English · Deutsch
Let the agent cook. Grill only what matters.
A Codex skill for developers and project owners who want to stress-test a plan without answering every reversible decision themselves. Let Him Grill researches the project, resolves low-risk choices, and stops when goals, architecture, security, or cost need your judgment. The decision path stays inspectable and can be resumed later.
Project site · Reference artifacts · Changelog
Six decisions evaluated · five resolved autonomously · one human gate. Compare the initial state with the post-reassessment poster, read the recorded branch reassessment, or inspect a complete portable decision artifact.
The demo and poster were rendered from the bundled interface; see the recording provenance and sequence.
This is not a blanket continue autonomously instruction. Let Him Grill
defines a decision boundary, keeps the portable source of truth in
.grill/decisions.json, and invalidates dependent branches when an earlier
choice changes.
npx skills add pengusto/let-him-grill -g -a codex -yRequires Codex with skill support and Node.js with npx.
Then open a new Codex task in your project and try:
Use $let-him-grill in compact mode to stress-test this plan:
Add saved filters to the existing dashboard.
Reuse project patterns and stop when product scope needs my decision.
The skill checks project evidence first, explains its choices, and collects unresolved human gates. Before implementation, it asks you to confirm the planning summary. No additional server is needed.
| Mode | Best for | Output |
|---|---|---|
| Compact | Short, linear planning questions | Decisions and reasons in chat; files only when useful |
| Visual | Branching plans and later reassessment | Interactive tree and portable state in .grill/decisions.json |
Both modes use the same decision rules. Switch at any time.
An installable skill with published versions and recorded Codex installation and resume checks. This is a planning aid, not an enforcement mechanism for permissions: review provisional AI choices before implementation. The Python path provides executable state validation; the native fallback follows the same contract without that validation.
The validation record covers Claude Code package installation, but not an authenticated live invocation. The roadmap tracks remaining validation and launch work.
In five scripted paired planning runs, median time to a usable plan fell from 455 to 54 seconds. Let Him Grill's final plans surfaced seven normalized material human gates while asking one immediate question. See the protocol, raw transcripts, and limitations. The timing includes Codex execution and benchmark-controller latency; it is product evidence, not a controlled model-performance benchmark.
The clean-install validation records Codex discovery and resume behavior plus Claude package installation; live Claude invocation remains unverified.
That instruction tells an agent to keep going, but not where to stop or how to recover when an earlier choice changes. Let Him Grill makes those boundaries explicit: it classifies decisions, resolves only low-risk reversible options, stops at genuine Human-Gates, and can resume the portable state in a later task.
- researches repository code and documentation before asking questions
- recommends an answer for every real decision
- triages every option by fit, risk, effort, and reversibility
- continues through reversible, low-risk choices automatically
- stops at architecture, product, security, cost, and other human gates
- invalidates dependent decisions when an earlier choice changes
- supports compact text output and a persistent interactive decision tree
Use Git as a fallback when the skills CLI is unavailable. Both modes use the
same installation; choose the mode when invoking the skill.
Available in every Codex project for the current user:
mkdir -p ~/.agents/skills
git clone https://github.com/pengusto/let-him-grill.git \
~/.agents/skills/let-him-grillPowerShell:
New-Item -ItemType Directory -Force "$HOME\.agents\skills" | Out-Null
git clone https://github.com/pengusto/let-him-grill.git `
"$HOME\.agents\skills\let-him-grill"Version the skill with one repository:
mkdir -p .agents/skills
git submodule add https://github.com/pengusto/let-him-grill.git \
.agents/skills/let-him-grillStart a new Codex task after installation so the skill is discovered.
Text-first. State and visualization are created only when branching or revisiting decisions makes them useful.
Use $let-him-grill in compact mode to stress-test this plan.
Continue autonomously until my decision is required.
Persists decisions in .grill/decisions.json and shows the interactive tree at
human gates and after changes.
Use $let-him-grill in visual mode to stress-test this plan.
Continue autonomously until my decision is required.
