/blog-post-GM — a Claude Code skill we evolved with our own Evolution engine to write every post in the Godmode voice.
What Are Claude Code Skills? The Complete Guide
📄 What: Markdown files that reprogram Claude Code's brain
📁 Where:
~/.claude/skills/🔒 Power move: "Try to write tests" → "NEVER complete without passing tests"
🏗️ Structure: Sequential phases + gates > flat rule lists
# My Coding Skill NEVER mark a task complete without passing tests. ALWAYS read existing code first. MUST follow camelCase for fns.
$ tree scripts/ scripts/ ├── audit.sh ├── score.mjs └── mutate.py
$ git log -1 SKILL.md commit a4f2c1 v3.4.1 — bumped tests finish-line gate
Claude Code out of the box is generic — it doesn't know your standards, your testing expectations, or your definition of "done." A skill is a Markdown instruction file that permanently changes how the AI thinks, plans, and works — think of it as a settings file for AI behavior.
How Skills Work Under the Hood
Claude Code loads information from your codebase, conversation history, and instruction files. Skills get pulled into this instruction layer through a reference (a line that points to the skill file) in your CLAUDE.md.
CLAUDE.md references skill file
↓
Skill loads into the AI's working memory
↓
Every response follows skill protocol
↓
Finish-line checklist checks quality before "done"
Instruction files can live in three places:
CLAUDE.mdin project root — project-level~/.claude/CLAUDE.md— global (all projects)CLAUDE.mdin subdirectories — scoped to that path
The AI treats skill instructions as hard rules, not suggestions. "NEVER mark a task complete without passing tests" gets followed almost every time. "Try to write tests" gets ignored half the time.
Skills vs. Prompts vs. Custom Instructions
| Method | Persistence | Enforcement |
|---|---|---|
| Single prompt | One response | Weak |
| Background instructions | One conversation | Medium |
| CLAUDE.md rules | Every session | Strong |
| Skill file | Every session | Very strong — structured steps with checklists |
A CLAUDE.md rule says "always write tests." A skill defines a step-by-step sequence where testing is one layer that can only run after earlier steps are done. The AI can't skip layers.
Anatomy of a Skill File
# My Coding Skill
## Context Protocol
Before writing ANY code, read all relevant files in full.
Map dependencies. Understand the existing patterns.
## Implementation Protocol
Follow existing code conventions exactly.
Use the same naming patterns, file structure, and abstractions.
## Testing Protocol
Write tests for every public function.
Cover: happy path, error cases, edge cases.
Run tests. If any fail, fix before completing.
## Completion Gate
NEVER mark a task as done unless:
- All tests pass
- No TypeScript/lint errors
- Code follows existing patternsSteps in order, with a finish-line checklist at the end. The AI treats this as a checklist and checks every item before saying it's done.
What Makes a Skill Effective
Absolute Language
NEVER, ALWAYS, MUST, REQUIRED. Softer words like "should" or "try" get treated as optional and ignored.
Step-by-Step Order
Structure as a sequence: read codebase, THEN plan, THEN implement, THEN test. Ordered steps beat unordered rules.
Finish-Line Checklists
Define exactly what "done" means. Without a checklist, the AI decides when to stop — usually too early. This is the single most impactful pattern.
Concrete Over Abstract
"Write good code" is useless. "camelCase for functions, PascalCase for classes, UPPER_SNAKE for constants" is actionable.
Negative Constraints
"NEVER use any as a TypeScript type." "NEVER leave a catch block empty." Clear boundaries the AI won't cross.
The Godmode Product Suite
Instead of writing your own skills from scratch, Godmode gives you battle-tested protocols that work out of the box. Three products, each building on the last:
An 8-step pipeline that runs every task. No skipping layers, no early "done" — read, plan, build, test, edge-case, secure, verify, document.
A skill that improves other skills. Forks a SKILL.md into variants, scores each on real tasks, keeps what wins. The original is never touched.
Fuses Godmode's 8-step execution with Evolution's scoring into a single loop: build → score → fix → re-score, until every check passes. One prompt in. Finished product out.
