| name | grill-with-docs-batch |
| description | Grouped grilling session that wraps grill-with-docs semantics while discussing related questions in batches, with explicit convergence checks so it stops when docs already contain enough decision quality. Use when the user wants to stress-test a plan against docs and domain language faster than one-question-at-a-time grilling, especially when asking whether remaining questions exist. |
Grill With Docs Batch
Challenge a plan against the repository's domain model and documentation like
grill-with-docs, but move in small coherent batches instead of one question at
a time.
First Decide Whether To Stop
Before asking any batch, run a saturation audit:
- Read the target plan/ADR/spec and the domain glossary or context file it
depends on.
- Check recent git history for the same target docs when available, but use it
only as supporting evidence. Commit count is not a stop condition by itself;
it helps find decisions that were already made and areas that may be getting
re-litigated.
- Separate durable decision questions from implementation defaults. Do not ask
the user to discuss things the implementation can decide locally without
changing public contracts, private-data boundaries, safety policy, cost/model
infrastructure, or irreversible file moves.
- If an ADR is accepted, a plan names what changes, what is out of scope, the
important boundaries, the acceptance gates, and the next execution step, and
local docs/code do not contradict it, answer that no more discussion is needed.
- Classify the documentation target before proposing updates: plans and ADRs
answer different questions, and some tasks legitimately need both.
When the user asks whether remaining questions exist, answer the yes/no first.
Only ask a batch if at least one unresolved question would materially change the
plan, contract, public/private boundary, or acceptance gate.
Convergence Loop
This skill should remove the burden of asking "are we done yet?" from the user.
After every accepted batch, do the loop yourself:
- Interpret the user's reply, including shorthand such as "LGTM", "all agree",
"others are fine", "2 no because...", or "do it".
- Apply the resolved updates to the plan, context doc, or ADR surface named in
the batch. If no document update is warranted, say why.
- Re-read the updated target and run a fresh saturation audit.
- Either ask the next material batch or stop with the no-more-discussion
response.
Do not wait for the user to ask whether there are more questions. The user
answers batches; the skill owns convergence.
When the previous assistant message offered exactly one next action, treat a
short approval reply such as "LGTM", "sounds good", "可以", or "do it" as
permission to perform that action. Examples:
- after a proposed batch, apply the accepted answers and continue the
convergence loop;
- after a stop response that recommends preflight, run or prepare
$intuitive-preflight instead of asking the user to restate it;
- after a stop response that recommends execution, hand off to
$intuitive-flow only when the plan/preflight contract is already approved.
If the user's short reply is ambiguous because multiple next actions were
listed, ask one concise clarification and then continue.
Plan vs ADR Routing
Do not treat docs/plans/* and docs/adr/* as interchangeable planning
surfaces.
Use a plan file for execution scope, non-goals, order, files, tests, gates, and
open implementation questions. Use an ADR only for durable decisions future
agents should not relitigate: public contracts, command surfaces, private-data
boundaries, safety policy, architecture layers, rejected alternatives, and
accepted consequences.
Some tasks need both. In that case, keep the ADR short and durable, then let the
plan reference the ADR while owning execution details. Do not create an ADR for
local implementation defaults, a checklist, progress notes, or a decision that
the current plan can reverse cheaply. Do not put phase checklists, verification
logs, or task sequencing into an ADR.
Before asking "should this be an ADR?", first state whether the current issue is
contract-shaped or execution-shaped. If it is execution-shaped, default to the
plan file. If it is contract-shaped but the exact public shape is not selected
yet, default to recording the current assumption in the plan and defer the ADR
until the public contract is chosen.
For new plan files, prefer the repo convention; otherwise use a date-prefixed
slug. Do not bulk-rename old plans only to add dates. Prefer concise plan
metadata over filename churn for active or recently reviewed plans.
Keep ADR numbering for durable decisions, but make the creation threshold strict.
Do not create ADRs for proof loops, reruns, local-dev evidence, benchmark runs,
one-off gates, phase checklists, task status, report wording, local artifact
regeneration, or reversible implementation details.
ADR And Plan Surface Cleanup
When the user's concern is that the repo has too many plans or ADRs, treat that
as a documentation-entropy problem before proposing another decision record.
Classify the existing files into:
- current execution plans;
- stale or superseded plans;
- durable ADRs future agents should obey;
- ADR-shaped execution records, proof logs, rerun notes, or status snapshots.
