grill
Skill5 starsChanged 26 days agoMITView on GitHub
Pressure the facts behind a claim before a decision closes — counter-question with a concrete example, capped attempts, escape hatches, evidence hierarchy — and route what comes out to the loaded node's decisions or backlog. Manually triggered, and also the method any other interviewing tool cites by path.
90 other projects use a version of this.
Adopted by
Most recently taken up by
- incredible404-rgb/zstacktheir own version, first seen Nov 15, 2026
- oberemokis/skillstheir own version, first seen Nov 15, 2026
- ychampion/ystacktheir own version, first seen Nov 15, 2026
- crbnos/carbontheir own version, first seen Nov 15, 2026
- Yesterday-AI/ytstacktheir own version, first seen Nov 8, 2026
- therealtinhtute/mono-harnesstheir own version, first seen Nov 8, 2026
Used with
Projects with it also use write-skill, root-cause and 10 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
NoaiRoxSkillwrite-skill
Use when creating a new skill, editing an existing skill, or verifying a skill works before deployment.
Also in 5 projects that use this
Ingramradical235Skillroot-cause
Structured root cause analysis using WHERE (MECE decomposition) then WHY (causal structure diagrams) pipeline. Includes coin-flip detection, logical-leap detection, and vicious cycle intervention. Use when buffer enters YELLOW/RED, when metrics stall, or when Actions execute but don't produce results.
Also in 4 projects that use this
Neko-Catpital-LabsSkillprove-it
Shared evidence rule for investigating, diagnosing, debugging, proving, or explaining what happened or why for any problem or situation: do not assert something is fixed, working, passing, merged, or running unless you personally observed it this turn and can show what you saw; do not assert it was caused by X until a repro and controlled isolation prove that cause.
Also in 4 projects that use this
sahinSkillexecute
Also in 5 projects that use this
assisjpSkillfile-pr
Prepare and open or update an authorized pull request with a concise problem-first description and verifiable evidence. Use when the user asks to create, file, open, or update a PR. Do not use merely to review a diff or when publication is not authorized.
Also in 3 projects that use this
Arman-KudaibergenovSkillwrite-plan
Этот скилл MUST быть вызван когда есть design.md и нужно разбить доработку на атомарные задачи (2-5 мин) с точными путями файлов и критериями проверки в tasks.md. SHOULD также вызывать как промежуточный шаг между brainstorm и subagent-dev. Do NOT использовать для обсуждения идеи — используй brainstorm; для выполнения задач — используй subagent-dev.
Also in 3 projects that use this
Other versions90 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 90
- kirodotdev/KiroCrew★ 4.8ksame name
- crbnos/carbon★ 2.8ksame name
- JuliusBrussee/cavekit★ 1.1ksame name
- SethGammon/Citadel★ 927same name
- nagi-studio/skills★ 470same name
- KerberosClaw/kc_ai_skills★ 79same name
- alamops/skills★ 53same name
- Kanevry/session-orchestrator★ 53same name
- oakreader/oakreader★ 46same name
- manusco/resonance★ 38same name
80 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository12 other files it gives its agents.
- cpoSkill
- ctoSkill
- ideation-engineSkill
- insight-engineSkill
- market-researchSkill
- prdSkill
- product-builder-resolverSkill
- product-metricsSkill
- product-strategySkill
- synthetic-usersSkill
- ux-researchSkill
- visual-designerSkill
The file
We split this skill into the rules it gives an agent
4 rules. 0 are common, 4 are distinctive, and 0 are very specific to this project.
0
Common
Most of the 70 edited versions we read keep these.
4
Distinctive
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 87
Every question carries the answer the agent recommends, worded so that "yes" accepts it.
- Distinctive: line 102
If the environment offers helpers that can take a search, the search goes to one and the round continues with the questions that do not …
- Distinctive: line 113
Before the first round, the glossary is read if it exists.
- Distinctive: line 163
Every decision is recorded.
“Common” is measured: kept by at least 6 in 10 of the 70 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
grill — the pressure interview
Two faces. Invocable on its own: it pressures the facts of the loaded node and distributes what comes out without inventing a new destination. Referenceable method: any tool that interviews cites the section below instead of copying it — varying who it points at or which examples it uses is parameterization by whoever cites, never a second copy of the method.
Method
Cite this section by its heading (.agents/skills/grill/SKILL.md#method) from any tool that interviews. The
ten ingredients come in two groups. 1-5 are how a single fact is pressured. 6-10 are how the whole
interview runs. In the order they are applied:
1. Counter-question with a concrete example
Faced with an answer that cannot be falsified ("everybody liked it", "it was a success", "the team is clear on it"), the answer is neither accepted nor rephrased: a counter-question is returned asking for the fact behind it, with an example of what would count as a good answer.
Bad: "could you give more detail?"
Good: "'they liked it' by what measure — did they finish the task unassisted, did they say it, or did you infer it? A good answer names the data: '3 out of 5 finished without asking for help, 2 dropped out at step 3'."
2. Pressure capped at 1-2 attempts
At most two rounds of counter-question on the same fact. A third one gets no more information: it saturates the interviewee and the method stops being an interview and becomes an interrogation.
3. Escape hatches
Two ways out, neither of which is "keep insisting":
- Second resistance — the answer to the second counter-question still has no fact. The attempt is neither discarded nor forced into a third round: the gap is recorded, with who closes it, and you move on.
- Strong evidence up front — if the first answer already carries the fact (observed behavior, measured data), there is no counter-question to make: it is confirmed and you move on.
