the interview does not
produce a durable decision, return a structured summary and do not write files.
Do not publish a spec or tickets and do not start an architecture scan. Use an
explicit $to-spec, $to-tickets, or $improve-codebase-architecture
invocation for those operations.