| name | athena-forge |
| description | Use when creating new Athena skills, editing existing skills, or building skills for any Claude Code plugin — guides skill authoring with TDD validation |
Athena Forge — Skill Smith
You create and refine Claude Code skills. Every skill follows TDD: test first, write minimal, refactor until bulletproof.
The Iron Law
NO SKILL WITHOUT A FAILING TEST FIRST
Write skill before testing? Delete it. Start over. No exceptions.
Skill Format
Every Athena skill is a SKILL.md file in its own directory:
skills/
skill-name/
SKILL.md # Main content (required)
supporting-file.* # Only if needed (heavy reference, tools)
SKILL.md Template
---
name: skill-name
description: Use when [specific triggering conditions — symptoms, situations, contexts]
---
# Skill Name — Role Title
[1-2 sentence core principle. What does this skill enforce?]
## Process
### Phase/Step 1: [Name]
[Steps with code blocks, gates, output formats]
### Phase/Step 2: [Name]
[Continue...]
## Rules
- [Enforced constraints]
## Anti-Patterns (NEVER)
- [What NOT to do and why]
Frontmatter Rules
name: Letters, numbers, hyphens only. Verb-first: athena-build not build-athena
description: Starts with "Use when..." — describe TRIGGERING CONDITIONS only, not the workflow
- BAD: "Systematic debugging with hypothesis tracking and checkpoints"
- GOOD: "Use when encountering any bug, test failure, or unexpected behavior, before proposing fixes"
- Why: If description summarizes the workflow, Claude shortcuts to the description instead of reading the full skill
Athena Style Guide
Athena skills share these patterns:
| Element | Convention |
|---|
| Title | # Name — Role (e.g., "Athena Debug — Investigator") |
| Phases | ### Phase N: Name or ### Step N: Name |
| Gates | **GATE:** before critical decision points |
| Output | ASCII box templates with exact format |
| Rules | Bullet list of enforced constraints |
| Anti-Patterns | ## Anti-Patterns (NEVER) with what + why |
| Code | Complete and runnable, never pseudocode |
| File paths | Always exact (src/config.yaml not "the config file") |
TDD Cycle for Skills
RED: Baseline Test
- Write 2-3 pressure scenarios — situations where an agent might skip or shortcut the behavior your skill enforces
- Run scenarios with a subagent WITHOUT the skill
- Document exactly what went wrong:
- What choices did the agent make?
- What rationalizations did it use? (capture verbatim)
- Which pressures triggered violations?
GREEN: Write Minimal Skill
- Write the skill addressing those specific failures — not hypothetical ones
- Run the same scenarios WITH the skill loaded
- Agent should now comply
GATE: If the agent still violates, the skill isn't clear enough. Revise before proceeding.
REFACTOR: Close Loopholes
- Look for new rationalizations the agent found
- Add explicit counters in a rationalization table:
| Excuse | Reality |
|--------|---------|
| "Too simple to need this" | Simple tasks are where shortcuts cause the most waste |
| "I'll do it after" | After = never. The gate is absolute. |
- Add a Red Flags section listing thoughts that mean "stop, you're rationalizing"
- Re-test until the skill is bulletproof
Router Integration
If your new skill is a core Athena workflow, add it to the /athena router:
- Add a row to the routing table with intent, skill name, and trigger words
- Add the skill to the menu display
- Add context awareness rules (e.g., "if X state exists, route here")
Skill Types and Testing Approach
| Type | Examples | Test With |
|---|
| Discipline (rules) | TDD, verification gates | Pressure scenarios — time, sunk cost, exhaustion |
| Technique (how-to) | Debugging, planning | Application scenarios — can agent apply it? |
| Pattern (mental model) | Architecture, design | Recognition scenarios — when to apply / not apply? |
| Reference (docs) | API guides, syntax | Retrieval scenarios — can agent find the right info? |
Quality Checklist
Persuasion Principles
When writing discipline-enforcing skills, apply research-backed persuasion techniques to make them rationalization-resistant. See persuasion-principles.md in this skill's directory for the complete guide.
Quick reference:
- Discipline skills → Authority + Commitment + Social Proof (bright-line rules, rationalization tables)
- Technique skills → Moderate Authority + Unity (collaborative framing)
- Reference skills → Clarity only (no persuasion needed)
Anti-Patterns (NEVER)
- Writing skill before testing — you don't know what to teach if you haven't seen the failure
- Multi-language examples — one excellent example beats five mediocre ones
- Narrative storytelling — "In session X we found..." is not reusable. Teach the pattern.
- Summarizing workflow in description — Claude shortcuts descriptions, skips the full skill
- Batching multiple skills without testing each — deploy and test one at a time
- 500+ line skills — if it's that long, split into skill + reference file