Skills

A skill is a Markdown file of instructions, such as "how we write database migrations" or "our API error format". CoderHelm reads skills from git repositories you connect, and the agents read the ones that matter for each task.

Open Skills to connect a repository and see every item it holds.

Repo layoutCopy Anchor Link

CoderHelm reads three kinds of item from a repository's default branch. Each can also sit under .claude/, so a repo you already use with Claude Code works as it is.

Nothing in a skill repo is ever run. Scripts next to a skill are read as text, never executed.

FrontmatterCopy Anchor Link

---
name: migrations
description: How to write a database migration in this codebase. Read before changing any table.
---

# Migrations

- One migration per pull request.
- Every migration has a down step...

Claude Code tool names map like this: Read → read files, Edit/MultiEdit → edit files, Write → write files, Grep → search code, Glob/LS → list files. Others (shell, web, notebooks, sub-agents) have no equivalent and are reported as missing.

LimitsCopy Anchor Link

Files next to a skill must be text (Markdown, JSON, YAML, SQL, code and similar). Anything skipped is listed with the reason.

Syncing and pinningCopy Anchor Link

A repository syncs when you connect it, on every push to its default branch, and when an admin clicks Sync now. Only merged code on the default branch is read; pushes to other branches are ignored. Each item records the commit it was read at and a hash of its content, so you can always see exactly which version an agent used.

The compatibility checkCopy Anchor Link

Skills written for Claude Code often mention tools CoderHelm doesn't have. After each sync, CoderHelm checks every step of every item against your team's real tools, connections and data sources, and marks each requirement:

• Available: CoderHelm can do it.
• Translated: it can, with a different tool (named).
• Missing: it needs something you haven't connected, with what to connect.
• Blocked: it needs to write somewhere no connected tool can.

The check runs again only when the text or your tools change. An admin can choose to use the translated version of an item.

Always applyCopy Anchor Link

By default agents see a list of skills and read the ones they need. An admin can mark a skill or command Always apply so it is given in full, choosing the repositories, the agents (planner, coder, reviewer, PR descriptions, review comments, replies, release notes) and file paths it applies to. Agents can also join reviews as steps (plan, code or prose).

AgentsCopy Anchor Link

A team agent answers questions from the main agents, in one turn. It only gets read-only tools, worked out from its own tools: list; write tools are always blocked. An agent with no tools: gets none. Admins can try an agent from its page under Skills.

Team toolsCopy Anchor Link

A team tool is a named, typed lookup, such as "find an order by id", that agents can call. Each tool reads one data source with a fixed query (SQL, GraphQL, a logs query or an HTTP GET), and its inputs are typed and checked before they are bound into the query, never pasted into it.

Find them under Data sources → Team tools. Define one in the dashboard, or commit it as .coderhelm/tools/<name>.yaml on a default branch; a pushed file becomes a proposal and is never switched on by itself.

name: order_by_id
use_when: You need an order's status and total from its id.
source: orders-db
kind: sql
read: SELECT id, status, total FROM orders WHERE id = :order_id
inputs:
  order_id:
    type: integer
    min: 1
    example: 1042
max_rows: 1

Every version runs once with its examples, and then another admin must approve that exact version (a team with one admin approves its own). Any edit makes a new version that needs approval again. A tool can also be set to ask me first, so each call waits for a person's yes (for up to ten minutes).