Agent-Skills.md

Agent Skills: Grill With Docs

Extend grill-me by producing or updating a CONTEXT.md or ADR-style decision record while stress-testing a plan. Use when the user wants to be grilled and wants the shared understanding persisted as a living document.

user-invokedID: v1truv1us/ai-eng-system/grill-with-docs

Repository

v1truv1us/ai-eng-system
v1truv1usLicense: MIT
73

Install this agent skill to your local

pnpm dlx add-skill https://github.com/v1truv1us/ai-eng-system/tree/HEAD/skills/grill-with-docs

Skill Files

Browse the full folder contents for grill-with-docs.

Download Skill

Loading file tree…

skills/grill-with-docs/SKILL.md

Skill Metadata

Name
grill-with-docs
Description
Extend grill-me by producing or updating a CONTEXT.md or ADR-style decision record while stress-testing a plan. Use when the user wants to be grilled and wants the shared understanding persisted as a living document.

Default output: return only the result, blockers, and required evidence. Omit preambles, process narration, repeated context, confidence scores, and follow-up offers. Use at most five bullets unless a required artifact or schema needs more.

Grill With Docs

Overview

Run the same relentless, one-question-at-a-time interview as grill-me, but write down what you learn as you go. The output is a living CONTEXT.md or Architecture Decision Record (ADR) that captures the problem, the explored branches, the resolved decisions, and the remaining open questions.

When to Use

  • The user wants to be grilled and wants a document to hand off
  • A design review needs a durable record
  • A decision affects multiple teams or future maintainers
  • The team needs a single source of truth for why a path was chosen
  • The user explicitly says "grill me and write it up" or "grill with docs"

Trigger Phrases

  • "Grill me and document it"
  • "Grill this and produce a CONTEXT.md"
  • "Grill with docs"
  • "Stress-test this and write an ADR"
  • "Interview me and capture the decisions"

Execution Rules

  1. Run grill-me first. Ask one question at a time, provide recommended answers, explore the codebase when applicable, and resolve each branch before moving on.
  2. Write incrementally. After resolving each branch, append the decision to the running document. Do not wait until the end.
  3. Use the agreed format. Default to CONTEXT.md in the project root or beside the relevant code. Use the project's ADR template if one exists in docs/decisions/ or adr/.
  4. Capture decisions, not chat. The document should stand alone: state the question, the options considered, the chosen answer, the reasoning, and the constraints it imposes.
  5. Update, do not duplicate. If a CONTEXT.md or ADR already exists for this topic, update it. Do not create a parallel file unless the user asks for one.
  6. Close with open questions. List any parked branches or unresolved risks with an owner and a trigger for revisiting them.

Document Structure

Default CONTEXT.md outline:

  • Problem — What are we deciding and why does it matter?
  • Goals and Non-Goals — What is in scope and what is explicitly out?
  • Explored Branches — For each branch: question, options, recommendation, decision.
  • Dependencies — Which decisions block which others?
  • Risks and Mitigations — What could invalidate this plan?
  • Decided Approach — The agreed path in one paragraph.
  • Open Questions — Parked items with owners and revisit triggers.
  • References — Files, docs, tickets, or conversations cited during the grill.

Example Invocation Flow

User: "Grill with docs on the new caching layer."

Agent: "I'll interview you about the caching layer and write the shared understanding to CONTEXT.md as we go. First question: what user-visible problem must the caching layer solve? My recommendation: state it as a latency or cost metric with a threshold."

User: "Reduce p95 API latency from 800ms to under 200ms."

Agent: "Noted. I'll append that as the goal. Next question: which data is cacheable? I explored src/data/ and see mixed read/write patterns. I recommend starting with the read-heavy catalog endpoints. Accept?"

User: "Accept."

Agent: "Appended. Next branch: cache invalidation strategy..."

[After all branches are resolved]

Agent: "Interview complete. I wrote the decisions to docs/decisions/2026-06-17-caching-layer.md. The open questions are listed at the bottom with owners. Please review and edit the reasoning if I misrepresented anything."

Verification

  • [ ] The document reflects every resolved branch
  • [ ] Each decision includes the reasoning and a cited constraint
  • [ ] Open questions have an owner and a revisit trigger
  • [ ] The document location is discoverable from the relevant code or README
  • [ ] The user has reviewed and accepted the final write-up

Anti-Rationalization Table

| Excuse | Counter | |---|---| | "I'll write it up later" | Later rarely comes. Writing incrementally captures reasoning while it is fresh. | | "The chat log is enough" | Chat logs are not searchable or maintainable. A decision record outlives the session. | | "This doc will go stale" | A document with open questions and revisit triggers stays alive. A missing document is already stale. | | "We do not have an ADR template" | Use the default CONTEXT.md outline here. Consistency beats perfection. |