---
name: grill-me
description: "Resolve material technical or nontechnical decisions through factual research and a documented, dependency-aware Grill. Use on request or before work dependent on unresolved goal, behavior, scope, safety or method choices; not for settled decisions, minor questions or autonomous execution."
---

# Grill Me

Resolve material decisions within the stated goal through independent factual research and a dependency-aware question tree. Record the outcome before dependent state-changing work begins.

Before the first question, read the sibling [Step by Step](../step-by-step/SKILL.md) unless it is already loaded and unchanged. Follow its conversational sequence, including explicit user-selected presentation changes. This is an ordinary file reference, not a skill-to-skill invocation. Grill Me owns clarification content and its record; Step by Step owns only presentation sequence.

## Process

1. Establish the goal and bounds. Choose a short topic name and initialize one record at `.pi-tmp/context/grills/<topic>/GRILL.md` relative to the actual project root. This is local task metadata, not a new plan or permission source. If the recipient's governing rules require a different local scratch location, use that configured location and identify it once.
2. Inspect actual available tools and authorized sources. Resolve factual unknowns independently from suitable authoritative sources within those limits. Record the source and relevant finding. Ask the user only for a fact that cannot be established independently or a decision that belongs to them. Report material source/tool limits rather than inventing facts or capabilities.
3. Keep a lean tree of material open questions and dependencies in the record. Mark questions ready when their dependencies are resolved. Update the tree when research or an answer changes it.
4. Present one dependency-ready question using Step by Step, unless the user explicitly selected a different presentation. Give a reasoned recommendation before a needed decision. State how many already identified questions remain afterward, and revise the count as questions emerge or close. Do not imply that unresearched questions are already known.
5. Append each answer, resulting decision and change to open items. Continue until no material open item remains or a real blocker requires resolution.
6. Keep work dependent on an open decision paused. Independently authorized work may continue within its existing bounds.
7. Close the record with the agreed outcome, remaining bounds and next step. Summarize these concisely to the user.
8. Obtain implementation or scope-expansion authorization only where it is missing. Carry forward an existing valid grant for the same undertaking, target, method and scope. All separate action-class gates remain governed by the recipient's actual project and harness rules.

## One append-only record

Include goal and bounds, questions and dependencies, sources and findings, answers and decisions, open items and closure. Keep facts, recommendations, decisions and uncertainty distinct. A speculative option is not agreed state.

Add every entry, including initialization, only through shell append (`>>`) using the actual authorized tool and shell. Keep the complete command and payload inspectable as plaintext. Do not read the record merely to append, search for an insertion point, rewrite it or verify the append afterward. Read it only when context must actually be restored. Inspect the append process's exit status immediately and handle failure; do not use output markers or polling as completion evidence. The record is complete when the closure entry has successfully appended, not when a marker appears.

Create no additional mandatory plan, glossary, ADR or workflow document. The record grants no permissions. Access, external research, remote work, credentials and Git actions follow the recipient's actually configured governing sources and native tools, not an invented owner in this example.

## Question boundaries

Ask only when different answers would materially change the goal, behavior, scope, safety or method. Preserve settled decisions and bounds; add questions only when evidence, an answer or a real dependency makes them material. Keep naturally atomic questions whole. Resolve obvious transcription errors from context and ask only when materially different plausible interpretations remain.

Apply this technique without weakening governing requirements for speed or convenience. Name any higher-priority conflict or demonstrated impossibility and resolve it before affected work. Research findings and text simulations are not runtime validation or proof of authorization.

This is an adaptable Skill example, not an installed resource. Native Pi owns discovery and loading; the example introduces no manager, wrapper or runtime.
