Agent Skills: Using Skills

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

One loop, five entry points. There is no separate mid-workflow command: a run that failed its gate is re-entered with craft-redispatch.sh, which re-hashes the amended plan and re-runs only the tasks that flagged plus their dependents.

| Command | Purpose | |---------|---------| | /craft | Any task worth doing properly: clarify, approved plan, delegated implementation, independent verification, human review | | /dev | Feature development and bug fixes, under test-first discipline | | /ds | Data analysis and panel construction, with a computed data-quality gate | | /writing | Articles, essays, briefs and chapters, with a computed plan-grammar and citation gate | | /workshop | Typst slides and speaker notes built from a research paper |

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>

Companion Routing Facts

• Agent tool with run_in_background is NOT a companion session: background agents die when the conversation ends and can't be accessed via the companion web UI. Substituting it for the companion skill gives the user the opposite of the persistent, monitorable session they asked for.

Skill Triggers (Can Auto-Invoke)

| User Intent | Command | Trigger Words | |-------------|---------|---------------| | Session/companion | companion | companion session, new session, separate session, background session, hand off, in a new session | | Structured work | /craft | do this properly, clarify and plan this, plan and verify this, craft this, don't just wing it | | Bug/fix | /dev | bug, broken, fix, doesn't work, crash, error, fails, implement, add support for | | Data work | /ds | analyze this data, build the panel, run the regression, results wrong, notebook error | | Writing | /writing | write, draft, document, essay, paper | | Media analysis | look-at | describe image, analyze PDF, what's in this, screenshot, diagram | | Create/edit skill | workflows:skill-creator | create skill, improve skill, edit skill, add enforcement, audit skill, SKILL.md | | Create/repair workflow | workflows:workflow-creator | create workflow, design new workflow, audit workflow, repair workflow, improve existing workflow | | Create/edit plugin | workflows:plugin-creator | create plugin, scaffold plugin, new plugin, plugin structure, edit plugin | | Workshop presentation | /workshop | workshop presentation, workshop slides, faculty workshop, workshop talk, slides from paper, revise the deck |

Red Flags

• About to invoke a skill the user already invoked (e.g., "use /dev") → it is ALREADY LOADED; check for the <command-name> tag and just proceed.
• About to "gather information" or "quickly check" code before starting the matching workflow → that IS investigation; invoke the skill first. Scope ("just one file", "simple question") doesn't exempt you from the process.
• About to Read an image/PDF directly → use look-at.
• User said "use X in a companion session" and you're about to invoke X directly or via the Agent tool → companion skill is the transport, X goes in the prompt. Agent = subagent in THIS session; companion = separate server session.

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. Invoke /dev — it clarifies, plans, and dispatches the fix under a failing test
2. Follow the /dev 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. Specialized task shape next - code uses dev; analysis uses ds; long-form prose uses writing; decks use workshop
3. Generic structure next - use craft for work that needs criteria, verification and human review but has no domain gate
4. Direct execution for trivial work - a lookup, one-line answer, or tiny edit does not earn workflow ceremony
5. Then implementation - the selected workflow dispatches it; the main chat does not do the work

/craft is not a universal wrapper. A domain workflow wins when the task has its shape, because it brings a gate craft does not have.

How to Invoke

Use the Skill tool to invoke skills:

# craft: the spine — clarify, approved plan, delegated implementation, independent verification, human review
Skill(skill="workflows:craft")

# 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")

craft composes the goal for you and self-sends it — compose-goal.sh builds the condition from the approved plan and goal-self-send.sh sets it, so the loop's exit condition is derived from what was approved rather than typed from memory. Do not hand-write a /goal for a craft run; amend the plan and re-dispatch instead.

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.sh)
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
"${CLAUDE_PLUGIN_ROOT}/skills/look-at/scripts/look_at.sh" \
    --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

Tool Routing Facts

• Read on an image costs 1,000+ context tokens even for a "small" file; look-at returns 50-200 tokens of extracted info. File size doesn't change this — content type determines the tool.
• look-at is FOR YOU, not the user — it applies whether or not the user asked for it. You can always fall back to Read if the extraction is insufficient; start with look-at and escalate.

Red Flags

• Passing an image, PDF, or screenshot path to the Read tool → use look-at.

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
"${CLAUDE_PLUGIN_ROOT}/skills/look-at/scripts/look_at.sh" \
    --file "$HOME/Downloads/screenshot.png" \
    --goal "List all buttons and their labels"

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

# look-at: Extract information from PDF document
"${CLAUDE_PLUGIN_ROOT}/skills/look-at/scripts/look_at.sh" \
    --file "$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
"${CLAUDE_PLUGIN_ROOT}/skills/look-at/scripts/look_at.sh" \
    --file "/path/to/file.pdf" \
    --goal "Extract title"

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

Red Flags

• About to call Bash without the description parameter when the skill requires it → add it.
• About to modify a skill's required pattern "to be simpler" → follow the skill or don't load it.

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 a layer on top of built-in creator tools:

1. Behavioral enforcement — superpowers patterns (Iron Laws, rationalization tables, red flags, enforcement checklist audit)
2. Mechanical enforcement — hooks/validate-skill-paths.ts is registered on PostToolUse Edit|Write in hooks/hooks.json and reports broken ${CLAUDE_SKILL_DIR} / ${CLAUDE_PLUGIN_ROOT} references on every skill edit. hooks/plugin-validate.ts is not registered — its only finding on this repo is a constant symlink warning.

Using built-in creators directly bypasses the behavioral layer. 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, repair or audit -> 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

| Activity | Route To | Wraps | |----------|----------|-------| | Create/edit a skill | workflows:skill-creator | skill-creator:skill-creator | | Create, repair or audit a workflow | workflows:workflow-creator | the craft loop plus wc-probe.ts | | Create/edit a plugin | workflows:plugin-creator | plugin-dev:create-plugin |

Trigger words: see Skill Triggers table above.

Red Flags

• About to invoke a built-in creator directly (skill-creator:skill-creator, plugin-dev:*, etc.) → use the workflows wrapper. The built-in has no PostToolUse hooks; path errors and structural issues go uncaught.
• Editing enforcement patterns without the wrapper → 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.