SKILL.md Authoring

This skill enacts the Endogenous-First axiom from MANIFESTO.md: skills are re-encodings of documented procedures and patterns from AGENTS.md, not independent inventions.

Skills are tactical-layer knowledge artifacts that encode how a task is done. They sit beneath agents (who encode who does a task) and above session behavior.

1. File Naming and Location

• Location: .github/skills/<skill-name>/SKILL.md
• Slug format: kebab-case, lowercase, e.g. deep-research-sprint, pr-review-reply, session-management
• One skill per domain — do not combine multiple domain procedures into a single skill file

To create a new skill, create the directory .github/skills/<skill-name>/ and add a SKILL.md file. You can copy an existing SKILL.md as a starting point — for example:

cp .github/skills/source-caching/SKILL.md .github/skills/<skill-name>/SKILL.md

Then update the frontmatter and sections according to the conventions in this document.

2. YAML Frontmatter

Every skill file must open with YAML frontmatter:

---
name: <skill name in kebab-case>
description: |
  <Clear description of what this skill encodes — ≥25 characters>
  USE FOR: [bullet list of appropriate use cases]
  DO NOT USE FOR: [bullet list of what this skill explicitly excludes]
argument-hint: "usage hint shown in chat input box"
---

Field requirements:

Discipline fields (encode project governance):

---
name: deep-research-sprint
description: "Orchestrates ..."

# Optional — encode milestone/effort/applicability
tier: Foundation                          # Which milestone this skill applies to
type: research                            # Type: research | feature | scripting | automation | tooling | validation
effort: L                                 # Effort for practitioner to execute: s/m/l/xl
applies-to:                               # Which agent roles use this skill
  - Executive Researcher
  - Research Scout
status: active                            # Status: active | beta | deprecated | blocked
requires:                                 # Dependencies on other skills or tools
  - name: source-caching
    description: "Must cache sources before scouting"
---

3. Required Sections

Every skill file must contain at least one of each:

Discipline rule for Endogenous Sources section: Always encode:

1. The governing axiom this skill enacts (from MANIFESTO.md)
2. The pattern or GitHub issue this skill implements (e.g., #45 Research: Product Definition)
3. Which agents and tools depend on this skill
4. Citation chain to foundational documents

Example Beliefs & Context:

## Beliefs & Context

This skill enacts the *Endogenous-First* axiom from [`MANIFESTO.md`](../../../MANIFESTO.md) by encoding the complete research sprint workflow as a reusable procedure.

**Implements**: The research orchestration pattern from issue [#45 (Research: Product Definition)](https://github.com/EndogenAI/dogma/issues/45)

**Used by**:
- Executive Researcher (orchestrates fleet)
- Research Scout (gathers sources)
- Research Synthesizer (transforms findings)
- Research Reviewer (validates output)
- Research Archivist (commits deliverables)

**Foundation documents**:
- [`AGENTS.md`](../../../AGENTS.md) — governance constraints
- [`docs/guides/deep-research.md`](../../../docs/guides/deep-research.md) — full investigation methodology
- [`docs/research/methodology-review.md`](../../../docs/research/methodology/methodology-review.md) — prior art survey

4. Skill vs Agent — Composition Rule

When to create a skill (not an agent):

Extraction test: If a procedure in an agent body would benefit a different agent or tool without requiring that agent's posture or handoff logic, move it to a skill and reference it from both agents.

5. Link Path Rule

All links in skill files that exit .github/skills/<name>/ must use ../../../ relative paths. This is consistent with the agent file convention (../../ for depth-2 files) and ensures VS Code's prompts-diagnostics-provider can resolve them.

/-rooted paths (e.g. /AGENTS.md) resolve to the OS filesystem root on macOS, not the workspace root — they produce VS Code Problems panel errors.

