▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=gaDdrDdczO4
• Original title :: New Skills! v1.2 brings /wait-what, /writing-for-agents, and fixes /grill-me

Matt Pocock walks through version 1.2.0 of his skills repo (about 204K GitHub stars). The release adds a documentation site at aihero.dev/skills, Claude Code plugin marketplace distribution, Codex compatibility, and five new or reworked skills: wait-what (plain-language rewrite of verbose model output), a graph-driven multi-question grill-me, writing-for-agents, wizard (interactive provisioning scripts) and to-questionnaire (turn a grilling session into a shareable document).

Documentation, marketplace and Codex

The skills repo now has proper docs at aihero.dev/skills. The main flow is grill-with-docs, then spec, ticket, implement, code review; a side panel references every skill, a FAQ is sourced from Matt's personal wiki of questions people ask, and terms link to an AI coding dictionary he is building.

The repo is now part of the official Claude Code plugin marketplace: search "Matt Pocock skills" under plugin and install a read-only bundle that updates automatically. For Codex, every skill now ships an openai.yaml sidecar so that "allow implicit invocation: false" (user-invoked skills hidden from the agent's context) works there too, which it did not before.

wait-what: taming verbose Opus output

Matt finds Opus 5 verbose, full of odd LLM phrasing and hard to read; output styles and agent.md tweaks did not help. The wait-what skill is what you invoke when you do not understand what the agent just said. It tells the model to use ASD-STE100 Simplified Technical English (clear declarative sentences) and to ground itself in the ubiquitous language from context.md. His view is that the real cure for verbosity is making the agent use your language, built during grill-with-docs, rather than just "simple language".

grill-me asks questions in rounds

Old grill-me asked one question per turn, which becomes slow and tedious at the end of a session when only easy questions remain. Now it asks several per turn, treating the questions as a dependency graph: each round contains only the questions currently unblocked, pushing down the frontier as fast as possible. Each question comes with a recommended answer and emoji markers for easy scanning, so with dictation he can reply "Q1 agree, Q2 agree, Q3 change this".

writing-for-agents

The former writing-great-skills skill is generalised, since Matt used it for anything an agent reads: agents.md, Claude.md, skills. Skill-specific mechanics live in a reference file. It is useful for pulling instructions out of agents.md into skills to avoid front-loading, and is model-invocable so it triggers whenever agent configuration is edited.

wizard

After painful AWS infrastructure provisioning, Matt built a skill that generates an interactive bash wizard walking the human through steps only they can do: it opens the right pages, guides logins and settings, takes pasted API keys and saves them to files and GitHub secrets. Being a deterministic script, no secrets go to an agent or Anthropic. He chose this over computer-use automation to keep control while making the process "weirdly joyful".

to-questionnaire

A plainly named skill that extracts the decisions from a grilling session into a document to walk through with another person. Matt used it while planning a garden office, sending a Google Doc to his wife and feeding the answers back to the agent. He hopes to delete it eventually, as it patches the weak collaboration story of agents (better: an agent in Slack that stakeholders can tag), but many people are not yet in that world.

Wrap up

Smaller changes are in the v1.2.0 GitHub release notes. Matt announces a cheaper, self-paced AI Coding Crash Course (waitlist open, due in a couple of weeks), reasoning that AI coding has settled enough since last December to support a stable foundation.