Grill-me Skill: Review Assumptions before Building
An idea may feel ready to build while its users, scope, and untested assumptions remain unclear. Grill-me uses related questions to expose those decisions before implementation.

grill-me uses questions to uncover decisions that remain open
Matt Pocock’s grill-me examines plans, designs, and ideas to uncover options and assumptions. Users can start even if they don't have a completed plan yet. Users should review the reasons for each decision on their own. Instead of agreeing to the plan that the agent wrote all for you.
Several developers publish skills named grill-me. This article references mattpocock/skills commit 5beb35b0 whose grill-me/SKILL.md is the entry point forwarding to grilling and disables automatic model invocation. Interview details reside in the subsequently invoked skill, so read both files for behavior not fully described in the short listing.
The reviewed version does not write files or save an interview workspace, and it does not require a repository. It suits broad topics such as product scope or article direction. The developer provides grill-with-docs for work that needs codebase inspection, a glossary, or decision records. Keep these skills’ behavior separate.
How grill-me uses earlier answers to choose its next question
Some decisions depend on earlier answers. For example, you need to know who the users are before choosing what data to collect, and whether data will be stored before setting a retention period. The skill calls the set of questions ready for a decision in the current round the frontier.
The cited grilling version asks the full decision frontier in numbered rounds with proposals, then waits before computing the next round. Guides claiming grill-me always asks one at a time may describe another version. Users can request that style, but it is an added preference, not this original's default.
Arranging questions according to conditions helps to agree on assumptions before answering matters that depend on those assumptions. But the quality of the decision tree also depends on the knowledge of the model. The user must reframe the question when it misses an option or expands beyond the question.
Install both the entry point and its dependent skill
Skills.sh provides the entry-point command below. Node.js and an agent supporting Skills CLI are required. Review the listed skills and dependencies, and check that grilling is installed and available. If it is not added automatically, select it from the same repository. The entry-point file does not contain the full interview guidance.
npx skills add https://github.com/mattpocock/skills --skill grill-me
How you invoke the skill depends on the agent. If it supports commands named after skills, the maintainer lists the command below and suggests starting a fresh chat so the agent’s existing plan does not frame the discussion. This example follows the maintainer’s documentation.
/grill-me
Hypothetical request: use grill-me to review a small-team scheduling tool. Decide what version one will and will not do. Ask in Thai, give reasoned proposals, separate researchable facts from my decisions, and do not start implementation yet.
Hypothetical example: From simple tools to verifiable boundaries
| Round | Questions ready to ask | Unlocked results |
|---|---|---|
| Define the users | One user or several? Is access from other devices required? | Know whether sharing and user identities need to be considered. |
| Set data | For personal use, must data persist across days or be exportable? | You can control local storage and file formats |
| Define error | If you delete a table or open a broken file What does the user need to recover? | Obtain acceptance criteria and fallback scope |
KrobClick created this hypothetical conversation plan without a live skill interview. Answers depend on the project. Switching to multiple users reopens branches such as editing permissions and concurrent changes. Continuing unchanged questions would address the wrong requirements.
Do not answer “yes” to every criterion just because a proposal sounds confident. If something is unknown, say so and identify the evidence needed. Users own decisions; agents should investigate facts verifiable from documents or the environment, rather than turning every uncertainty into a question for the user.
Separate questions answerable through discussion from those needing actual inspection
The developer guide highlights questions discussion alone cannot settle, such as whether a long form works better on one page or several. Without seeing both, repeated questions may elicit guessed preferences. Build a small prototype or collect user evidence before deciding.
KrobClick suggests three statuses for each issue: decided with reasons, awaiting evidence, or deferred as out of scope. For repeated questions, identify missing evidence rather than adding detail without new knowledge. Ask whether each question enables a decision or more evidence. Conversation length and question count alone do not establish a complete plan.
Finish with decisions, unknowns, and acceptance criteria
Once aligned, request agreed decisions, reasons, rejected alternatives, unknowns, and acceptance criteria. Ask explicitly as a next task if these should be saved to a file: the referenced grill-me version does not create permanent files itself. Verify that the document is actually saved in the repository if it is needed beyond the conversation.
If you need to move to another session, read the Handoff guidance to pass only necessary context, evidence, and outstanding issues. Interview findings still require feasibility checks and real experiments. Start with a small idea whose users and problem you can explain, then use questions to uncover hidden decisions.
Sources
- mattpocock/skills: source pinned to a commit
- grill-me: trigger point forwarded to grilling
- Matt Pocock: Current edition of the grill-me manual.
Source reviewed October 21, 2026, pinned to cited commits. This explains documentation without installing or testing the skill. Thai examples are hypothetical KrobClick scenarios. The AI-generated cover is not output from the described skill.