Using Skills

Invoke relevant skills BEFORE any response or action.

This is non-negotiable. Even a 1% chance a skill applies requires checking.

CRITICAL: Skill Already Loaded - DO NOT RE-INVOKE

DO NOT:

• ❌ Use the Skill tool to invoke it again
• ❌ Say "I need to invoke the skill"
• ❌ Call Skill(skill="dev") or similar

DO INSTEAD:

• ✅ The skill instructions follow immediately in the next message
• ✅ Just proceed to the next step
• ✅ Follow the loaded skill's instructions directly

If you catch yourself about to invoke a skill that's already loaded, STOP. Just go to the next step. </EXTREMELY-IMPORTANT>

The Rule

User message arrives
    ↓
Is user explicitly invoking a skill (e.g., "use /dev")?
    ↓
YES → SKILL IS ALREADY LOADED
      ↓
      DO NOT invoke again with Skill tool
      ↓
      Proceed to next step (follow skill instructions)
NO  → Contains session keyword? (companion, new session, background session, etc.)
    ↓
YES → Invoke COMPANION skill FIRST — put everything else in the session prompt
NO  → Check: Does this match any other skill trigger?
    ↓
YES → Invoke skill FIRST, then follow its protocol
NO  → Proceed normally

Workflow Commands

Each workflow has two entry points — start fresh or re-enter mid-workflow:

IRON LAW: Companion Transport Priority

When the user's request mentions any session keyword — 'companion session', 'new session', 'separate session', 'background session', 'parallel session', 'companion', 'hand off to a session', 'in a new session' — the companion skill MUST be invoked FIRST, regardless of what other skills are mentioned.

The companion skill is a TRANSPORT mechanism. It launches the session. Other skills/tasks go inside the session's prompt.

"use workflows creator in a new companion session"
    ↓
WRONG: invoke workflows:skill-creator directly (or Agent tool)
RIGHT: invoke companion skill, put "use workflows:skill-creator" in the prompt

The rule: 'do X in a companion session' = companion launches, X goes in the prompt. NOT 'do X directly'.

Invoking X directly when the user said 'in a companion session' is NOT HELPFUL — the task runs in your current context, dies when the conversation ends, and the user cannot monitor or interact with it in the companion web UI. You did the opposite of what was asked. </EXTREMELY-IMPORTANT>

Rationalization Table — Companion Routing

Skill Triggers (Can Auto-Invoke)

Red Flags - You're Skipping the Skill Check

If you think any of these, STOP:

Bug Reports - Mandatory Response

When user mentions a bug:

DO NOT:
1. Read code files
2. Investigate independently
3. "Take a look" without structure

INSTEAD:
1. Start ralph loop:
   ralph-loop: Start Ralph Loop in current session with bug debugging
   Skill(skill="ralph-loop:ralph-loop", args="Debug: [symptom] --max-iterations 15 --completion-promise FIXED")
2. Inside loop, follow /dev-debug protocol

Any code reading before starting the workflow is a violation.

Skill Priority

When multiple skills could apply:

1. Transport skills first - companion session routes EVERYTHING else inside the session prompt
2. Process skills next - debugging, brainstorming determine approach
3. Then implementation - dev, ds, writing execute the approach

How to Invoke

Use the Skill tool to invoke skills:

# dev-debug: Midpoint entry for dev workflow - debug, fix, re-test
Skill(skill="dev-debug")

# dev: Feature development workflow with 7 phases and TDD enforcement
Skill(skill="dev")

# ds: Data analysis workflow with 5 phases and output-first verification
Skill(skill="ds")

Or start ralph loop first for implementation/debug phases:

# ralph-loop: Per-task ralph loop pattern for implementation and debugging
Skill(skill="ralph-loop:ralph-loop", args="Task description --max-iterations 15")

IRON LAW: Multimodal File Analysis

NO READING IMAGES/PDFS WITH Read TOOL. USE look-at INSTEAD.

The Rule

User asks about image/PDF/media content
    ↓
Is it a media file requiring interpretation?
    ↓
YES → Use look-at skill (bash call to look_at.py)
NO  → Use Read tool for source code/text

When to Use look-at

ALWAYS use look-at for:

• .jpg, .jpeg, .png, .webp, .gif, .heic - Images
• .pdf - PDFs requiring content extraction
• .mp4, .mov, .avi, .webm - Videos
• .mp3, .wav, .aac, .ogg - Audio
• Any file where you need to UNDERSTAND content, not just see raw bytes

Pattern:

# look-at: Extract information from media file with specific goal
python3 "${CLAUDE_PLUGIN_ROOT}/skills/look-at/scripts/look_at.py" \
    --file "/absolute/path/to/file" \
    --goal "What specific information to extract"

When NOT to Use look-at

Use Read tool instead for:

• Source code files (.py, .js, .rs, etc.) - need exact formatting for editing
• Plain text files (.txt, .md, .json, etc.) - preserve exact content
• Config files requiring exact formatting preservation
• Any file that needs editing after reading

Rationalization Table - STOP If You Think:

Red Flags - STOP If You Catch Yourself:

• "Let me read this image..." → NO. Use look-at.
• "I'll use Read to see what's in the PDF..." → NO. Use look-at.
• "Just quickly checking this screenshot..." → NO. Use look-at.
• Passing image path to Read tool → STOP. Use look-at instead.

