Skills

A skill is procedural knowledge stored as a SKILL.md file: how to adapt a project into an MCP server, how a team performs a review, how a repository is released. Skills are loaded into the agent's context on demand rather than being always present.

Trust is earned

Accumulating a skill is not the same as improving. A skill may have been derived from one lucky run, may not apply to this repository, or may have started failing. CoreLayer therefore evaluates every skill before it may run without being asked:

Skill requested
  → evaluateSkillTrust(manifest)
      deprecated?                → suggest only
      autonomy: "suggest"?       → suggest only
      3 consecutive failures?    → demoted to suggest only
      never succeeded?           → suggest only
      otherwise                  → may run automatically

"suggest only" does not disable the skill. It stays available; the agent must request it explicitly.

Execution history

Every execution appends an outcome (success, error, runId, executedAt) to the skill's provenance, capped at 20 records. That history is what the trust rule reads, so a skill that starts failing loses autonomy without anyone editing it.

Format

skills/
└── my-skill/
    └── SKILL.md

# My Skill

Short description of what this skill does.

## When to use it

The conditions under which this skill applies.

## Steps

### Step 1: Gather context
### Step 2: Execute
### Step 3: Verify and report

Built-in skill

Authoring guidance

• State the conditions for use, not just the steps.
• Include a verification step; a skill with no checkable outcome cannot earn trust.
• Prefer deterministic steps (commands, files) over prose instructions.
• If the skill needs elevation, say so explicitly rather than requesting broad permissions.

Next Steps

• Tool Registry — how tools differ from skills
• Add a Custom Tool — build a deterministic tool instead
• Connect an MCP App — the MCP adapter skill in practice