New Skills! v1.2 brings /wait-what, /writing-for-agents, and fixes /grill-me

Docs site and Claude Code marketplace

Source statementthe skills repo is mature enough to need real documentation, so Matt built a site with his team. He notes it was once his full-time job to write docs. The best entry point is the /skills page and its groupings; the left panel is a full reference of every skill. There is a common-questions page fed by a personal wiki of questions people actually ask him, and the skills link into an AI coding dictionary. [00:18-01:18]

Source statementthe second big change is an official Claude Code plugin marketplace listing. You open Claude Code, run the plugin command, search "Matt Pocock" or "Matt Pocock skills," and install from there. The result is a read-only bundle with no extra steps, and any updates Matt makes are pulled down automatically. [01:18-01:43]

Codex compatibility: openai.yaml sidecars

Source statementCodex users previously could not benefit from user-invoked skills. The value of a user-invoked skill (seen in the SKILL.md frontmatter) is that it stays hidden from the agent's context window until you invoke it. That works for Claude Code, pi, and a couple of other harnesses, but not Codex. v1.2.0 ships an openai.yaml sidecar for every skill so that allow_implicit_invocation: false travels to Codex too, and the skills work out of the box in Codex UI. [01:43-02:32]

Inferencethe sidecar is a portability shim. The same "hide until invoked" intent is expressed once per harness format (Claude's SKILL.md frontmatter, Codex's openai.yaml) so one skill behaves the same across both.

wait-what: a fix for Opus verbosity

Source statementMatt finds Opus, especially Opus 5, "talking garbage": very verbose, using weird LLM phrases, hard to read, and many people feel the same. Output styles and agents.md tweaks did not fix it. What he needed was a way to say "wait, what did you just say?" You use the skill whenever the agent produces something incomprehensible, and it replies with a much better alternative. [02:32-04:01]

Source statementthe real cure for verbosity is not "use simple language," it is "use your language," the ubiquitous vocabulary you built up in grill-with-docs. The simplified-English instruction is only part of it. [03:20-04:01]

grill-me: multi-question rounds

Source statementgrill-me is Matt's most popular skill. It used to grill one question per turn (question, answer, question, answer). That has a bad failure mode: near the end, the hard questions are done and you are left clicking "yes, sounds good" through a slow tail of easy ones. The fix he wanted was multiple questions per turn, so you can answer a batch of easy ones at once. [04:01-04:32]

Source statementthe problem with a naive batch is that question three may depend on the answer to question one. The solution: treat the questions as a graph. A critical question at the start opens up a set of others, which may open still more. grill-me now follows this graph, asks only the questions answerable now, and asks them in rounds, always pushing you down the frontier as fast as possible. [04:32-05:40]

writing-for-agents

Source statementMatt broadened his "writing great skills" skill because he was using it for anything an agent reads, not just skills, to make the doc more concise and the agent's output more predictable. It is now writing-for-agents: use it for agents.md, claude.md, or when creating or editing skills. It has extra skill-mechanics stapled on in a reference file. Crucially, it helps pull content out of agents.md into skills to stop "frontloading" (over-stuffing the always-loaded config). It is model-invocable with a clean description, so it pulls itself in whenever an agents.md is being modified. [06:27-07:27]

wizard: a human-run provisioning wizard

Source statementMatt had to provision infrastructure and walk through AWS steps he hated. Rather than let the agent click through with computer use (which felt "icky" and gave up control), the wizard skill generates an interactive bash wizard that walks the human through the steps only they can perform. It is a deterministic script, so nothing is sent to an agent or to Anthropic. It sends you to the exact page, gets you to log in, has you paste API keys, saves them to the files it needs (including GitHub secrets), and walks staged UI until done. It turned provisioning from painful to "weirdly joyful." [07:27-08:58]

to-questionnaire

Source statementthe most boring skill name he has produced, but one he really needed. It takes the decisions inside a grilling session and pulls them into a document you can walk through with someone else. His example: a wayfinder session to design a garden office, where the person he actually needed to consult was his wife, who will use the space. He had the agent turn the grilling questions into a Markdown doc, put it in a Google Doc, shared it, walked through it together, then pulled the answers back into the agent. [08:58-10:28]

Source statementMatt hopes to delete this skill someday, because it patches the fact that agents are hard to collaborate with. The better setup, which some teams already have, is an agent in Slack that stakeholders can tag into a shared conversation. But many people are not there yet, so a document you can hand someone to comment on in Google Docs is genuinely useful now. [09:40-10:28]

