Context-Extraction Skills — Making the Agent Interview You

Source: wiki synthesis: mattpocock-skills, claude-one-person-marketing-team, team-skills-plugin-marketplace, nine-ai-techniques-late-2026, new-rules-context-engineering-claude-5

There is a skill shape that inverts the usual direction. Most skills encode what the agent should do; this one encodes what the agent should ask — it interviews the human relentlessly before any work begins, and writes the answers down. The wiki has documented it since May 2026 as an engineering tool. In the week of 2026-08-19 it turned up in three more places, including a marketing workflow aimed at people who have never opened a terminal.

The interesting thing is not that four creators invented the same idea. It is that they did not — this is one skill diffusing across a domain boundary, and the receiving domain does not know where it came from.

Key Takeaways

• The canonical implementation is Matt Pocock’s, in mattpocock/skills (MIT): grill-me — “intensive planning interviews, forces decisions out of the human via Socratic pressure” — and its engineering sibling grill-with-docs, “deep planning that updates domain-language docs as it goes.” Both were already documented here in May 2026.
• The defining behaviour is relentlessness plus one-at-a-time. From the verbatim prompt already captured in this wiki: “Interview me relentlessly… walk down each branch of the design tree… for each question provide your recommended answer… ask one at a time… if a question can be answered by exploring the codebase, explore the codebase instead.” That last clause is the discipline — do not ask the human what you can go and find out.
• The leave-behind is what separates it from a conversation. The AI Daily Brief’s emphasis on grill-with-docs: it “writes the vocabulary and the hard decisions into your repo while it does” — “importantly, it is not just an interview, but an interview that comes with a leave-behind.” An interview that dies with the session solved nothing.
• It has now crossed into marketing, and the provenance was lost in transit. Nate Herk’s non-technical marketing build ships a skill called “Grill Me,” described as one that “relentlessly interviews the user about every aspect of a topic until you reach shared understanding,” distributed through his own community — with no attribution to Pocock in the video. ^[inferred — the descriptions match almost verbatim and the wiki’s 2026-06-05 note traces an earlier grill-me walkthrough to what appears to be the same creator community, but the video does not name a source, so shared origin is likely rather than established]
• It is now also recommended as the exemplar of skill design, by The AI Daily Brief pointing at aihero.dev/skills, and installed as third-party stock by at least one team-skills operator, who credits “Matt PCO’s grill me skill” — the auto-caption’s rendering of Pocock.
• The reason it spread is the same reason context scaffolds spread. Every playbook in this wiki says “give the agent your brand positioning / coding standards / definition of done.” None of them says what to do when you cannot write those down. This is that step, and it works for anyone who can answer a question even if they cannot author a document.

Why it crossed the boundary now

Two forces met.

From the tooling side: the file-backed context scaffold became the default. CLAUDE.md as a router into a context/ folder, progressive disclosure as the organising principle, skills as the unit of packaged procedure. All of these assume the context files exist. The scaffold’s quality ceiling became the quality of what you put in it.

From the practitioner side: the people now hitting that ceiling are not writers. Herk’s audience is explicitly non-technical marketers; the team-skills playbook’s is teammates in marketing and finance. Asking someone to author a positioning document, a voice guide, and a persona set as a precondition for using an agent is asking for the hardest deliverable first.

The interview skill resolves both. The agent supplies the structure and the persistence; the human supplies only answers, one at a time, which is a far lower-effort mode than authorship. The artifact appears as a by-product.

Which is why the engineering-to-marketing crossing is unsurprising in hindsight. grill-me was built to extract a spec from a developer who knew the system but had not written it down. Extracting brand positioning from a founder who knows their customer but has not written it down is structurally identical. The domain changed; the shape of the missing information did not.

What combining them enables

Read together, the four instances describe a build that no single one states.

Stage 1 — extract. Run the interview against the domain you cannot articulate. In engineering that is the plan or design: vocabulary, hard decisions, and the constraints obvious to you and invisible in the repo. In marketing it is the Three Ps — pain, person, promise — plus voice, objections, and what you have already tried.

Stage 2 — encode. The leave-behind becomes the context/ files that CLAUDE.md points at, which every later prompt inherits for free. From that point the prompts get short, because the context stopped being something you retype.

Stage 3 — survive the interview. Already solved and already recorded here: an hour-plus interview overflows context. The continuous-checkpoint variant writes after every answer into a brainstorms/ folder — one document per session holding the algorithm, the key decisions, a step-by-step Q&A log, and “open flags” for things you must go and get from other people. That turns the interview into a resumable artifact rather than a session you must finish in one sitting. Anyone adopting the pattern in a new domain should take this variant, not the original.

