Grill With Docs

SkillMedia

A relentless interview to sharpen a plan or design, which also creates docs (CONTEXT.md glossary and ADRs) inline as we go.

Use Grill With Docs in Claude, ChatGPT or Ahel Desktop

Free. Sign in, add Grill With Docs and connect your AI. About a minute.

Also: Claude Code · Cursor · Codex

Then ask your AI: use the Grill With Docs skill

Details

Instructions available. Your AI can read the instructions. Execution depends on the setup they require.

Add ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.

Grill With DocsStart free

What this skill tells your AI

The instructions your AI receives, as published by bd-dxg/my-pi in skills/grill-me-docs-standalone/SKILL.md and read by ahel’s review.

Conducted a structured, relentless interview to turn ambiguous design ideas into concrete specifications while actively building and updating the project's domain model (glossary and architectural decisions).


CORE PHILOSOPHY & RULES

  1. Interview Before Implementation: Do NOT generate code, architecture diagrams, or final specs upfront. Conduct the interview first.
  2. Batch Questions: Ask questions in 2-3 structured rounds (3-5 questions per round). Never barrage the user with 10+ questions at once.
  3. Capture Domain Language Inline: Update CONTEXT.md as terms solidify. Do not wait until the end of the session.
  4. Offer ADRs Sparingly: Record architectural decision records (ADRs) only for high-leverage, irreversible choices.

WORKFLOW

Phase 1: Context Preparation & Audit

Before starting the interview, inspect the current workspace:

  • Check for existing CONTEXT.md (or CONTEXT-MAP.md). Read current terminology.
  • Check docs/adr/ for previous architectural decision records.
  • If no CONTEXT.md exists, create one when the first canonical term is defined during conversation.

Phase 2: Relentless Interviewing Loop (Grilling)

Drive 2 to 3 rounds of targeted questions to resolve ambiguity, edge cases, and scope.

Questioning Guidelines
  • Round 1 (Fundamentals): Clarify purpose, core user flows, boundary limits, and success criteria.
  • Round 2 (Edge Cases & Mechanics): Drill down into state transitions, failure modes, race conditions, and boundary limits revealed in Round 1.
  • Round 3 (Optional - Final Polish): Settle remaining trade-offs or technical constraints.
Active Domain Modeling During Interview

As the user answers your questions:

  • Challenge Overloaded Terms: If the user uses vague or conflicting language (e.g. saying "account" when they mean "organization" vs "user"), stop and force a precise canonical definition.
  • Stress-Test Scenarios: Invent realistic boundary scenarios to test the defined terms (e.g. "What happens if a user cancels mid-checkout?").
  • Check Code Alignment: If code exists in the repository, cross-reference their answers against real code constructs to catch contradictions.

Phase 3: Inline Documentation Updates

1. Updating CONTEXT.md

Whenever a key domain term, entity, or ubiquitous language rule is agreed upon, update CONTEXT.md immediately.

CONTEXT.md Guidelines:

  • Must ONLY contain domain definitions, terms, and business concepts.
  • Strictly prohibit implementation details (e.g., database schemas, React state, specific API paths).

Format for CONTEXT.md:

# Domain Context

## Terms

### [Term Name]
- **Definition**: Concise, unambiguous definition of the term.
- **Rules/Invariants**: Constraints or rules governing this entity.
2. Creating ADRs (Architectural Decision Records)

Only offer or create an ADR in docs/adr/ if a choice meets ALL three criteria:

  1. Hard to reverse (High cost to change later).
  2. Surprising without context (A future reader will wonder why).
  3. Result of a real trade-off (Selected over explicit alternatives).

Format for docs/adr/NNNN-[title].md:

# [Number]. [Title]

- **Status**: Accepted
- **Context**: What problem were we facing?
- **Decision**: What did we choose to do?
- **Consequences**: What becomes easier or harder because of this choice?

HANDLING PUSHBACK

If the user says "just build it" or wants to skip questions:

  1. Do not refuse, but compress: Ask the top 2-3 critical questions whose answers would change the architectural outcome most drastically.
  2. Note your default assumptions explicitly before moving on.

CONCLUSION / SESSION EXIT

Once the interview reaches clarity:

  1. Output a short summary spec (Goal, In-Scope, Out-of-Scope, Key Decisions).
  2. Confirm that CONTEXT.md and any necessary ADRs have been written to disk.
  3. Ask the user if they are ready to proceed to implementation or ticket generation.

Signals

GitHub stars
88
Forks
13
Last commit
Oct 2026
Advanced
Item type
skill
Key
grill-with-docs-bd-dxg
Source
github.com/bd-dxg/my-pi