Skills

Skills are reusable instruction packs. They are different from tools, agents, and project memory:

Surface	Best for	Loaded when
Agent	Role and authority for the whole turn	You select the agent
Skill	Reusable workflow guidance for a specific kind of task	You mention it or the agent discovers it
Project memory	Small durable facts about this workspace	Runtime injects saved memory
Tool	Runtime action such as read, patch, search, or MCP call	The model calls it and runtime approves it

Use skills when a workflow is useful across projects but should not apply to every turn.

Where skills live

Global skills:

~/.config/kodacode/skills/<skill-id>/SKILL.md

Project skills:

<workspace>/.kodacode/skills/<skill-id>/SKILL.md

Each skill directory needs a SKILL.md file. Keep the description short and specific, because KodaCode lists available skill metadata in the prompt without loading every full skill body.

Minimal skill

---
name: migration
description: Plan and implement small database migrations with tests.
---

Use this skill when a task changes schema, migrations, seed data, or persisted
records.

Workflow:

1. Inspect the existing migration framework and test setup.
2. Identify backward-compatibility constraints before editing.
3. Add the migration and the smallest useful test coverage.
4. Run the relevant migration or persistence tests before finishing.

Then mention it in the composer:

$migration add a nullable billing_status column and update the tests

Or use it from the one-shot CLI path:

kodacode --skill migration "add a nullable billing_status column and update the tests"

Mention syntax

Use either form:

$migration
${migration}

The mention selects that skill for the turn. This is the most predictable way to use skills because the full skill content is loaded only when requested.

Skill discovery tools

KodaCode also has search_skills and skill tools. Agents can use those tools to discover and load a relevant skill during a turn.

That is useful when you have a large catalog and do not know the exact skill name. It is less predictable than an explicit $skill mention because it gives the agent discretion to search and load guidance.

Good skill shape

Good skills are compact and operational:

one clear trigger
a short workflow
repo-agnostic guidance
references to scripts or templates when needed
explicit stop conditions for risky work

Avoid skills that duplicate your always-on repo instructions. Put permanent project policy in AGENTS.md instead.

Example prompts

Use a project-local skill for a release checklist:

$release prepare the 1.8.0 release notes from the current branch

Use a global skill for a repeated debugging workflow:

$perf-debug investigate why the dashboard query regressed after this branch

Let the agent discover a skill when you only know the domain:

Find the right skill for accessibility review, then audit the changed React components.

Cost behavior

Available skill metadata is cheap compared with loading full skill content. Explicit $skill mentions load the full SKILL.md for that turn, so long skills still have input-token cost.

Keep large reference material in normal files inside the skill directory and point to it from SKILL.md. The agent can read those files only when they are actually needed.