Wrap and course

Source statementa few smaller changes are in the v1.2.0 GitHub release change log. Matt is launching an AI coding crash course: instead of a time-boxed group class run over a couple of weeks, it is a cheaper self-paced course available year-round. His reasoning is that AI coding has "settled down" with little fundamental change since roughly December, so a stable foundation is now teachable without going stale. A waitlist link is in the description. [10:28-11:37]

Tools and resources

Claims to verify

• The "24K stars, 24th most-starred repo of all time" claim is stated verbally and not shown on screen; verify against the GitHub repo before repeating it (the "of all time" ranking especially).
• The captured screenshots of SKILL.md and openai.yaml are from the video; the exact frontmatter keys (allow_implicit_invocation, disable-model-invocation) should be confirmed against the installed v1.2.0 skills before relying on them in your own authoring.
• Whether the marketplace bundle is strictly read-only and how updates are delivered (auto-pull cadence, pinning) is stated but not demonstrated; check the plugin docs before assuming you cannot edit a local copy.
• wait-what's effectiveness is anecdotal ("saves your ass a couple of times"); it is a prompt-only skill, so results will vary by model and context.
• The garbled ~02:50-03:05 passage in the auto-captions is Matt reading an example of Opus verbosity, not a real instruction. The video's own on-screen text is the reliable source for the skill bodies.

Glossary

• skill: A packaged, reusable instruction set an agent loads for a specific kind of task; the unit this whole plugin ships.
• SKILL.md: The Markdown file that defines a skill: frontmatter (name, description, invocation policy) plus the instruction body.
• frontmatter: The YAML block at the top of a SKILL.md that carries machine-read fields like name, description, and invocation flags.
• user-invoked skill: A skill hidden from the agent's context until the human explicitly calls it, keeping the context window clean.
• model-invoked skill: A skill the agent can pull in on its own when its description matches the task.
• disable-model-invocation: The SKILL.md frontmatter flag (true) that makes a skill user-invoked only.
• allow_implicit_invocation: The openai.yaml equivalent flag; false means the same "only when the human invokes it" behavior in Codex.
• openai.yaml sidecar: A per-skill file that restates the skill for Codex so Claude-only invocation rules carry over to that harness.
• harness: The runtime an agent runs in (Claude Code, Codex, pi, opencode); each reads skills in its own format.
• marketplace: The official Claude Code plugin registry you install skills from; the mattpocock-skills bundle now lives there.
• read-only bundle: An installed plugin you do not edit locally, so updates from the author apply cleanly.
• the main flow: Matt's core skill pipeline: grill-with-docs, to-spec, to-tickets, implement, tdd, code-review.
• grill-with-docs: The skill that interviews you about a plan and records the decisions, seeding the project's vocabulary.
• ubiquitous language: The shared project vocabulary (captured in CONTEXT.md) that both humans and the agent use; the real cure for vague output.
• CONTEXT.md: The file holding a project's ubiquitous language, which wait-what and others ground their output in.
• ticket: A handoff artifact scoping one session of work; stands alone or hangs off a spec, and can block or be blocked by sibling tickets.
• spec: A written specification a grilling conversation is turned into, upstream of tickets.
• wayfinder: A slower, denser shaping skill that charts a large effort as a map of decisions; drops into a grilling session for the parts of the map that fit.
• smart zone: The range in which an agent works well; a ticket is sized so a session stays in it rather than degrading before the work is done.
• ASD-STE100: Simplified Technical English, a controlled-language standard wait-what invokes to push clear, declarative sentences.
• question graph: The dependency structure grill-me now follows: answering one question unlocks the next set, asked in rounds.
• round: One batch of currently-answerable questions in the new grill-me; each round unlocks the next.
• frontier: The leading edge of still-answerable questions grill-me pushes you toward as fast as possible.
• recommended answer: A suggested response grill-me shows beside each question so you can accept or override quickly.
• context pointer: A reference an agent holds (a skill description, a line in AGENTS.md) whose wording decides how reliably the agent reaches the named material.
• frontloading: Over-stuffing an always-loaded config (AGENTS.md/CLAUDE.md) with content that should live in on-demand skills.
• SKILL-MECHANICS.md: The reference writing-for-agents reads for skill frontmatter, invocation choice, and router skills.
• wizard (skill): A skill that generates a deterministic interactive bash script walking a human through provision steps only they can perform.
• deterministic script: A plain script with no agent in the loop, so nothing is sent to a model; the wizard's output is one.
• GitHub secrets: Encrypted repository settings the wizard can write provisioned keys into.
• to-questionnaire: A skill that turns grilling decisions into a Markdown document a stakeholder fills in.
• the send: In to-questionnaire, the only thing the user is interviewed about: who the document goes to and what is needed back.