🛡️ Godmode — "Do it right the first time"
Think of default Claude Code like a talented freelancer who rushes. Godmode turns it into a disciplined team that follows a checklist before calling anything "done."
How it works, step by step:
↓
🏗️ 2. Plan architecture
↓
⚙️ 3. Build following existing patterns
↓
🧪 4. Write thorough tests
↓
🔍 5. Hunt edge cases
↓
🛡️ 6. Check security
↓
✅ 7. Verify everything runs
↓
📝 8. Document what and why
Eight layers, fixed order, no skipping. Claude can't jump to "done" without passing through every step. Read the full deep dive →
🧬 Evolution — "Make the skills themselves better"
Godmode gives Claude a great process. But what if the process itself could improve over time? Evolution is a skill that improves other skills — like hiring a coach for your coach.
How it works, step by step:
↓
🎯 Run it against real test tasks
↓
📊 Score it across 8 quality categories
↓
🔍 Find the weak spots
↓
🔧 Propose a fix (tested before applying)
↓
🔄 Repeat until the skill converges
Three modes: Evolve (keep improving in place), Mutate (fork a new version), or Splice (cross-breed two skills together). Original files are never touched. Read the full deep dive →
🚀 One-Shot Beta — "One prompt. Finished product."
Our most advanced skill. One-Shot takes Godmode's 8-layer execution and Evolution's quality scoring and fuses them into a single loop. You send one prompt, and it keeps working — building, scoring, fixing, re-scoring — until every quality check passes. No follow-ups needed.
How it works, step by step:
↓
⚙️ Full 8-layer execution runs
↓
📊 Output scored across 8 quality categories
↓
❌ Any category below threshold?
↓
🔧 Fix just the weak areas & re-score
↓
🔄 Loop until ALL categories pass
↓
✅ Deliver finished product
Verified at 0.920 across 1,000 honest test iterations and 60 benchmarks. The only skill that checks its own work and fixes it before you see it.
📊 How They Compare
| Godmode | Evolution | One-Shot Beta | |
|---|---|---|---|
| What it does | Makes Claude follow an 8-step process | Makes skill files improve over time | Executes + self-corrects in one shot |
| Analogy | 📋 A strict checklist | 🧬 A breeding program | 🤖 An autonomous engineer |
| You send | A task + it follows protocol | A skill file + it returns a better one | A task + it delivers a finished product |
| Follow-ups needed? | Sometimes — you review the output | Yes — you approve each improvement | No — it loops until quality passes |
| Best for | Consistent, thorough code every time | Making any skill measurably better | High-stakes tasks that need to be right |
Think of it like cooking: Godmode is a great recipe. Evolution is a system that taste-tests and improves the recipe over time. One-Shot is a chef who follows the recipe, tastes the dish, adjusts the seasoning, and keeps going until it's perfect — all before it reaches your table.
The free tier (Godmode Lite) includes 4 of the 8 layers — enough to see the difference. Paid tiers add the full protocol, Evolution, and One-Shot. See pricing.
Building Your Own Skills
- Start with your biggest pain point — what does Claude consistently skip or get wrong?
- Write a protocol, not a wish list — steps in order with a finish-line checklist
- Save as .md and reference from your
CLAUDE.md - Test on real tasks — if the AI skips something, strengthen the language ("should" becomes "MUST")
- Iterate — good skills are refined over dozens of sessions
Common Mistakes
Do
Focus on 5-10 rules that matter most
Use consistent, non-contradictory instructions
Define a clear finish-line checklist
Scope rules to the right language/context
Don't
Write 50 rules — they get diluted
Say "move fast" AND "be thorough"
Let the AI decide when it's done
Use TypeScript rules in a project with multiple programming languages without scoping
Try a Skill Right Now
Godmode Lite is a free, fully functional skill file. Download it, drop it in your project, and see the difference in one session.
Download Lite (Free) See Full TiersThe Bottom Line
Skills are the single highest-impact thing you can do with Claude Code. A well-written skill turns the AI from a reactive code generator into a disciplined engineering partner. Stop giving one-off instructions for every task — bake your expectations into reusable, structured rules that stick around forever.