Skip to content
OpenAgentSkill
Browsing is open · Catalog captured 2026-11-21 · Automatic updates and account services are paused.
Development

grilling

When asked to grill, stress-test, or pressure-test an idea, or before a decision map or spec rests on an untested conversation, interview the decision maker in small rounds to expose assumptions and settle material choices.

timharris707 · Static checks on file

Installation

Read the source and dependencies before running a command. Copying does not install or execute anything.

npx skills add timharris707/skills --skill grilling

About this skill


name: grilling description: "When asked to grill, stress-test, or pressure-test an idea, or before a decision map or spec rests on an untested conversation, interview the decision maker in small rounds to expose assumptions and settle material choices."

Grilling

Read the Codex desktop binding when this workflow needs harness mechanics, model routing, or recovery.

A grilling session converts a half-formed idea into shared understanding by interviewing the decider until nothing load-bearing is still assumed. It is the pack's live, single-session instrument: no tickets, no map doc, no branch: the deliverable is agreement, reached out loud and then written down (see the closing record below).

Map the work as a design tree: every decision branches into the decisions that hang off it. You do not know the tree in advance; each answer grows it.

Rounds and the frontier

The frontier is every decision whose prerequisites are already settled: the questions askable now without guessing at answers you have not heard. Record the whole frontier, then ask it in manageable independent batches. Where the repo binds domain-memory, read the relevant decision records before composing the first round; a recorded decision is settled, not frontier, and reopens on new evidence, never on repetition.

The word means the same thing here as it does on the tracker, the edge of what is takeable, where decision-map works a frontier of gated tickets. A question whose answer depends on another question still open in this round belongs to a later round. Putting it in this round forces the decider to answer twice: once on a guess, once for real.

Each answer reshapes the tree. Settled decisions push the frontier outward and unblock what depended on them. Recompute and ask the next round.

How to ask a round

Use the available Codex question tool and its current schema. In Default mode, request_user_input_async can present choices and free-text questions while independent fact gathering continues. Put the recommendation first with a short reason. Use a Plan-only question tool only in Plan mode.

Ask a small, coherent batch of independent decisions. Keep a wide frontier in the working record and present the next manageable batch after answers arrive; do not dump every open question at once or assume Claude's four-card limit. Open questions can use the same tool without invented answer choices. If no question tool is available, use a concise numbered text round.

Use the user's recorded answers and decisions. A preselected choice, silence, or elapsed time does not answer a required question. Ask only questions whose prerequisites are settled; keep unrelated fact gathering moving while waiting.

Facts are yours, decisions are theirs

Finding facts is your job, never the decider's. When a frontier question needs something the environment already knows, go find it: what the code does, what the config says, what the vendor documents. Dispatch a subagent for anything that takes real digging, and follow the research skill's contract when the answer wants a durable citation.

Do not block on it. A running investigation is an unsettled prerequisite, so only the questions downstream of it wait. Ask the rest of the frontier now and fold the finding in when it lands.

The decisions are the decider's. Put each one to them and wait. A grilling agent that answers its own questions has produced nothing but a transcript of itself.

Where it sits in the pack

  • Before a decision map. Charting needs a destination and a first read on the fog; grilling produces both. Reach for decision-map when the answer is "this is too big for one sitting and the open questions gate each other"; grilling is the engine that surfaces that, not a replacement for it.
  • Before a spec. A conversation nobody pressure-tested makes a spec full of silent assumptions. Grill first, then write.
  • Beside prototype and research. When a question turns out to be "I need to see it in action," that is a prototype; when it is "somebody documented this already," that is research. Hand off and keep grilling the rest of the frontier.
  • Not the advisory board. advisory-board puts a finished artifact in front of several models for independent review. Grilling is one agent interviewing one human about something that is not finished yet. Grill to reach a position; convene the board to test one you already hold.

Close with a record

Agreement that lives only in the transcript evaporates with the session. Once the decider confirms shared understanding, write it down somewhere durable: a comment on the driving ticket where one exists, otherwise a dated summary in the repo's docs home (the binding doc names it). The record is short and it is not the interview: it enumerates the settled decisions, the design tree as visited, and marks each recommendation accepted or overridden. It is what a later spec or ticket names as its plan source, so to-tickets links a citable decision instead of a conversation nobody can reopen. Where the repo binds domain-memory, the close-record also mints memory: each settled decision becomes a decision record at the memory home and new or sharpened terms enter the glossary, the same store the pre-round read consults so a settled question is never re-asked.

Done when (checkable: verify each line before reporting complete)

  • The frontier is empty, and the closing record enumerates the settled decisions: the visited tree is readable from the record, with nothing load-bearing left assumed.
  • Every question that needed a fact got one you found, not a fact the decider was asked to supply.
  • Every rec
Read the complete source ↗