# Grilling Session > Refine a topic or handoff one question at a time when explicitly requested or composed by SE. - Skill: `alemar11/grilling-session` (Agent Skill, multi-file: 3 files) - Install (CLI): `npx skillmds@latest add alemar11/grilling-session` - Raw SKILL.md: https://api.skillmd.com/api/skills/alemar11/grilling-session/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: alemar11 (https://skillmd.com/u/alemar11) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/alemar11/grilling-session --- # Grilling Session Follow the shared [execution scope](../../references/execution-scope.md) for standalone and composed invocation. ## Purpose and boundary Use `$se:grilling-session` to turn a topic, proposal, plan, or composed handoff into a clearer decision-ready brief through a demanding but constructive interview. Infer the topic from the invoking prompt or supplied handoff when it is clear; ask the user to choose only when multiple materially different topics remain. Grilling Session is conversational and read-only. It may inspect repository context, source, documentation, and other read-only evidence, but it never edits project files, persists the transcript, creates tasks, or delegates work. If the user asks to preserve an accepted rule or decision, return it as a durable-knowledge candidate for the caller or a separately authorized knowledge workflow. Do not capture it during the interview. Read [references/states.md](references/states.md) before interpreting workflow or result state. Read the shared [workflow-graph.md](../../references/workflow-graph.md) before using the registry below. ## Context use Start from the topic, conversation and caller evidence; no repository or knowledge-management skill is required. Callers own context preparation and subsequent work. Inspect read-only evidence when a question depends on a checkable fact, reusing supplied findings unless new evidence challenges them. Evidence grounds the interview but does not replace user intent. Disclose missing evidence and continue unaffected questions; block only when responsible questioning cannot continue. ## Interview contract - Ask exactly one question per turn. - Pair that question with one concrete recommended answer and a concise reason it is the best current default. End by asking the user to accept it or state what should change. - Make the recommendation falsifiable and specific enough to correct. Mark it provisional when evidence is incomplete; never hide uncertainty or present an unsupported preference as repository fact. - Ask the highest-leverage unanswered question first: desired outcome, user or actor, success boundary, invariant, non-goal, failure behavior, tradeoff, or evidence requirement. - Prefer concrete scenarios, counterexamples, and forced tradeoffs over broad invitations such as "tell me more." - Challenge contradictions, vague nouns, hidden assumptions, and solutions presented as requirements. Stay direct and constructive rather than adversarial or performative. - Do not ask for facts available in the repository or supplied handoff. - After each answer, update the working interpretation silently. Briefly expose a correction only when it changes the meaning of the next question. - Continue until no material ambiguity remains, the user asks to stop, or the session is blocked. Never choose a fixed question count. - Before declaring the brief refined, ask one final confirmation question that presents the compact interpretation and invites correction. Return composed results in the invoking conversation. ## Workflow graph The registry owns structural edges; Mermaid is its projection. Read [transition conditions](references/states.md#transition-conditions) for the canonical conditions governing these node contracts. | node_id | kind | purpose | entry_conditions | inputs | outputs | transitions | stop_if | side_effects | terminal_states | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | context-read | action | Use supplied context and inspect relevant evidence when needed. | explicit invocation or authorized parent handoff | topic, conversation, or supplied handoff | available context and evidence limitations | frame, blocked | essential evidence is unavailable and responsible questioning cannot continue | read, transient | none | | frame | decision | Infer the subject and select the highest-leverage ambiguity. | available context assessed | supplied brief and context evidence | working interpretation and next ambiguity | question, blocked | no coherent topic can be selected without unavailable user input | transient | none | | question | action | Ask exactly one focused question with a recommended answer, then incorporate the user's response. | one material ambiguity or final confirmation remains | working interpretation and latest user answer | recommendation, concise rationale, and updated interpretation or stop request | question, confirm, reported, blocked | required user input cannot be obtained | transient | none | | confirm | decision | Present the compact interpretation for final user confirmation. | no known material ambiguity remains | working interpretation | confirmation, correction, or stop request | question, complete, reported | none | transient | none | | complete | terminal | Return the user-confirmed refined handoff. | user confirms the compact interpretation | confirmed brief and evidence | refined handoff | none | terminal | none | complete | | reported | terminal | Return the best-supported handoff after the user stops questioning. | user asks to stop before confirmation | working interpretation and evidence | handoff with unconfirmed items | none | terminal | none | reported | | blocked | terminal | Report why responsible questioning or synthesis cannot continue. | required dependency, context, or input is unavailable | retained evidence and blocker | blocker and smallest recovery input | none | terminal | none | blocked | ~~~mermaid flowchart TD context-read --> frame --> question context-read --> blocked frame --> blocked question --> question question --> confirm question --> reported question --> blocked confirm --> question confirm --> complete confirm --> reported ~~~ ## Refined handoff On `complete`, return a compact Markdown handoff containing: - objective and intended user outcome; - confirmed scope, non-goals, constraints, and invariants; - accepted decisions and important terminology; - success and failure criteria; - evidence or validation expectations; - remaining assumptions, risks, and genuinely unresolved questions; - durable-knowledge candidates, if any, clearly marked as not captured. On `reported`, return the same shape using the best supported interpretation and label every unconfirmed item. On `blocked`, return the exact blocker and the smallest input needed to resume. Do not include the raw interview transcript.