.github/skills/my-skill/SKILL.md
  │
  ├─ ../              → .github/skills/   ← WRONG (wrong target)
  ├─ ../../           → .github/          ← WRONG (wrong target)
  ├─ /AGENTS.md       → OS filesystem root ← WRONG (resolves to /AGENTS.md on macOS)
  └─ ../../../        → (repo root)       ← CORRECT (depth 3 from repo root)

Correct usage for skills:

[`MANIFESTO.md`](../../../MANIFESTO.md)
[`AGENTS.md`](../../../AGENTS.md)
[`docs/guides/agents.md`](../../../docs/guides/agents.md)
[`.github/agents/README.md`](../../agents/README.md)

Within-directory links (to sibling files in .github/skills/<name>/) remain relative:

[`./other-file.md`](./other-file.md)   ← same directory: OK

This convention is enforced by the no-absolute-path-links-in-agent-files pre-commit hook.

6. Cross-Reference Density

Every skill file must contain at least one back-reference to the foundational document that governs it:

• Pure procedural skills → reference AGENTS.md
• Research-oriented skills → reference MANIFESTO.md (Endogenous-First axiom)
• Automation/scripting skills → reference AGENTS.md § Programmatic-First

Minimum pattern (place in Endogenous Sources):

This skill enacts the *<Axiom Name>* axiom from [`MANIFESTO.md`](../../../MANIFESTO.md).
It is governed by [`AGENTS.md`](../../../AGENTS.md) and associated procedural guidelines.

7. Encoding Inheritance Declaration

The skill's first substantive section must ground it in the encoding chain:

## Beliefs & Context

This skill enacts the *<Axiom>* axiom from [`MANIFESTO.md`](../../../MANIFESTO.md).

**Implements**: [Pattern or issue reference]

**Governed by**: [`AGENTS.md`](../../../AGENTS.md) § [relevant section]

Read [`docs/guides/agents.md`](../../../docs/guides/agents.md) before modifying this skill.

Governing axioms by skill type:

8. Completeness Template

Use this template structure for new skills:

---
name: <skill-name>
description: |
  <Clear description>
  
  USE FOR: [bullet list]
  DO NOT USE FOR: [bullet list]
argument-hint: "<usage hint>"

# Governance metadata
tier: <Foundation|Wave 1|Wave 2|...>
type: <research|feature|scripting|automation|tooling|validation>
effort: <s|m|l|xl>
applies-to:
  - <Agent Name>
status: active
---

# <Skill Title>

## Beliefs & Context

[Declare axioms, pattern, dependent agents, foundational docs]

---

## <Core Workflow/Procedure>

[Step-by-step instructions, code examples, decision trees]

---

## <Quality Gate/Completion Criteria>

[What constitutes successful execution]

---

## Guardrails

- [Never do X]
- [Always do Y]

9. No Heredocs in Skill Bodies

Like agents, skill files must not contain cat >> file << 'EOF' patterns or inline Python heredoc writes in example steps.

Instead: Instruct practitioners to use create_file (new files) or replace_string_in_file (edits) — both are safe for Markdown content.

10. Validate Before Committing

Before committing a new or modified skill:

uv run python scripts/validate_agent_files.py --skills
# To validate a single skill:
uv run python scripts/validate_agent_files.py .github/skills/<skill-name>/SKILL.md

Pre-commit checklist:

• ✅ Frontmatter has all required fields (name, description with USE FOR / DO NOT USE FOR)
• ✅ Optional governance fields present (tier, type, effort, applies-to, status)
• ✅ All repo-root paths use ../../../ prefix (not ../../)
• ✅ Endogenous Sources declares governing axiom
• ✅ Endogenous Sources references defining issue or pattern
• ✅ Completion Criteria section defines success
• ✅ No heredoc patterns in workflow steps
• ✅ All cross-references to foundational docs are correct links
• ✅ No orphaned or broken internal links

Guardrails

• Never use ../../ for repo-root references in skills — always ../../../ (one extra level up from agents)
• Never omit the Endogenous Sources section — encoding fidelity depends on it
• Never combine multiple domains in one skill — one skill per bounded domain
• Never create a skill without understanding the agent(s) that will use it — ask yourself: which agents will benefit from this?
• Do not add a skill without running the CI linter and passing all checks
• Do not introduce a procedure that is not grounded in an existing AGENTS.md principle or a prior-art pattern