Skills and rules

A skill is a short Markdown instruction that FreeGent adds to the prompt only when it is relevant. This keeps the base system prompt small, which saves tokens on every request, and still lets the agent follow a detailed procedure for debugging, testing, PDFs and so on when the situation calls for it. A rule is the same kind of file, but a standing instruction rather than a procedure.

What is built in

• Skills such as debug, test, search, deep-research, failure-recovery, pdf, sql, image, audio, datetime and location.
• Rules for git, HTTP, JavaScript, documents, AST queries, file persistence and data fallbacks.
• In the repository's skills/ folder: commit, plan, research, tasks and verify-with-tests.

Switch them on and off in Settings → Skills, where you can also edit the triggers of each.

When a skill is added

Triggers are checked at the start of each turn, and some fire in the middle of a turn after a tool result.

Trigger	Fires when
trigger	A whole-word match of a keyword in your message (sql does not match mysql)
trigger_on_filetype	The message mentions, or the project contains, a file with this extension
trigger_on_media	The message has an attachment of this type: image, audio, video, pdf or file
trigger_on_message_pattern	A regular expression matches the message
trigger_on_event	session-start (first message) or task-start (a task-completion request)
trigger_on_tool	A tool ran and its result contains a string or satisfies a comparison, for example execute_code=>exit_code!=0
trigger_on_failure	N failures of a kind (http_error, tool_error) in one turn
trigger_on_repeat, trigger_on_history_tool, trigger_on_turn, trigger_on_file_present, trigger_on_completion	A tool repeats N times, a tool appears in the history, the first turn, a path exists, or the model declares the work completed or blocked

A skill can also be restricted: exclude_roles stops it for given main roles, requires_tools skips it when a tool it needs is disabled, and roles injects its body into the prompts of workers with those roles regardless of triggers.

Always-on skills stay in the system prompt, which keeps the start of the prompt stable and cacheable. Triggered skills are prepended to your message as a guidance block, and mid-turn triggers arrive as a short note. A shared budget caps the total text injected, because everything that triggers costs context on every request.

Writing your own

A skill is skills/<name>/SKILL.md in the project, or created in Settings → Skills (a form, or Import from file). Front matter, then the body:

---
name: my-skill
description: One line shown in the Skills list and in the prompt
trigger: release notes, changelog
trigger_on_filetype: .md
---

## Steps

1. Read CHANGELOG.md.
2. …

• name must match the folder name. Everything after the front matter is the body, injected as written.
• Keep skills short. Prefer a specific trigger over session-start.
• Writing a file under skills/ reloads the list at once.