Three skills
Each one is a folder with a SKILL.md in the open Agent Skills format. An agent loads a skill when your request matches it.
✎ eddie-review
The remark loop: list what is open, read each remark next to the lines it marks, edit, reply, resolve. It applies the editor's mechanical edits, places unmatched marks, and comes with a full command reference.
⚓ eddie-anchors
The rules for // eddie: marker lines. They keep every mark attached to its paragraph however much the prose is rewritten, so an agent must carry them through untouched.
⇄ eddie-pipeline
The round trip: map a newly returned annotated PDF, start a new round, export the report, stamp the answers into a fresh asciidoctor-pdf render, and gate CI on open remarks.
Install the skills
Eddie asks which tools you use, then writes the skills where each one looks.
# the CLI: npm install -g @eddie-doc/cli, or the copy bundled with the VS Code extension
$ eddie skills
Which AI coding tools do you use? Eddie's skills will be installed where each one looks.
1) Claude Code (found here)
2) OpenAI Codex
3) Cursor
4) Gemini CLI
5) GitHub Copilot
6) Other (any tool that reads .agents/skills: OpenCode, Amp, Kiro…)
Numbers separated by commas, or "a" for all [1]: 1,3
Install for:
p) this project — /work/book (commit them so collaborators get them)
u) your user account — /Users/me (every project on this machine)
p or u [p]: p
.claude/skills
installed eddie-review
installed eddie-anchors
installed eddie-pipeline
Or from where you already are
- VS Code: Eddie Doc: Install Agent Skills for AI Tools… (command palette or the review view's menu)
- Obsidian: Install agent skills for AI tools… (installs into the vault)
- Terminal UI: press
Sineddie tui - Scripts:
eddie skills install --tool claude,codex --scope project
Keeping them current
eddie skills listshows what is installed, where, and which version.- Run
eddie skillsagain after upgrading to update them. - A skill of the same name that you wrote yourself is never replaced or removed.
eddie skills removetakes Eddie's skills out again.
Where they go
Eddie writes the fewest copies that reach every tool you pick. Claude Code plus Cursor needs one copy; Claude Code plus Codex needs two.
| Tool | Project folder | For your account |
|---|---|---|
| Claude Code | .claude/skills | ~/.claude/skills |
| OpenAI Codex | .agents/skills | ~/.agents/skills |
| Cursor | .agents/skills (also reads .claude/, .cursor/, .codex/skills) | same under ~ |
| Gemini CLI | .agents/skills or .gemini/skills | same under ~ |
| GitHub Copilot | .agents/skills (also .github/, .claude/skills) | ~/.agents/skills or ~/.copilot/skills |
| Others (OpenCode, Amp, Kiro…) | .agents/skills | ~/.agents/skills |
Using Eddie with Claude Code
Claude Code picks up skills from .claude/skills and uses one when your request matches it.
- Map the editor's PDF, in the editor or the terminal:
eddie open ch04.adoc ~/Downloads/ch04-acme.pdf --origin Acme. You can also ask Claude to do it; theeddie-pipelineskill covers it. - Install the skills with
eddie skillsand pick Claude Code. Commit.claude/skillsso everyone on the book gets them. - Let it run
eddiewithout asking each time. Add this to.claude/settings.json:
Commands that change the manuscript in bulk ({ "permissions": { "allow": ["Bash(eddie *)"] } }apply) or delete review state still need--yes, and the skill tells Claude to ask you first. - Ask for the work in plain words:
manuscript/ch04.adoc. Fix what you agree with, reply to everything else, and leave anything that needs my decision open.~/Downloads/ch05-acme.pdf as a new round and tell me what changed since the last one.What Claude does
For the first prompt, Claude follows eddie-review. It runs eddie list ch04.adoc --bucket open,review --json and, for each remark, eddie show … --json, then reads the lines it points at and edits the prose, keeping every // eddie: line in place as eddie-anchors requires. It then runs eddie reply and eddie resolve. Its summary tells you what it resolved, what it answered but left open, and what it could not place. Each step lands in the review file, so the VS Code tree, the Obsidian panel and the stamped PDF all show it.
You can also invoke a skill directly, for example /eddie-review, or run Claude headless in a script: claude -p "Answer the open remarks on ch04.adoc".
eddie is on the
PATH of every integrated terminal and the review tree reloads whenever a review file
changes. Run Claude Code in that terminal and watch remarks move to Resolved
as it works.
Other coding agents
Every tool here loads the same skills. Only the folder differs, and eddie skills takes care of that.
◎ OpenAI Codex
Reads .agents/skills in the repository (and ~/.agents/skills). Run codex in the manuscript repo and ask it to work through the remarks; it picks the skill from its description, or you can name the skill in your prompt.
▣ Cursor
Its agent loads skills from .agents/skills, and also from .claude/skills, so one install serves Cursor and Claude Code. Ask in the agent chat; it runs eddie in Cursor's terminal.
✦ Gemini CLI
Reads .agents/skills or .gemini/skills, in the project or your home folder. Start gemini in the repo and ask as you would ask Claude.
⌘ GitHub Copilot
Agent mode in VS Code and the Copilot coding agent read .agents/skills, .github/skills or .claude/skills. For the cloud agent, install the CLI in .github/workflows/copilot-setup-steps.yml (npm install -g @eddie-doc/cli) so it has eddie.
… OpenCode, Amp, Kiro and others
Any agent that supports the Agent Skills standard reads .agents/skills. Choose Other in eddie skills.
›_ Agents without skills
Any agent that can run shell commands can still use Eddie. Point it at eddie help and --json, and paste the anchor rules from MANUSCRIPT-CLAUDE.md into its AGENTS.md.
Why it works well for agents
The CLI was built to be scripted, not just to be typed.
{ } JSON everywhere
Every reading command takes --json. Results go to stdout and messages to stderr. Exit codes: 0 done, 1 failed, 2 wrong usage.
# Stable remark numbers
A remark is #7 for good, across re-maps, reports and rounds. An agent and the editor refer to the same thing.
⚑ Safe by default
Bulk edits to the manuscript and deletions of review state ask first and need --yes when no one is at the terminal. --dry-run shows the plan.
⇆ One source of truth
The agent writes the same review files the editors read, so you can stop it at any point and carry on by hand in VS Code, Obsidian or eddie tui.