Skip to content

Agent Skill

MFS includes a reusable agent skill at:

skills/mfs/

The skill teaches an agent when to use MFS search, when to browse with bounded context, and when native shell tools are still the better choice.

What the skill contains

skills/mfs/
  SKILL.md
  references/
    command-reference.md
    workflow.md
    candidate-selection.md
  • SKILL.md is the entry point: trigger guidance, decision tree, anti-patterns.
  • command-reference.md explains command syntax and options.
  • workflow.md explains the search-plus-browse loop.
  • candidate-selection.md explains how to choose between close candidates.

The skill intentionally does not include scripts or platform-specific metadata. It only needs shell access to the mfs command.

Install for Codex

Codex reads user skills from $HOME/.agents/skills and repo-scoped skills from .agents/skills.

User-level install:

mkdir -p ~/.agents/skills
cp -R skills/mfs ~/.agents/skills/mfs

Repo-scoped install:

mkdir -p .agents/skills
cp -R skills/mfs .agents/skills/mfs

Then invoke it explicitly with $mfs, or let Codex choose it when a task matches the skill description.

Install for Claude Code

Claude Code supports local skills in a .claude/skills directory.

mkdir -p .claude/skills
cp -R skills/mfs .claude/skills/mfs

Use the skill when asking Claude Code to search, inspect, or verify local indexed files.

Use with other shell-based agents

For agents without native skill discovery, provide skills/mfs/SKILL.md as a system or developer instruction and make the files under skills/mfs/references/ available as supporting references.

The agent needs:

  • shell access
  • mfs on PATH
  • an indexed target folder, or permission to run mfs add <path>

Example prompts

Use the MFS skill to find where PDF conversion cache eviction is implemented.

Use MFS to answer this question from the local docs corpus. Verify the final
answer with file paths and line ranges.

Use the MFS skill to search memory logs for the prior decision about queue
payloads, then inspect the surrounding context.

Operating model

The skill follows the same model as MFS itself:

  • use semantic search to locate candidates
  • use progressive browse to understand structure
  • use line ranges to verify exact evidence
  • use native tools for exact literals and filename patterns

This keeps the agent from choosing between two poor extremes: reading whole files too early or trusting a single search chunk without context.