activation

When there is even a remote possibility that a skill is relevant to the current task, you MUST invoke it immediately.

YOU DO NOT HAVE THE OPTION TO SKIP A RELEVANT SKILL. IT IS MANDATORY.

This requirement is absolute. It cannot be reasoned away or deferred.

Accessing Skills

Plugin installation (standard setup): Invoke skills directly through the Skill tool. All protocols are auto-registered. Example: Skill("godmode:intent-discovery")

When the Skill tool cannot locate a protocol: Fall back to reading the file directly with the Read tool:

1. Protocols live at: [plugin-root]/skills/
2. Each protocol has a definition: [plugin-root]/skills/[protocol-name]/SKILL.md
3. Read the SKILL.md and execute its instructions

Locating the plugin root:

• Running from the repository: the working directory is the plugin root
• Installed as a plugin: the CLAUDE_PLUGIN_ROOT environment variable contains the path

Other platforms: Consult platform-specific documentation for skill loading mechanisms.

Working With Skills

The Mandate

Invoke applicable skills BEFORE generating any response or taking any action. Even a marginal chance of relevance is sufficient reason to invoke. If an invoked skill turns out to be inapplicable, you can simply move on.

digraph protocol_flow {
    "Message arrives" [shape=doublecircle];
    "About to enter planning?" [shape=doublecircle];
    "Intent discovery done?" [shape=diamond];
    "Run intent-discovery" [shape=box];
    "Could any skill apply?" [shape=diamond];
    "Invoke Skill tool" [shape=box];
    "Skill tool errored?" [shape=diamond];
    "Read SKILL.md fallback" [shape=box];
    "Announce: 'Applying [skill] for [goal]'" [shape=box];
    "Contains checklist?" [shape=diamond];
    "Track each checklist item" [shape=box];
    "Execute skill instructions" [shape=box];
    "Generate response" [shape=doublecircle];

    "About to enter planning?" -> "Intent discovery done?";
    "Intent discovery done?" -> "Run intent-discovery" [label="no"];
    "Intent discovery done?" -> "Could any skill apply?" [label="yes"];
    "Run intent-discovery" -> "Could any skill apply?";

    "Message arrives" -> "Could any skill apply?";
    "Could any skill apply?" -> "Invoke Skill tool" [label="yes, any chance"];
    "Could any skill apply?" -> "Generate response" [label="clearly not"];
    "Invoke Skill tool" -> "Skill tool errored?";
    "Skill tool errored?" -> "Read SKILL.md fallback" [label="yes"];
    "Skill tool errored?" -> "Announce: 'Applying [skill] for [goal]'" [label="no"];
    "Read SKILL.md fallback" -> "Announce: 'Applying [skill] for [goal]'";
    "Announce: 'Applying [skill] for [goal]'" -> "Contains checklist?";
    "Contains checklist?" -> "Track each checklist item" [label="yes"];
    "Contains checklist?" -> "Execute skill instructions" [label="no"];
    "Track each checklist item" -> "Execute skill instructions";
}

Guardrails

These internal thoughts are warning signs that you are rationalizing:

Thought	Truth
"This is a straightforward question"	Questions are tasks. Look for applicable skills.
"I should gather more context first"	Skill invocation precedes clarification.
"Let me scan the codebase first"	Skills define HOW to scan. Invoke first.
"I can check files and history quickly"	Files lack conversational context. Invoke skills.
"Let me collect some data first"	Skills prescribe data collection methods.
"This doesn't warrant a formal skill"	If a matching skill exists, it must be used.
"I recall what this skill says"	Skills get updated. Read the current version.
"This isn't really a task"	Any action is a task. Look for skills.
"The skill is too heavy for this"	Seemingly simple tasks often have hidden complexity. Use it.
"I'll do this one thing first"	Invoke BEFORE doing anything.
"I'm making good progress"	Undirected effort creates waste. Skills prevent this.
"I understand that concept already"	Understanding a concept is not the same as following a skill. Invoke it.

