| name | decision-coach |
| description | Use at the start of MATERIAL or PROGRAM work when the user has a rough request, unclear success criteria, hidden trade-offs or missing decisions. Ask only high-impact questions, recommend defaults and produce a better execution brief. |
Decision Coach
- Identify the actual decision and desired outcome.
- Inspect available project state before asking questions.
- Separate known facts, assumptions and material unknowns.
- Ask at most five questions whose answers can change the solution.
- For each question provide:
- why it matters;
- recommended default;
- consequence of another choice.
- Produce an improved execution brief containing scope, non-goals, users,
domain, constraints, outputs, acceptance and verification.
- Stop discovery when planning can proceed safely.
Do not ask the user to rewrite their prompt from scratch.
A recommended default is a proposal, not a substitute for asking
A default exists for unattended runs (CI, no human in the loop) and to
make an interactive answer faster to give -- it is not permission to skip
asking.
When a live user is present:
- Present the material questions directly, with each default and its
consequence stated -- do not silently substitute the default and
continue as if they had answered.
- Wait for a real answer, or explicit instruction to proceed with the
stated defaults, before advancing past
DISCOVER on MATERIAL or
PROGRAM work.
- The bigger the blast radius -- an architecture choice, a compliance
rule, anything hard to reverse -- the less acceptable it is to have
decided it without them, even under time pressure.
Silently self-answering is only correct when no user is available at all.
Surface Relevant Capabilities
Name matched, not-yet-installed capabilities in the same response as the
material questions, not a separate report (AGENTS.md, step 8).
- Check
capability-audit for the actual request; name each match in one
line -- what it is, why it matched, and a recommended yes/use-it or
no/skip-it.
- "Yes" is an explicit approval gate (
AGENTS.md, Approval Gates) -- a
registry name is a reviewed candidate, never an installed default.
- No strong match: offer, do not silently perform, a research pass via
reference-router for current external options; wait for go-ahead.
Skip this section for MICRO work.
Grilling Intensifier
Explicit "grill me" / "stress-test this plan" / "challenge every
assumption" deepens this same procedure -- decision-coach stays canonical,
never activates on an ordinary feature request
(catalog/overlap-matrix.json, group decision-interview).
- Map the plan's real branch points; inspect the repository per branch
before asking.
- Still cap at five material questions -- highest blast radius first,
defer the rest.
- State each branch's hidden assumption explicitly.
Adapted, original wording, from Matt Pocock's MIT grilling skill
(github.com/mattpocock/skills/tree/main/skills/productivity/grilling).
