Hey I'm @mg. Here are a few skills for getting better work out of agents and better words out of them, too. These grew out of how I work with my agents. Take the ones that might help and give them a test run.
Check out the skills for why each one exists, when to use it, and what to expect. Here's the tl;dr:
Audit the instructions behind your agents for conflicts, stale workarounds, and accidental busywork. Especially useful with frontier models like GPT-6 Astra and Claude Fable, where older prompting tricks can get in the way.
🖍️ Try: Fire up a session with Astra or Fable, and tell them to run /redliner in a project, or globally across your system. And don't worry about wasting context, the skill has been designed to keep usage under control.
Help agents say more with less, without losing the plot. Agentish is a language style to use where agents are writing for other agents, e.g. AGENTS.md, CLAUDE.md, skills, and prompts.
✍🏻 Try: Ask your agent "How would you write @AGENTS.md using /agentish?"
When every detail is "load-bearing" and every insight is "hiding in plain sight," the agentisms have taken over. Cut the mannered prose while keeping the substance, technical detail, and your voice.
🤦🏻 Try: After seeing annoying outputs, just run /people-words and watch the agent speak more naturally.
Pull request titles and descriptions a reviewer can act on: what the change does, what happens today, what changes, and how to check it. Covers issue links for GitHub and Linear, before-and-after media, and everything that should stay out.
📬 Try: Before your agent opens a PR, tell it to use /pr-craft for the title and description.
npx skills
npx skills@latest add galligan/skillsPick the skills you want and the agents you want to use them with. To install one directly:
npx skills@latest add galligan/skills --skill redlinerOr browse the collection without installing anything:
npx skills@latest add galligan/skills --listThe installer copies ordinary skill files into your project. Pull later updates with npx skills update.
Codex or ChatGPT desktop
Use the built-in skill installer:
$skill-installer Install redliner from https://github.com/galligan/skills/tree/main/skills/redliner
Swap redliner for any of the other skill names.
Claude Code
Ask Claude Code to install the bundle in its native skills folder:
Install the `/redliner` skill from https://github.com/galligan/skills/tree/main/skills/redliner into `~/.claude/skills/redliner`. Include all supporting files without cloning the repo. Check with me before replacing an existing installation.
Then invoke /redliner. Use .claude/skills/redliner for a project-only install. Swap the skill name and paths to install another.
Each skill is a complete bundle. You don't need Bun or Skillset to use one. ChatGPT web and mobile use plugins; this collection currently ships standalone skills.
I built Skillset to manage agent instructions and skills, and use it to author and build this collection. Source files for the skills live in .skillset/; the complete, installable versions land in skills/.
Three references keep the instruction work grounded:
- Instruction Selection: does this directive earn its place?
- Instruction Placement: where should it live?
- Agentish: how do we make the intended behavior clear?
agentish uses these for writing; redliner uses them for review. Each gets its own bundled references, so you can install either on its own. people-words stays focused on expression.
We use the guidance here, too. Our generated AGENTS.md and .claude/CLAUDE.md route substantive instruction changes through a scoped Redliner review. The review notes and follow-up corrections show what the first passes found and missed.
Use Bun 1.4.0. Skillset is pinned to @skillset/cli@0.26.1.
bun install --frozen-lockfile --ignore-scripts
bun run skillset new skill my-skill --yesEdit .skillset/skills/my-skill/SKILL.md and keep its supporting files alongside it. Then rebuild and check the result:
bun run skillset change status
bun run build
bun run checkRecord meaningful changes through Skillset's change commands, using the scope reported by change status. Commit source, change records, and generated output together. The publishing guide covers change evidence, release commands, tests, and compiler details.
Edit the source and rebuild; let Skillset manage skills/, AGENTS.md, .claude/CLAUDE.md, and the locks. Building doesn't publish or install anything.
MIT. Use what helps, make it your own.