Dev Grill

Status — dormant (2026-08-10). This plugin is no longer part of the author's own daily stack, and no further development is planned for now. The reason is specific and worth stating: the two plugins that were meant to consume this engine — dev-spec-driven at planning time and dev-guardian at diff time — each carry the interrogation method inline and run fine without it, so the shared-engine indirection stopped earning its keep.

It still works. The method below is complete and the code is fixed as of v0.2.2. If you don't already run those two plugins, this stands on its own — it is the only member of the family that needs no MCP server, no toolchain, and no dependencies at all.

The AI grills your understanding instead of you reviewing its code.

Code is now cheap to generate and expensive to stay connected to. Reading 10k lines of AI-generated code a day is neither fun nor sustainable. dev-grill inverts the review: it interviews you, one decision-branch at a time, until you both reach shared understanding of what's being built and why each significant decision was made — then hands back a compact, language-agnostic summary you can drop into an RFC, a spec, or a PR gate.

It is the shared engine. Two adapters plug into it:

Where	Moment	Input	Output goes to
dev-grill (/grill)	any	plan / design / diff / module / incident	a file you can save
dev-spec-driven (/grill)	before code	requirements	folded into requirements.md as EARS
dev-guardian (/guardian-grill)	before merge	the diff/PR	a 🟢/🟡/🔴 understanding verdict

The interrogation method lives in one place (skills/dev-grill/SKILL.md). The adapters are thin: they only decide what to feed the engine and where the answer goes. Tune the method once, everything improves.

What it grills

Domain-significant decisions only — business-rule branches, validation & failure paths, in/out of scope, unit contracts (inputs/outputs), state changes and edge cases. It deliberately does not grill formatting or naming — that's what a quality gate / linter is for (dev-guardian).

Modes & depth

Five modes (it infers if you don't say):

• plan — grill an idea before code exists.
• design — pressure-test a written design/architecture.
• diff — confirm you understand the decisions a diff/PR introduced.
• onboarding — grill your understanding of an existing/inherited module.
• postmortem — grill your understanding of an incident's actual causal chain.

Depth: --quick (~3 questions, riskiest branch), default (5–15 adaptive), --deep (exhaustive).

Usage

/grill plan     let's build a notification service
/grill diff     # grills the domain decisions in your current git diff
/grill onboarding src/billing   # grills what you understand about inherited code
/grill diff --deep              # exhaustive
/grill          # infers the mode from what you paste

Every grill ends with a Shared Understanding summary — Purpose, a language-agnostic Contract, Domain decisions & rationale, Edge cases, out-of-scope, and a Decision-coverage map (✅/❓/⚠️) that makes the result auditable. You can pause a long grill and resume it later.

Works in every AI (no MCP needed)

dev-grill is a pure prompt engine, so it ports to any host as plain text — the method into the host's rules file, and /grill into its command surface. Supported: Claude Code, Cursor, Gemini CLI, GitHub Copilot / VS Code, Windsurf, OpenAI Codex, and any generic AGENTS.md reader (Cline, Roo, Zed, Continue).

Three ways to wire it up:

1. Auto-setup (easiest). Install in Claude Code; the first time you open a project it asks once whether to wire /grill into the other AIs, and does it for you on yes. Writes nothing without your consent; disable with DEV_GRILL_AUTOSETUP=off. Or trigger it any time with /grill-setup.
2. Installer CLI. node cli/dev-grill.mjs install all --project /path/to/repo (idempotent, reversible, --dry-run to preview). node cli/dev-grill.mjs list shows every host.
3. By hand. Copy the files from host-rules/ — see INTEGRATIONS.md.

Install

Drop this folder next to your other plugins and enable it as a Claude Code plugin. It has no MCP server and no dependencies — it's a prompt engine (skills + commands) plus a dependency-free installer and a fail-open auto-setup hook.

Design notes

• One question at a time. Always. A questionnaire dump defeats the purpose.
• Language-agnostic output. The summary captures inputs/outputs/validation/domain rules, not implementation — so the same understanding can seed Go, Rust, or TypeScript.
• An honest 🟡 beats a false 🟢. If you can't explain a branch, that's a finding, not a failure.
• Trilingual. Responds in your language (EN/PT/ES).

MIT · Carlos Pereira