Skills system

Skills are self-contained folders that teach Seepient how to complete specific tasks. They package instructions, scripts, schemas, and reference material that load into the agent context on demand.

Skill folder layout

A skill lives in a directory containing at minimum a SKILL.md file:

• 📁 skills/my-code-reviewer/
  • 📄 SKILL.md
  • 📁 references/
    • 📄 style-guide.md
  • 📁 scripts/
    • 📄 check-syntax.sh

The SKILL.md file

The main instruction file uses markdown with YAML frontmatter:

---
name: code-reviewer
description: Review pull requests for architectural boundaries, test coverage, and security regressions.
---

# Code Reviewer Skill

When reviewing pull requests:
1. Verify that changes follow responsibility layers.
2. Confirm that tests accompany new logic.
3. Check for unhandled error cases.

Where skills live

Seepient searches for skills in three locations:

1. Workspace skills: .agents/skills/ or skills/ within your current project root.
2. User global skills: ~/.seepient/skills/ available across all workspaces on your machine.
3. Bundled skills: Built-in skills packaged directly with Seepient.

AGENTS.md standard

Seepient supports the AGENTS.md standard. When starting up, Seepient searches the repository hierarchy for AGENTS.md files:

• Project root AGENTS.md
• Subdirectory AGENTS.md files for package-specific rules in monorepos
• User global rules in ~/.seepient/AGENTS.md

Rules found in these files are parsed and assembled into the agent system prompt using a four-layer trust hierarchy:

1. System runtime invariants (immutable security policies)
2. Global user rules
3. Repository-level project instructions
4. Turn-level user instructions

Turn-scoped skill switching

Skills do not bloat the system prompt permanently. During execution:

1. Seepient registers only the names and descriptions of available skills in its initial context.
2. When the model decides it needs a skill (or when you invoke it with /skill name), Seepient loads the full SKILL.md content into that specific turn.
3. Once the task finishes, the heavy context unloads, keeping prompt token costs low and preventing unrelated instructions from polluting subsequent queries.

Ephemeral injection and mandatory skills

You can declare mandatory skills in your configuration or command-line flags:

seepient --skill code-reviewer "Review the changes in src/domain"

Mandatory skills load before any user prompt executes and remain pinned throughout the task.