| name | omg-deep-interview |
| description | Deterministic, resumable requirements convergence before planning or implementation. |
omg-deep-interview — requirements first
Use this skill when a request is vague, brownfield, or missing explicit
boundaries and acceptance. The omg CLI owns the interview state and final
specification; an agent may provide repository facts or suggested wording, but
must not write authoritative interview state.
Contract
- One pending question per round, always targeting the weakest unresolved
clarity dimension after intent-first stage priority.
- Profiles:
quick (threshold .30, 5 rounds), standard (.20, 12), and
deep (.15, 20).
- Brownfield scoring uses intent, outcome, scope, constraints, success, and
context. Repository evidence is gathered before asking for discoverable facts.
- Non-goals, decision boundaries, acceptance, and one explicit pressure pass
are mandatory even when the numeric threshold passes.
- A clear labeled task may take a zero-question path, but it still requires a
pressure pass before close.
- Until
status is complete, do not plan or implement. waiting_input
always includes the exact resume command.
- Wrong-run, stale-question, corrupt-state, and artifact-hash mismatches fail
closed.
CLI
omg interview start "task" --profile standard
omg interview status --run RUN
omg interview answer --run RUN --question-id QUESTION --text "answer"
omg interview pressure-pass --run RUN --text "assumption and trade-off result"
omg interview close --run RUN
start, answer, and pressure-pass print the current state and an exact
resume_command. The final CLI-stamped JSON spec and transcript live under the
run stages/ directory; a human-readable projection is written under
.omg/plans/.
Labeled zero-question intake
For already-clear work, the task text may provide these labels (one per line or
separated by semicolons): Intent, Outcome, Scope, Constraints,
Success, Context, Non-goals, Decision boundaries, and Acceptance.
The CLI validates completeness and ambiguity; labels do not bypass the pressure
pass or close gate.
Hard rules
- Do not invoke an LLM from this primitive and do not build a parallel chat UI.
- Do not mutate product code during an incomplete interview.
- Do not self-declare completion or write
passes/verified; only the CLI may
stamp artifacts and lifecycle state.
- Model/agent prose is a proposal. It becomes authoritative only through the
CLI command path.