4. Evidence hierarchy
behavior > data > claim > assumption. Faced with two answers competing over the same fact, the one
higher in the hierarchy wins — an assumption does not weigh the same as observed behavior even when
both are stated with the same confidence.
5. Every gap is recorded, never filled in
A gap the pressure did not close is not completed with an inference so the document looks tidy. It is recorded as open, with who closes it. Filling it in silently is the same class of incomplete scan presented as complete that the rest of the system forbids.
6. Rounds over the decision tree
The decisions of the interview are ordered as a tree, by what each one depends on. Each round brings together, numbered, every question that can already be answered. A question that depends on another one still open in the same round goes to the next round. After each round, what became answerable is recomputed.
Format of a round (plain text, readable in a terminal without formatting):
Round 1
1. Who pays for the product, the person who uses it or their manager?
Recommended: the manager pays, the analyst uses it.
2. Does the first version include a free tier?
Recommended: no, a free tier is not part of the first version.The person answers with "yes" to take every recommendation, or by number to change some. A question number, the question and its recommendation are always on separate lines.
Bad: one long message with six questions, three of which depend on the answer to the first.
Good: round 1 asks who the buyer is and what they use today. Round 2, once those are answered, asks about pricing, which depends on both.
7. A recommended answer that "yes" accepts
Every question carries the answer the agent recommends, worded so that "yes" accepts it. Agreeing never requires answering "no" to anything.
Bad: "Recommended: don't limit it to one segment?" (agreeing takes a "no")
Good: "Recommended: limit the first version to one segment, the finance teams of mid-size companies."
8. Facts are looked up, decisions are asked
A fact the environment can answer (files, earlier research, the body repo, another tool) is looked up by the agent and never asked of the person. If what is found contradicts what the person said, the contradiction is shown to them. The counter-question of ingredient 1 stays for the facts only the person has. Decisions are always the person's.
If the environment offers helpers that can take a search, the search goes to one and the round continues with the questions that do not depend on that fact. If it offers none, the search happens before the round that needs it.
Bad: "How many customers does the research say you interviewed?" (the research file says it)
Good: "The research file says 8 interviews, 5 of them with finance leads. You mentioned 'a dozen conversations': do the extra four sit outside that file?"
9. The glossary is challenged and written as terms settle
Before the first round, the glossary is read if it exists. If the person uses a word that clashes with it, they are stopped right then, naming the definition that clashes. A vague word, or one with two meanings, is resolved by proposing the precise term and an edge case that tests it.
A term is written the moment the person accepts its exact text in a round, never at the end. The
glossary lives in context/glossary.md of the loaded product node; without a product node, in
product-context/glossary.md of the current folder. The file is created when the first term is
agreed. Each entry has the term, one or two sentences saying what it is, and the words to avoid. It
carries no implementation detail: a glossary is not a spec.
**Account** — The company that signs the contract and pays for the seats. _Avoid_: client, customer.Bad: the person says "customer" and the glossary defines it as the end user; the interview goes on.
Good: "The glossary defines Customer as the end user, and you just used it for the company that pays. Proposed term: Account, the company that signs. Edge case: a freelancer who pays for themselves, is that an Account? Recommended: yes."
10. The interview closes when nothing is open and the operator confirms
It ends when no answerable question is left unanswered and the person confirms that both understand the same thing. Before that confirmation, the only things written are glossary terms (ingredient 9), gaps (ingredient 5) and decisions; no deliverable is produced and nothing agreed is acted on.
Bad: the last question is answered and the agent starts writing the deliverable.
Good: "Nothing answerable is left open. Two gaps recorded. Do we understand the same thing: a per-team product for finance leads, sold to the manager? Yes to close."
Invocable face
It runs on its own over the node that is already loaded. It pressures its facts with the method above and distributes what comes out across the canonical files that already exist — it invents none:
| What comes out | Where it goes |
| A decision | A file of its own in decisions/ of the workspace, written by the session close with about: of the loaded node when it is a product, a course, a channel or an account |
| A gap | The backlog of the loaded initiative —or, when the loaded node is a product, a course, a channel or an account, of that node's standing initiative (capture --about)—, recorded as open with who closes it named in the text — the same free-text, verbatim line format the session close already uses |
| A term the operator accepted | The product glossary (context/glossary.md), at the destination and in the format of ingredient 9 |
| A gap, if the loaded product node declares its own open-questions layer | That layer (context/open-questions.md of the node), instead of the backlog |
The decision and gap destinations exist before this tool; the glossary is born when the first term
is agreed. If pressuring a fact were to lead to writing into a destination that is not in this table,
grill does not invent it: it says so and stops — that row is added by the operator, not by the session.
Key decisions
Every decision is recorded. A key decision is one that is also marked as key where it is recorded, so whoever opens the work next reads it before the rest instead of waiting for a search to find it. It is key only when all three conditions hold; two out of three is not key:
- Hard to reverse: undoing it costs real work, money or trust.
- Surprising without context: someone who arrives without this conversation would not guess it.
- Born of a real trade-off: a real alternative was on the table and was discarded.
Key: "The company sells B2B first" — undoing it means rewriting pricing and the pitch (1), a newcomer would assume B2C (2), and the B2C path was weighed and dropped (3).
Not key: "The weekly report goes out on Mondays" — trivial to change (fails 1), nobody is surprised by it (fails 2), no alternative was weighed (fails 3).
The mark is proposed by the agent and confirmed by the operator. Before marking a decision as key, show the operator the decision with the three conditions side by side, say which holds and which does not, and mark it only if they accept. A decision that fails one condition is not proposed: say which one fails and leave it unmarked.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

