Skill Design Patterns
Benchmarked against: Anthropic — Skills Best Practices Principle: Skills should be self-contained, focused, and token-efficient
Pattern 1: Decision framework
For skills that help agents make decisions:
## Decision criteria
| Factor | Weight | Evaluate |
|--------|--------|----------|
| Cost impact | High | What does each option cost? |
| Reversibility | High | Can we undo this? |
| Quality | Medium | Which option produces better results? |
## Decision process
1. List all options
2. Score each against criteria
3. Present recommendation with rationale
4. Wait for Captain approval
Used by: /66s-review, HITL decision format
Pattern 2: Step-by-step procedure
For skills that guide a specific workflow:
## Steps
### Step 1: Prepare
- Check prerequisites
- Gather required information
### Step 2: Execute
- Perform the main action
- Validate results
### Step 3: Verify
- Run checks
- Compare against expected output
### Step 4: Report
- Summarize what was done
- Note any issues
Used by: /create, /orchestrator
Pattern 3: Mode switch
For skills that change agent behavior:
## Activation
When this skill is active:
- Use [specific engine] for all tasks
- Limit output to [constraints]
- Prioritize [specific behavior]
## Deactivation
Return to normal when:
- Quota restored
- Captain says otherwise
Used by: /brain_lite, /brain_mid, /brain_pro
Pattern 4: Template generator
For skills that produce structured output:
## Template
Generate the following structure:
\```
[Section 1: Title]
Content...
[Section 2: Analysis]
| Column 1 | Column 2 |
|----------|----------|
| ... | ... |
[Section 3: Recommendations]
1. First recommendation
2. Second recommendation
\```
Used by: /ui, GSTA alignment
Anti-patterns
| Anti-pattern | Problem | Better approach |
|---|---|---|
| Kitchen sink skill | 3,000+ tokens, tries to do everything | Split into focused sub-skills |
| Vague instructions | "Analyze the thing" | Specific steps with templates |
| Duplicate rules | Skill repeats what's in rules | Reference rules, don't copy |
| No output format | Agent produces inconsistent output | Always include format template |
| Hardcoded values | Dates, versions in skill text | Use "current" or "latest" |
Token optimization
| Technique | Savings |
|---|---|
| Reference rules instead of copying | -500-1,000 tokens |
| Use tables for decision criteria | More compact than prose |
| Include only essential examples | -200-500 per example |
| Link to UB entries for deep details | Skill stays lean |
Related pages
| Page | Relationship |
|---|---|
| Creating a Skill | How to build |
| Skills Overview | What skills are |
| CLAUDE.md + PD | PD architecture |