Quizzing

Quizzing is an agent skill + system for getting a design right on the first go — by front-loading an agent's clarifying questions into one convenient, well-designed questionnaire instead of a long back-and-forth chat.

It is the inverse (in cadence, not rigor) of the grill-me skill. Grilling interrogates you relentlessly, one question at a time. Quizzing keeps grilling's best principles but changes the interaction model.

How it works

1. Look up facts. The agent discovers everything it can from the environment (filesystem, tools) instead of asking.
2. Bearing phase (grilling-seeded). The agent asks 1–5 main "bearing" questions, grilling-style — one at a time, each with a recommended answer — to get direction.
3. Generate the form. From those bearings, the agent generates a tailored, capped questionnaire (20 questions by default, or the number you asked for) so the whole thing takes minutes, not 30.
4. Fill the web form. A clean, Material Design, dark-themed web form presents the questions one at a time (Next / Prev, progress indicator) so you can focus. Each question offers pre-made answers, ranges (1–5 / 1–10) for calibration, custom input, and examples/diagrams for context.
5. Persist as JSON. The form saves your answers as a JSON file — the canonical contract the skill reads to produce the design.
6. Confirm, then loop. The agent summarizes its understanding and waits for confirmation before acting. If you challenge the direction or more info is needed, it generates another targeted form and repeats until you're aligned. The next round is derived from the last one, not retyped: anything you deferred, every recommendation you overrode, and every trade-off you were warned about but left open comes back as a question, so nothing you asked to revisit gets quietly dropped.

Question types and checks

Quizzing supports several question types so each decision is asked in the clearest way. Any screen can carry a recommended answer, an example, and an inline diagram, and any question can invite free text alongside its options — so you are never forced to pick one of the choices offered. Because the questions are asked one at a time, the form also checks answers against each other: a combination that cannot work is caught and explained, and one that is merely risky can be accepted deliberately.

Install

Quizzing works with the GitHub Copilot CLI, Claude Code, and the Gemini CLI.

GitHub Copilot CLI

copilot plugin marketplace add Tranquility2/quizzing
copilot plugin install quizzing@quizzing

Claude Code

/plugin marketplace add Tranquility2/quizzing
/plugin install quizzing@quizzing

Then invoke it with /quizzing.

Gemini CLI

gemini extensions install https://github.com/Tranquility2/quizzing

Working from a clone or hacking on Quizzing itself? See the development guide for local install, the repository layout, and the make tasks.

Status

Early development. See docs/design.md for the full design, and docs/plan/architecture-quizzing-roadmap-1.md for the delivery ledger tracking what has actually shipped.

Design principles

• Fast: capped questions, recommended defaults, minutes not meetings.
• Focused: one question per screen.
• Decisions, not facts: only ask what the agent can't discover itself.
• JSON is the contract: the answers file drives everything downstream.
• Iterative: converge in a few short loops.
• Local & private: offline-capable, no third-party network calls.

Credits

Quizzing was inspired by grill-me by Matt Pocock (from mattpocock/skills) — "a relentless interview to sharpen a plan or design." grill-me has been heavily used and genuinely helpful, and it directly inspired this project. Quizzing keeps its core principles (look up facts, ask the questions that matter, recommend answers, don't act until confirmed) but flips the cadence: instead of a one-at-a-time interrogation, it front-loads the questions into a single, capped, convenient form. Huge thanks to Matt for the original idea. 🙏