Skills

A skill is a reusable instruction file — a Markdown SKILL.md with a name, a description, and a procedure — that an agent can pull into a task on demand. Where an MCP server gives an agent new tools and data, a skill gives it know-how: how to run a disciplined code review, how to design an experiment, how to build a .docx from scratch. You invoke one in a conversation by typing / and picking it from the menu.

Skills are the agent's own feature — each supported agent reads them from its own skills directory. Codeg is the management layer on top, and it gives you two ways to put skills there:

• Skill packs — curated bundles Codeg ships (Experts, Science, Office). You enable a pack's skills for the agents you choose, and Codeg links them in.
• Your own skills — write a SKILL.md by hand for any agent, at global or per-project scope.

Either way the result is the same: the skill shows up in that agent's / menu, ready to invoke. The two live under adjacent Settings screens — Skill Packs and Skills — and the rest of this page covers each in turn.

Enable a curated skill pack

Open Settings → Skill Packs. It's a tabbed hub over three curated, Codeg-managed bundles, all sharing one central store (~/.codeg/skills) and one enabling model:

Pack	What's in it	Depth
Experts	A software-engineering workflow library (drawn from the open-source superpowers collection) — grouped by discovery, planning, execution, quality, debugging, review, and meta.	Ready out of the box
Science	A research toolkit — ideation, design, analysis, visualization, evaluation, and literature. → Scientific Research
Office	Document skills that create and edit .docx, .xlsx, and .pptx, powered by the bundled officecli tool. → Office Documents

The skill-and-agent matrix

Each tab shows a grid: skills down the side (grouped by category), your agents across the top. Tick a cell to enable that skill for that agent, untick to remove it. Everything is batchable:

• A skill's ⋯ menu enables or disables it across every agent at once.
• An agent's column header does the reverse — every skill for one agent.
• Select several skills (the checkboxes) to flip a chosen set for a chosen subset of agents, and the top-right Everything menu toggles the whole grid.
• Click a skill's name to open a preview drawer — its full SKILL.md, plus a per-agent switch list.

Enabling is instant and reversible; disabling in bulk asks first, since it's the destructive direction.

How enabling actually works

The central store is the single source of truth. When you enable a skill for an agent, Codeg drops a symlink (a junction on Windows) from that agent's skills directory into the central copy — so nothing is duplicated, and there's no separate database of what's on: the link either exists or it doesn't. The Open central folder button in the toolbar reveals ~/.codeg/skills if you want to look.

A cell isn't always a plain on/off. You may see:

• Blocked — a real folder of the same name already sits in the agent's skills directory, so Codeg won't overwrite it. Move or rename it, then enable.
• Linked elsewhere — that name is already a link pointing somewhere other than Codeg's store.
• Copy mode (amber ring) — the filesystem refused a symlink, so Codeg copied the files instead. It still works; it just won't auto-update in place.

Packs update themselves

Experts (and any bundled skill) are baked into Codeg and re-extracted to the central store when you upgrade — so enabled skills stay current with no action from you. Edit a bundled skill in place and Codeg flags it user modified and preserves your version. Science skills may carry a needs key or needs setup badge — they'll link fine, but the skill itself needs an API key or a Python environment before it runs; see the skill's own instructions and where credentials live.

Write your own skill

For a skill that isn't in any pack — an internal convention, a workflow you keep repeating — open Settings → Skills and author it directly.

Pick a managed target (only agents that actually support skills appear in the dropdown), then choose a scope:

• Global — the agent's home skills directory (Claude Code's is ~/.claude/skills), available in every folder.
• Folder — a project-scoped directory ({folder}/.claude/skills) that only shows up when you're working in that folder. Pick the folder from the list of ones you've opened in Codeg.

Click New Skill, give it an id (which becomes the on-disk file or {id}/SKILL.md directory, and the /-name you'll invoke), and write the body. The editor is split: a Markdown source pane and a live preview that renders the content and pulls out the YAML front-matter fields. A SKILL.md is just front-matter plus a procedure:

---
name: pr-review
description: Review a pull request for correctness, tests, and style
---

When asked to review a PR:
1. Read the diff and summarize the intent in one line.
2. Check each changed file for logic errors and missing tests.
3. Flag anything that breaks the project's conventions.

Save, and it lands right beside the pack-installed skills. From the list you can preview, edit, open the folder it lives in, or delete it. Skills shipped as part of the system are marked read-only — you can read them but not change them. Because you're writing straight into the agent's own skills directory, this is the agent's native skill format; Codeg just gives you an editor and a preview for it.

Invoke a skill in a conversation

Enabled skills surface in the composer the moment you need them. Type / and they appear in the autocomplete right next to slash commands — pick one to drop it into your message. (Codex uses $ instead of /, matching its own convention.) The + menu also carries shortcuts for the three skill families — Experts, Office, and Research — so you can reach a common skill without remembering its name.

Two things to know:

• A skill that isn't enabled for the current agent shows as locked — enable it for that agent first (in Skill Packs or Skills), and it'll light up.
• Folder-scoped skills you wrote only appear while you're working in their folder; global ones are always there.

Good to know

• Skills are per agent, not per Codeg. Enabling one for Claude Code doesn't enable it for Codex — that's the whole point of the matrix. Turn on the same skill for as many agents as you like.
• Changes apply on the next session. A skill is read when the agent starts a conversation; for one that's already open, Reconnect to apply picks it up without losing history. → Configure an agent
• Disabling is safe. It only removes the link from that one agent — the central copy, and every other agent's link, stay put. Removing a bundled skill entirely takes it out of all agents at once.
• It works the same on a server. Both Settings screens are in the browser build; skills live on the machine Codeg runs on.
• A custom Pi agent directory opts out. If you've pointed Pi at a non-default PI_CODING_AGENT_DIR, it isn't managed by the shared store and won't appear as a target. → How agents differ

Next steps

• Office Documents — put the Office pack to work: build and edit .docx, .xlsx, and .pptx with live preview.
• Scientific Research — the Science pack in depth, from hypothesis to write-up.
• MCP Servers — the other way to extend an agent: give it new tools and data sources.
• Multi-Agent Collaboration — distinct from the Experts pack: one agent delegating whole sub-tasks to another agent type.