Skill
Supplemental decision-filtering policy for grill-with-docs. Use together with the primary skill when a user wants fewer, higher-value questions: automatically resolve routine decisions with the recommended default, reserve 5%-20% of decision points for major or genuinely controversial questions, and ask exactly one question at a time. This skill is not standalone and must not replace the primary skill's exploration, terminology, documentation, ADR, or decision-recording workflow.From its SKILL.md
npx -y skills add sinberCS/grill-with-docs-qa-filter --skill skillAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 26 days oldThe repository was created 26 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
3.9 KB, 693 tokens by cl100k_base, as published. Nobody here has run it
Grill with Docs: Q&A Filter
Keep the primary workflow in control
Use this Skill only as a question-selection layer for grill-with-docs.
- Keep
grill-with-docsas the primary workflow and source of truth. - Apply this Skill only while selecting and presenting decisions.
- Do not repeat or replace code exploration, domain analysis, terminology work, documentation, ADR creation, or session closure.
- Return every resolved decision and its rationale to the primary workflow for recording.
- Follow
grill-with-docswhenever instructions conflict outside the question-selection stage.
Classify genuine decisions
Use the decision points and evidence already produced by the primary workflow. Do not manufacture extra decisions to satisfy a quota.
Classify each decision as:
auto: adopt the recommended answer without interrupting the user.ask: present the decision because it is major or genuinely controversial.
Prefer ask when a decision:
- defines agent autonomy or human-control boundaries
- changes core business behavior or user experience
- affects permissions, safety, releases, irreversible data, or external side effects
- is expensive to reverse
- has multiple defensible business answers
- blocks several downstream decisions
Treat reversible defaults, naming, internal interfaces, component reuse, and ordinary implementation details as auto unless they materially change business behavior.
Enforce the question budget
Let N be the number of genuine decision points and Q the number classified as ask.
For N >= 5, keep:
Q >= ceil(N * 5%)Q <= floor(N * 20%)
For fewer than five decisions, ask at most one question and only when it is genuinely necessary.
Do not inflate N by splitting one decision into cosmetic sub-decisions.
For every auto decision:
- Adopt the recommended answer.
- Return the decision and rationale to
grill-with-docsfor unified recording. - Continue without asking the user.
Review only high-value questions
When the host permits subagents and the user has authorized delegation, use one or two independent, read-only reviewers immediately before presenting an ask decision. Do not use reviewers to replace the primary workflow or run a parallel interview.
Give every reviewer the same neutral evidence:
- the precise decision
- established facts and constraints
- concrete business scenarios
- available alternatives
Do not reveal the primary agent's recommendation before reviewers respond.
Ask reviewers to provide:
- whether the decision is truly major or controversial
- their recommended answer
- the strongest counterargument
- missing evidence or risks
Use the results as an objectivity check. The primary agent remains responsible for the recommendation and the user interaction. If reviewers agree that the decision is not major, downgrade it to auto.
Ask exactly one question
Present one decision per turn:
Question: <clear business decision>
Current state: <necessary facts and scenarios>
Recommendation: <recommended answer>
Trade-off: <main benefit, cost, and independent review signal>
Should we proceed with <recommended answer>?
Do not append secondary questions.
After the user answers:
- State the resolved decision briefly.
- Return it to
grill-with-docsfor recording. - Recalculate affected downstream recommendations.
- Continue with the next
askdecision.
What ships with it: 1 file
327 B alongside SKILL.md
agents/
- openai.yaml327 B