Skill Ordering

When several skills might be relevant, follow this sequence:

1. Challenge skills first (rationale) - question whether the work is worth doing before committing
2. Process skills second (intent-discovery, fault-diagnosis) - establish the approach
3. Reference skills third (reference-engine routes to: ux-patterns, design-research, github-search, codebase-research, system-design, deployment-advisor) - locate proven solutions before designing
4. Design skills fourth (ui-engineering, design-integration, specification-first) - apply references to guide decisions
5. Implementation skills fifth (test-first, project-bootstrap, pattern-matching, environment-awareness) - drive execution
6. Quality skills sixth (quality-enforcement, security-protocol, completion-gate, comprehension-check, error-recovery) - enforce standards
7. Orchestration skills (team-orchestration, delegated-execution) - leverage parallelism

"Let's build X" -> rationale -> intent-discovery -> reference-engine (routes to appropriate references) -> design skills -> implementation skills. "Build a website" -> rationale -> intent-discovery -> reference-engine -> design-research + ux-patterns -> ui-engineering -> implementation. "Fix this bug" -> fault-diagnosis first, then domain-appropriate skills. (rationale skipped for bugs) "Big project" -> rationale -> intent-discovery -> reference-engine -> task-planning -> team-orchestration (when warranted). "Agent is stuck" -> error-recovery activates proactively based on failure count detection. "What database/hosting should I use?" -> environment-awareness -> deployment-advisor -> system-design.

Execution Mode Guidance

When a user says "use godmode" for a task, evaluate the task and suggest the optimal execution strategy:

digraph execution_guidance {
    "Evaluate task" [shape=box];
    "Single atomic step?" [shape=diamond];
    "Handle directly" [shape=box];
    "Multiple independent parts?" [shape=diamond];
    "Requires peer coordination?" [shape=diamond];
    "Suggest Agent Teams" [shape=box style=filled fillcolor=lightgreen];
    "Suggest Parallel Delegation" [shape=box];
    "Multi-step sequence?" [shape=diamond];
    "Suggest Delegated Execution" [shape=box];
    "Suggest single-agent workflow" [shape=box];

    "Evaluate task" -> "Single atomic step?";
    "Single atomic step?" -> "Handle directly" [label="yes"];
    "Single atomic step?" -> "Multiple independent parts?" [label="no"];
    "Multiple independent parts?" -> "Requires peer coordination?" [label="yes"];
    "Multiple independent parts?" -> "Multi-step sequence?" [label="no"];
    "Requires peer coordination?" -> "Suggest Agent Teams" [label="yes"];
    "Requires peer coordination?" -> "Suggest Parallel Delegation" [label="no"];
    "Multi-step sequence?" -> "Suggest Delegated Execution" [label="yes"];
    "Multi-step sequence?" -> "Suggest single-agent workflow" [label="no"];
}

Present the suggestion before acting:

Based on my analysis, I suggest:

[Suggested strategy] because [reasoning]

- Agent Teams: [X] agents collaborating on [modules] concurrently
- OR Delegated Execution: [N] sequential tasks with review gates
- OR Parallel Delegation: [N] independent tasks running simultaneously
- OR Single agent: manageable enough to handle in one flow

Shall I proceed this way, or do you prefer a different approach?

Skill Categories

Rigid (test-first, fault-diagnosis): Follow precisely. Do not deviate from the discipline.

Flexible (design patterns): Adapt the principles to fit the situation.

The skill definition itself indicates which category applies.

YoloMode

When the user signals they want autonomous execution, skip all confirmation prompts and compress workflows — but never auto-pick labeled options.

Trigger phrases (case-insensitive):

• "yolo", "yolomode", "yolo mode"
• "just go", "just do it", "go for it"
• "skip the questions", "don't ask"
• "auto-pick", "you decide"

Behavior when active:

• Skills that present A/B/C/D options ALWAYS pause for user selection — no exceptions
• Intent-discovery compresses: survey + research + present recommended approach, skip per-section confirmation
• Rationale analysis still runs (safety check) but is compressed — present analysis and options concisely
• Completion-gate and test verification are NEVER skipped (quality is non-negotiable)

How skills reference YoloMode:

If YoloMode is active:
  -> State: "YoloMode active — proceeding with recommendation [X] because [reason]"
  -> Execute immediately
  -> Log which option was auto-selected

If YoloMode is NOT active:
  -> Present options with labels and recommendation
  -> Wait for user selection

YoloMode does NOT bypass:

• Test execution (completion-gate)
• Security checks (security-protocol)
• Destructive operations (merge-protocol abandon confirmation)
• Quality enforcement (quality-enforcement)

YoloMode is about speed, not recklessness. It trusts the agent's recommendation for design choices, not for safety checks.

User Directives

User directives specify WHAT needs to happen, not HOW. "Add X" or "Fix Y" does not mean bypass established workflows.