For plan or ADR overload, default to reversible organization: index/metadata
updates, archive moves for misfiled execution/proof records, and no deletion,
renumbering, gap filling, or broad filename churn unless explicitly accepted.
If the cleanup policy itself is unsettled, ask one focused batch. Once accepted,
implement the cleanup directly instead of repeatedly grilling the same
classification question.
Decision-Impact Test
Before asking a question, state why its answer matters. A question is worth
asking only when a concrete answer could change at least one of:
- the public API, MCP/tool contract, file layout, or command surface;
- private-data, safety, security, credential, cost, or external-infrastructure
boundaries;
- acceptance criteria, verification gates, rollout gates, or hard blockers;
- phase ordering, ownership, or whether a feature belongs in the current slice;
- glossary/ADR language whose meaning affects future implementation choices.
If the answer would only choose a local implementation default, test detail,
wording polish, or "nice to record" preference, do not ask. Pick the conservative
default, or patch the plan directly if the user asked to record defaults.
Core Rule
Keep the quality bar of grill-with-docs:
- Challenge vague or overloaded terms against
CONTEXT.md or CONTEXT-MAP.md.
- Explore code and docs instead of asking questions that local context can answer.
- Use concrete scenarios and edge cases to force precise boundaries.
- Update
CONTEXT.md inline as soon as glossary terms are resolved.
- Offer ADRs only when a decision is hard to reverse, surprising without
context, and the result of a real trade-off.
- Stop once the remaining items are implementation defaults, test details, or
local wording polish rather than unresolved product/domain decisions.
The behavioral changes are pacing and convergence: ask grouped questions when
the questions belong to the same decision layer, and stop when the remaining
items no longer need user decision.
Stop Conditions
Stop grilling and say so when any of these are true:
- The target docs let an implementer answer: what changes, what does not change,
what boundary must be protected, how it will be verified, and what the next
execution step is.
- Remaining candidate questions fail the Decision-Impact Test.
- Git history or prior discussion shows repeated refinement of the same area and
the new pass would not change what gets built, verified, or protected.
- The question would produce another planning-document edit but would not change
what gets built, verified, or protected.
- The user shows process-fatigue signals such as "we have done this multiple
times", "is this just not stopping", or "can we move on".
At a stop condition, answer directly that no more discussion is needed, name
what is already decided, list remaining implementation defaults, then include
Plan state:, Recommended next action:, and Shortcut:.
Do not convert implementation defaults into another batch. If the user wants the
defaults recorded, patch the plan directly with concise defaults instead of
asking more questions.
Batch Shape
Each batch should contain 3-6 tightly related questions. Use fewer when the
decision is risky or highly dependent.
For each batch, include theme, assumptions verified from docs/code, 3-6
numbered decision questions with recommended answers, update targets, and a
promise to apply accepted updates before rerunning the saturation audit.
Wait for the user's response before applying docs or moving to the next batch.
Accept shorthand answers such as "all agree", "1 yes, 2 no because...", or
"change 3 to...".
Do not impose a fixed batch limit on a first-pass grill of an unclear plan. After
each batch, re-run the saturation audit and either stop or explain which
Decision-Impact Test item justifies another batch. For a target that is already
accepted, repeatedly refined, or close to execution, default to zero or one batch
unless the user explicitly asks to keep exploring.
When To Fall Back To One Question
Ask one question at a time when:
- A term conflicts with the current glossary and affects every later question.
- The user's answer could materially change the batch structure.
- A decision touches public contracts, private data boundaries, safety policy,
security, irreversible file moves, or external paid/model infrastructure.
- Local docs/code contradict the user's premise.
Documentation Discipline
After each accepted batch:
- Apply only the resolved
CONTEXT.md glossary/relationship updates.
- Keep
CONTEXT.md free of implementation details, plans, and progress notes.
- If a plan update is warranted, keep it focused on scope, execution order,
acceptance gates, verification, and open implementation questions.
- If an ADR is warranted, create or update it separately with clear context,
decision, alternatives, and consequences. Link it from the plan when both
surfaces are needed.
- If ADR or plan cleanup is warranted, prefer archive moves plus README/index
updates over deletion, renumbering, or broad filename churn.
- Report exactly what changed, then run the saturation audit before asking any
next batch. If the audit finds no more decision-impact questions, stop and
recommend the next workflow step.
For plan-backed work, prefer updating the existing plan over scattering
resolved decisions through chat. Refresh any existing lifecycle header
concisely; do not add one for tiny local changes.
Language
Mirror the user's language for the discussion. Keep questions direct and include
your recommended answer for each question.