Cost & Context Benefits

• Read tool on image: ~1,000-5,000 context tokens
• look-at extraction: ~50-200 output tokens
• Savings: 95%+ token reduction
• Speed: Faster responses, less context bloat

Example Usage

# look-at: Extract specific information from image file
python3 "${CLAUDE_PLUGIN_ROOT}/skills/look-at/scripts/look_at.py" \
    --file "$HOME/Downloads/screenshot.png" \
    --goal "List all buttons and their labels"

# look-at: Analyze diagram to understand data flow
python3 "${CLAUDE_PLUGIN_ROOT}/skills/look-at/scripts/look_at.py" \
    --file "$HOME/Documents/architecture.png" \
    --goal "Explain the data flow between components"

# look-at: Extract information from PDF document
python3 "${CLAUDE_PLUGIN_ROOT}/skills/look-at/scripts/look_at.py" \
    --file "$HOME/Downloads/report.pdf" \
    --goal "Extract the executive summary section"

Enforcement

Using Read on images/PDFs when look-at should be used results in:

1. Wasting context tokens unnecessarily
2. Making conversations slower
3. Ignoring available optimization tools
4. Violating the tool selection protocol

Validate before calling Read: Ask "Is this a media file?" If yes, invoke look-at instead.

IRON LAW: Following Skill Instructions

WHEN A SKILL LOADS, YOU MUST FOLLOW ITS EXACT INSTRUCTIONS.

Skills contain specific patterns, required parameters, and enforcement rules. Skipping these requirements defeats the purpose of loading the skill.

The Rule

Skill loads successfully
    ↓
Read the skill's requirements carefully
    ↓
Follow ALL instructions, including:
    - Required tool parameters (descriptions, timeouts, etc.)
    - Specific command patterns
    - Enforcement patterns (Iron Laws, Red Flags)
    - Step sequences
    ↓
Execute using the skill's exact patterns

Common Violations

Bash Description Parameter:

When a skill requires description parameter on Bash calls (like look-at), you MUST include it:

# ❌ WRONG: No description parameter
python3 "${CLAUDE_PLUGIN_ROOT}/skills/look-at/scripts/look_at.py" \
    --file "/path/to/file.pdf" \
    --goal "Extract title"

# ✅ CORRECT: With description parameter as skill requires
Bash(
    command='python3 "${CLAUDE_PLUGIN_ROOT}/skills/look-at/scripts/look_at.py" --file "/path/to/file.pdf" --goal "Extract title"',
    description="look-at: Extract title"
)

Rationalization Table - STOP If You Think:

Red Flags - STOP If You Catch Yourself:

• About to call Bash without description when skill requires it → STOP. Add the description parameter.
• Thinking "I'll skip this requirement" → STOP. Skills don't have optional requirements.
• "The skill says to do X but I'll do Y" → STOP. Follow the skill or don't load it.
• Modifying the skill's pattern "to be simpler" → STOP. The pattern exists for a reason.

Why This Matters

Skills encode:

1. Tested patterns - Proven to work in production
2. Optimization - Context/token savings, clean output
3. Enforcement - Prevent common mistakes
4. UX standards - Consistent, professional output

When you skip skill instructions:

• ❌ You waste the effort of loading the skill
• ❌ You create messy, unprofessional output
• ❌ You miss optimizations (context savings, speed)
• ❌ You violate user expectations
• ❌ You make debugging harder

The skill loaded for a reason - follow it completely.

IRON LAW: Creator Activity Routing

NO CREATING OR SUBSTANTIALLY EDITING SKILLS, PLUGINS, OR WORKFLOWS WITHOUT THE WORKFLOWS WRAPPER.

The workflows plugin provides wrapper skills that add two layers on top of built-in creator tools:

1. Behavioral enforcement — superpowers patterns (Iron Laws, rationalization tables, red flags, enforcement checklist audit)
2. Mechanical enforcement — PostToolUse hooks (plugin-validate.py, validate-skill-paths.py) that fire on every Write/Edit

Using built-in creators directly bypasses both layers. This applies globally — any project, not just the workflows plugin.

"Substantial" means: adding/removing sections, changing enforcement patterns, altering process flow, adding/modifying hooks. Typo fixes, version bumps, and single-line clarifications are exempt.

The Rule

About to create or substantially edit a skill/plugin/workflow
    ↓
What type of creator activity?
    ↓
    +---> Skill creation/editing ------> Invoke workflows:skill-creator
    |
    +---> Workflow creation/editing ----> Invoke workflows:workflow-creator
    |
    +---> Plugin creation/editing ------> Invoke workflows:plugin-creator
    ↓
Follow the wrapper skill's full process
    ↓
DO NOT invoke built-in creators directly (skill-creator:skill-creator,
plugin-dev:skill-development, plugin-dev:plugin-structure, etc.)

Creator Routing Table

Trigger words: see Skill Triggers table above.

Rationalization Table - STOP If You Think:

Red Flags - STOP If You Catch Yourself:

• About to invoke a built-in creator directly (skill-creator:skill-creator, plugin-dev:*, etc.) → STOP. Use the workflows wrapper.
• Editing enforcement patterns without the wrapper → STOP. This is exactly when the audit matters most.

Advanced Agent Harnessing Patterns

For detailed oh-my-opencode production patterns including:

• Background + parallel execution (3x speedup)
• Tool restrictions for focused agents
• Structured delegation templates
• Failure recovery protocol
• Environment context injection
• Cost classification system
• Metadata-driven prompts

See: references/agent-harnessing.md

Quick reference:

• Tool restrictions: references/tool-restrictions.md
• Delegation template: references/delegation-template.md
• Metadata infrastructure: references/skill-metadata.py

Based on: obra/superpowers and oh-my-opencode production patterns.