Stage 4 — distribute the interview, not just the answers. From the team-skills playbook: an interview skill that produces a brand-voice file is one any teammate can run for a new client, with structurally comparable outputs. The extraction is the reusable asset.

And there is a documented chain that strings these together — /grill-with-docs → /handoff → /prototype → /handoff back — recorded in this wiki as a real session loop, which is the same durable-state thesis as the checkpoint variant.

The bookend: interview before, footer after

A second pattern in the same batch is the mirror image of this one, and they compose.

	Interview skill	Self-improvement footer
When	Before the work	After every run
Extracts	Tacit knowledge the human already has	Corrections the human just made
Trigger	You cannot write the context down	A step failed, needed a workaround, or was rejected
Output	context/ files	An edit to the skill itself

Together they close a loop: the interview gets the agent to a competent first attempt, and the footer captures every correction after that. The team-skills playbook’s worked example — a teammate hits an image-upload error, finds the workaround, the skill updates, everyone receives it — is the second half working. The interview is what makes the first attempt worth correcting rather than restarting.

Honest limits

• Nobody has measured this. Every source reports it as valuable; none compares interviewed context against hand-written context on the same task.
• A relentless interviewer is a cost. The skill is designed to keep asking. The checkpoint variant exists precisely because sessions run over an hour, and none of the sources says when the interview should stop. An interview that outlasts the human’s patience yields worse context than a shorter one they completed.
• Garbage in still applies. It extracts what you believe, including what you believe wrongly. There is no external check on whether your positioning is correct — only on whether it is stated.
• This is diffusion, not convergence. Reading four sightings as four independent discoveries would overstate the evidence considerably. What the week actually shows is one well-designed skill escaping its origin domain — which is a claim about distribution, not about inevitability.

Try It

1. Start from the canonical version. mattpocock/skills is MIT-licensed; take grill-me or grill-with-docs rather than writing your own from the descriptions.
2. Take the checkpoint variant with it. Write after every answer into a brainstorms/ folder, with an “open flags” list for anything you must go and ask someone else. Without this, a long interview is a session you cannot resume.
3. Pick the context file you have been avoiding writing. Brand voice, definition of done, the persona, the architecture decision — the one that has been a to-do for a month. That is the interview’s first target.
4. Insist on the leave-behind by name. “Write the vocabulary and the hard decisions to context/<topic>.md.” Without this instruction the session was a conversation.
5. Keep the explore-first clause when you adapt it. “If a question can be answered by exploring the codebase, explore the codebase instead” generalises directly — if it can be answered by reading the site, the CRM, or last quarter’s campaigns, it should not be asked.
6. Run it before kickoff, not after the first bad output. The failure this fixes is invisible at prompt time and obvious in the deliverable.
7. Then attach the self-improvement footer to the skills that consume that context, so corrections accumulate instead of repeating.

Related

• Matt Pocock’s Skills Repo — the canonical grill-me and grill-with-docs, the verbatim prompt, and the continuous-checkpoint variant
• Turn Claude Into a One-Person Marketing Team — the pattern arriving in a non-technical marketing workflow
• Distributing Team Skills as a Plugin Marketplace — the distribution layer, and the self-improvement footer
• Nine AI Techniques Worth Trying — grill-with-docs named as the exemplar of skill design
• New Rules of Context Engineering — why the scaffold this feeds is shaped the way it is
• CLAUDE.md Primer — the router the extracted files hang off
• Agent Skills Overview — authoring constraints for the skill itself
• R.I.T(E) Framework — the single-prompt ancestor of the same “get the requirements out of the human” move
• Cross-Topic Connections — topic index

Open Questions

• Is Herk’s “Grill Me” Pocock’s grill-me? The descriptions match closely and this wiki’s 2026-06-05 note traces an earlier grill-me walkthrough to what appears to be the same creator community, but the video credits no source. Checking the distributed markdown file against mattpocock/skills would settle it in a minute, and matters for whether “independent adoption” or “redistribution without attribution” is the right description.
• Does an interviewed context file actually outperform a hand-written one? The obvious A/B — same task, same agent, interviewed versus authored context — has not been run by anyone in these sources.
• When should the interview stop? No source specifies a termination condition beyond “shared understanding,” which the agent judges.
• Is there a domain where this fails? Every instance is a soft-context domain — brand, plan, design. Whether it helps for context that is already written down somewhere (an API contract, a schema) is untested and probably no; the explore-first clause suggests the skill’s own authors think so.