claudify

Claudify

Guide users to convert functionality into the appropriate Claude Code automation type (Agent, Skill, Rule, Command, or Hook).

Topic Dispatch

When this skill is invoked with a topic specifier (e.g., Skill("claudify", "improve") or /claudify improve), load and follow ONLY the matching topic file below — do NOT default to the creation-wizard body (Decision Matrix / Workflow) in this SKILL.md. A bare invocation (no topic, agentify, or create) uses the creation workflow below.

For improve / persist / background-polling, Read that topic file and follow it — the Decision Matrix and creation Workflow below apply only to the create/agentify/(none) creation path. For any other/unrecognized topic arg, fall back to the create/agentify/(none) creation workflow below rather than assuming a matching topic file exists.

Decision Matrix

Full comparison: automation-decision-guide.md

Workflow

Step 0: Output-size routing (HARD STOP — before Step 1)

Before any large-output operation, decide: inline vs subagent vs script. Parent context bloat from claudify itself reduces the budget available for the actual automation creation work.

Routing decision matrix

Don't / Do table

Self-check (before each Read/Glob/WebFetch in claudify workflow)

1. Is the read result the deliverable, or intermediate context that the subagent's parent doesn't need?
2. Will this Read + subsequent Reads cumulatively add >2K tokens to parent context?
3. Is the operation parallelizable (multiple Reads, multiple files)?
4. If 1=intermediate OR 2=yes OR 3=yes → dispatch subagent

Subagent return contract: subagent returns only the deliverable path(s). No template dump, no intermediate analysis, no confirmation text.

Step 1: Identify Candidates

If no target specified ("agentify" alone):

• Review conversation for automation candidates
• Look for: verbose outputs, multi-step workflows, repeated patterns
• MUST use multiSelect: true when presenting candidates (users often want multiple)

If target specified:

1. Check local marketplaces first:
   • ~/.claude/plugins/marketplaces/*/plugins/*/
   • Use /skill-dedup command to find overlaps
2. If not found locally, search remote: WebFetch https://claudemarketplaces.com/?search=[keyword]
3. If found, recommend existing or extend. If not, proceed to create

Step 1 Guards (HARD STOP):

• No immediate-execution options: AskUserQuestion options must all be automation types (Skill/Agent/Rule/Hook/Command). "Just execute now" or "Execute without automation" options are forbidden — claudify's purpose is creating/extending automation, not executing the underlying task.
• Scope check when extending existing skills: Before proposing to extend a global skill (~/.claude/skills/), verify the change is project-agnostic. Project-specific variables, hostnames, or environment values must NOT be added to global skills. If the automation is project-specific, recommend a project-level skill (.claude/skills/) or a project rule (.claude/rules/) instead.

Step 1.5: Merge logic / Grouping / AskUserQuestion

When duplicate or similar automation candidates are found, confirm grouping and merge options with the user using AskUserQuestion.

Grouping criteria:

• Similarity in model or functionality
• Contextual similarity
• Keyword grouping (e.g., "openclaw management")

Merge criteria:

AskUserQuestion (merge options): "Candidates found. How should I structure them?" options:

• "Merge related topics into one multi-topic skill (e.g., openclaw: exec/gateway/test)"
• "Create separate agents for each functionality"
• "Skill + Agent combination (instruction=skill, implementation=agent)"

PROHIBITED: Do not create separate agents without merging candidates into a logical structure if they are related.

Step 2: Gather Requirements

Use AskUserQuestion to clarify:

• Triggers: Commands or conditions that activate the automation
• Tools: Required tools or skills
• Scope: Global (~/.claude/) / Project (.claude/)
• Language: code comments, variable names, documentation

Question patterns: askuserquestion-patterns.md

Step 3: Recommend Type

Use the automation-decision-guide.md to recommend the best type.

Step 4: Create

CRITICAL: Follow the creation method for each type. Apply Step 0 routing first — multi-template Reads or transcript scans dispatch to subagent.

Skill with scripts (If scripts are required, use skill directory structure):

skill-name/
├── SKILL.md          # frontmatter + documentation + node scripts/xxx.js (mode)
├── topic-a.md        # topic file
├── topic-b.md        # topic file
└── scripts/
    └── xxx.js        # actual logic (temp location, permanent storage)

• Do not create tmp_*.js in current directory; move to scripts/
• Call in SKILL.md: node /scripts/xxx.js
• Use relative path from __dirname in scripts
• Separate implementation logic and topic files

Skill (no scripts): MUST use skill-writer (do NOT create directly)

Skill tool: skill: "project-automation:skill-writer"

Agent: Create in ~/.claude/agents/ or .claude/agents/

• agent-templates.md

Rules: Create in ~/.claude/rules/ or .claude/rules/

• rules-guide.md

Slash Command: Create in ~/.claude/commands/ or .claude/commands/

• slash-command-syntax.md

Hook: Add to settings.json

• hook-examples.md

Plugin (open source):

• plugin-creation.md

Step 5: Validate

1. Register or copy to target location
2. Reload Claude Code:
   1. Manual sync to cache, OR
   2. New session to reload

Auto-sync hook: plugin-cache-sync.sh syncs marketplace to cache on Edit/Write

Output Guidelines

Keep responses concise:

1. List identified candidates (with multiSelect)
2. Summarize the recommended structure (Merge logic)
3. Provide the creation plan

Success Case

Scenario (2026-03-09):

• Found 3 openclaw-related functions
• Proposed 3 options for merging
• Result: Implementation success, user satisfied

Key factors:

1. Identification of 3 functions
2. "Merge?" AskUserQuestion
3. merging skills using skill-writer (multi-topic)

Ralph Mode (AskUserQuestion bypass)

Detection condition: do not judge by .ralph/ presence alone. Ralph mode requires all of:

1. .ralph/ directory exists AND
2. Environment variable RALPH_LOOP=1 is set (Ralph autonomous loop sets this)

Even when .ralph/ exists, an interactive user session is normal mode — AskUserQuestion works as usual.

Workflow Change:

Scope note (HARD STOP): this table governs the creation workflow only (new agent/skill/rule/hook). It does not extend to passive, no-ask persistence steps elsewhere in this skill's topics (improve/persist RAG session/discovery-chunk store) — those follow the narrower carve-out in persist.md's own "Ralph Mode" section, not this table. Do not generalize "Ralph Mode = record-only, nothing runs" from this table alone.

improvements.md recording example

## Agentify Candidate (Implementation)

### [Candidate Name]
- **Context**: [Why it was found]
- **Recommended Type**: [Skill/Agent/Hook/Slash Command]
- **Recommended Structure**: [Topics]
- **Rationale**: [Why it's recommended]
- **Tag**: [NEEDS_REVIEW]

Self-Improvement

After changes are complete, Self-improve based on conversation:

1. Identify failure and workaround patterns