Visual mode uses the Python standard-library backend when available and falls back to native Codex file and visualization tools otherwise. Both backends populate the same bundled HTML template, so the interface does not change with the renderer. To choose explicitly:
Use $let-him-grill in visual mode with the Python backend.
Use $let-him-grill in visual mode with the native Codex fallback. Do not use
Python or another runtime.
Codex announces Visual mode · Python backend or
Visual mode · Native Codex fallback before the first decision.
.grill/decisions.json is the portable source of truth. In a later task with
Let Him Grill installed, ask the agent to resume that file. The Python backend
can inspect the deterministic next action without changing state:
python3 <skill-dir>/scripts/decision_state.py resume .grill/decisions.jsonResume re-checks provisional AI choices against current repository evidence, reassesses contradicted or invalidated branches, and continues autonomously to the next human gate. HTML and Markdown exports are derived views and should be regenerated in the new workspace.
Use $let-him-grill in the best fitting mode to stress-test this
plan until my decision is required.
Codex chooses compact mode for short linear discussions and visual mode for branching or revisitable decisions. It states the selected mode once. You can switch modes at any time.
Three complete bundles include the starting prompt, portable JSON state, interactive tree, Markdown handoff, and a documented branch reassessment:
Six use cases with prompts and interface examples
Use $let-him-grill to choose a budgeting and reporting approach for our SaaS business. Stop before compliance or spending decisions.
Example decisions: financial priority, forecast cadence, and spending approvals.
Use $let-him-grill to decide whether this B2B product should start as a modular monolith or separate services. Stop at material scaling or ownership trade-offs.
Example decisions: system shape, API contracts, and delivery workflow.
Use $let-him-grill to plan a domain-model training workflow. Stop at privacy, licensing, or budget gates.
Example decisions: measurable objective, evaluation-data governance, and the first adaptation method to test.
Use $let-him-grill to shape the save system and multiplayer scope for this game prototype. Stop where platform, networking, or player-experience goals differ.
Example decisions: core player loop, save format, and multiplayer timing.
Use $let-him-grill to create a twelve-week language training plan. Continue until motivation, certification, or professional priorities need my judgment.
Example decisions: primary learning outcome, weekly practice rhythm, and correction timing during speaking practice.
Use $let-him-grill to choose deployment, authentication, backups, and observability for this internal portal. Stop before accepting security exposure or recurring cost.
Example decisions: deployment target, employee authentication, and recovery evidence required before launch.
At shared understanding, Codex summarizes confirmed human decisions, provisional AI choices, assumptions, remaining risks or blockers, and the ordered implementation plan. It asks for confirmation before implementation.
After confirmation, it updates an existing canonical planning, specification, or decision document when the repository already uses one or documentation was requested. It does not create a duplicate plan file by default.
- Codex with skill support
- Node.js with
npxfor the primary installation command - Git only for the manual installation fallback
- Python 3 recommended for deterministic visual state updates
- no virtual environment,
pip install, server, or network service
Let Him Grill extends the companion grilling and domain-modeling skills; it
does not vendor or modify them. The clean-install evidence verifies this
repository's discovery and portable-state path, not every live combination of
companion skills and host runtime.
Compact mode works without Python. The native visual fallback applies the same state and invalidation rules through Codex file tools, but does not have the Python backend's executable validation. Hosts without inline visualization support fall back to the same decision content as text.
For a global installation made with the skills CLI, rerun the installation command. Save any local customizations first; reinstalling replaces the installed skill files:
npx skills add pengusto/let-him-grill -g -a codex -yFor a global installation made with the manual Git clone:
git -C ~/.agents/skills/let-him-grill pull --ff-onlyProject-local submodule:
git submodule update --remote --merge \
.agents/skills/let-him-grillpython3 scripts/test_decision_state.pyThe state engine uses only the Python standard library. See the roadmap for launch and follow-up work. User-facing changes are tracked in the changelog.
If Let Him Grill helps you work through project decisions, you can support my open-source work. Your support helps me maintain the skill, improve validation, and document useful examples.
Inspired by Matt Pocock's Grill with Docs workflow. Let Him Grill is an independent project and is not affiliated with or endorsed by Matt Pocock or OpenAI.