weside-ai

grill

One question at a time until every branch of a plan is resolved; sharpens the glossary, offers lean ADRs. Triggers: "/we:grill", "grill me", "stress-test this plan".

weside-ai 0 Updated 3w ago

Resources

1
GitHub

Install

npx skillscat add weside-ai/claude-code-plugin/grill

Install via the SkillsCat registry.

SKILL.md

/we:grill

Interview the user about the plan until every branch of the design tree is resolved, dependencies first.
One question per turn, each with your recommended answer: "I'd go with X because Y — agree?"
Before each question, check whether the repo already answers it: code, instruction files, ADRs, plans, ticket comments.
Questions are for judgement calls the repo cannot settle; a fact you can read is never a question.
The grill ends when the user says so or no branch is left; then summarise the decisions in at most five lines.

Discipline

  • Wait for the answer before the next question. One question with a recommendation beats four options.
  • A decision on record is cited, not asked. "ADR-0065 decided X — I build on that." Ask only when two records
    disagree or the record predates the change under discussion. A wrong assumption about the environment or an
    earlier decision is the most frequent correction in past sessions.
  • Stress-test with concrete scenarios. Invent the edge case that forces a precise boundary between two
    concepts.
  • Cross-check what the user says against the code. "The code does X, you said Y — which is right?"

Glossary

The glossary is the file the repo's instruction file names (for example GLOSSARY.md); only when none exists,
CONTEXT.md at the repo root, created with the first resolved term. Format for a new file:
${CLAUDE_PLUGIN_ROOT}/skills/grill/references/context-format.md; an existing glossary keeps its own format.

  • A term that conflicts with the glossary is called out at once: "The glossary defines X as A; you mean B?"
  • A vague or overloaded term gets a proposed canonical term and the words to avoid.
  • A resolved term is written as soon as it is resolved, with a one-line note to the user. When you run inside
    /we:story, hand the lines to its step 5 commit instead.
  • The glossary holds definitions only: no implementation detail, no spec, no scratch notes.

ADRs — offer sparingly

Offer an ADR only when all three hold: hard to reverse, surprising without the context, a real trade-off between
genuine alternatives. Follow the repo's ADR directory: the newest ADR's file naming and its TEMPLATE.md if one
exists. Without either, one paragraph (context, decision, why) in docs/adr/NNNN-slug.md.

The summary

At most five lines, one decision each, with who decided. Inside /we:story the summary becomes part of the
refiner's brief.


Adapted from Matt Pocock's skills (MIT).