reusable instruction packages

A skill is a folder with a SKILL.md that teaches the agent how to do a particular kind of work — a house style, a workflow, a checklist. DGC ships a collection of 22 coding workflows. You can add project or personal packages, including portable .agents/skills directories.

• /skills — browse instructions, source, diagnostics, and enabled state. The editor's Create / install action creates or copies a package; the terminal uses a.
• Type $ or / after your draft text to choose skills. The editor attaches removable chips without replacing the draft; the terminal inserts an exact $name mention.
• Explicit selections load the full instructions for native and delegated subscription turns. Up to eight selections share a bounded context allowance. Disabled or removed selections are rejected before model execution. Scripts are never executed on install.
• The standard and full tool profiles advertise every enabled skill that permits implicit invocation; the adaptive profile advertises only skills that narrowly match the request. Explicit-only skills remain available through $name. The user request and normal permissions always take precedence.
• Project skills override personal skills, which override built-ins. Metadata refreshes at the next turn; use Reload to update the picker after external edits.
• dgc-design ships by default but stays dormant for normal coding — artifact frontend work activates it automatically. It preserves the requested brand and theme; DGC's purple palette is for DGC-branded or otherwise unbranded standalone DGC artifacts.

• browser-test — real browser flows, failure recovery, responsive and keyboard checks.
• fix-ci — inspect the failed CI job and commit, reproduce the cause and verify a fix.
• pr-feedback — evaluate review comments against current code and address justified changes.
• skill-author — create focused, portable skill packages with meaningful validation.
• mcp-builder — implement and test the tools/resources/prompts an MCP server advertises.
• Also included: batch, code-review, dataviz, debug, deep-research, dgc-design, handoff, loop, onboard, plan, refactor, security-review, setup, ship, ui-review, verify, write-tests.

These packages provide instructions, not tools or account access. Browser work needs an available browser harness; remote CI/PR work needs the appropriate client and account; MCP development uses the project's SDK. Missing capabilities are reported. Skills do not install dependencies, change permissions, publish or send messages merely by being selected. For example: Check failed-save recovery $browser-test or Create a project API migration skill $skill-author.

A skill is a directory containing a SKILL.md: YAML frontmatter, then the instructions themselves.

~/.dgc/skills/commit/SKILL.md

---
name: commit
description: Write a conventional commit message for the staged changes
---

Read the staged diff with the native `git_diff` tool's staged view and write a single conventional
commit message for it. Focus (optional): $ARGUMENTS

Describe what changed and why, never how. No trailing period on the subject.

name is how you invoke it, description is what the agent matches against when deciding whether the skill applies. $ARGUMENTS receives the invocation request. Known metadata fields support plain, quoted, folded (>), and literal (|) scalar values. DGC does not execute YAML tags, aliases, or objects. The when field is not used.

Discovery is project-first, so a repo can override a personal skill of the same name:

• <project>/.dgc/skills/<name>/SKILL.md
• <project>/.agents/skills/<name>/SKILL.md
• ~/.dgc/skills/<name>/SKILL.md
• ~/.agents/skills/<name>/SKILL.md
• DGC's bundled skills

Optional agents/openai.yaml metadata can set interface.display_name, interface.short_description, interface.default_prompt, and policy.allow_implicit_invocation: false. Invalid optional metadata is diagnosed and fails closed to explicit invocation. Supporting files resolve relative to the skill directory and retain normal filesystem permissions.

A plugin is now the most common way a skill arrives: installing one copies its skills into the same discovery root, where they behave like any other. See Plugins, apps and MCP.

dgc skills list
dgc skills create review-api
dgc skills install ./team-skills/review-api
dgc skills disable review-api
dgc skills enable review-api
dgc skills show review-api

The same actions work through /skills ACTION in the terminal. Use --user with create/install for a personal package. A source outside the project requires --allow-external, or an explicit selection in the editor's folder dialog. The editor also supports /skills list|reload|show NAME|enable NAME|disable NAME. Creation produces an explicit-only scaffold; edit SKILL.md before using it.

Installation never overwrites or merges an existing destination. A local package can contain up to 128 files/directories, 4 MiB total, eight directory levels, and 512 KiB per supporting file. SKILL.md is limited to 64 KiB and 48,000 instruction characters. Symlinks, special files, .env files, and generated dependency directories are rejected. The terminal also supports raw-URL installation of a single SKILL.md; use local package installation when supporting files are required. Installed scripts retain private file permissions and can be run through their interpreter after normal approval.

Disablement is shared in DGC's configuration. The terminal's x disables a skill; it does not delete source files. /skills lists the built-ins as examples you can copy.