<EXTREMELY-IMPORTANT> Every recommendation must be justified by Claude Code runtime behavior, not taste.

Non-negotiable rules:

1. Read references/claude-skill-runtime.md before planning or rewriting.
2. Default to --mode plan unless the user explicitly asks to rewrite now.
3. Treat frontmatter as the highest-leverage surface. Claude routes and budgets skills from metadata first.
4. Keep SKILL.md focused on workflow. Move bulky examples, framework variants, and edge-case catalogs into references/ or scripts/.
5. Do not add context: fork, paths:, or user-invocable: false without a concrete runtime reason.
6. Never add agent-only frontmatter to a skill. tools, disallowedTools, skills, initialPrompt, permissionMode, maxTurns, background, memory, isolation, color, and mcpServers belong to agents, not skills.

</EXTREMELY-IMPORTANT>

Normalize Skill For Claude

Inputs

• $target: Skill directory or direct path to SKILL.md
• $mode: Optional. plan or rewrite. Default: plan

Goal

Produce a Claude Code optimized skill that is easier to auto-invoke, cheaper to keep in prompt budget, and safer to execute. Always start by creating a per-skill DAG plan unless the user explicitly asks for direct rewrite.

Step 1: Resolve the target

1. Accept either a skill directory or a direct SKILL.md path.
2. Normalize to the skill root and confirm SKILL.md exists.
3. Determine mode:

- --mode rewrite means plan first, then rewrite in the same run - missing mode means plan

1. Inventory the current skill:

- line count of SKILL.md - existing frontmatter fields - whether references/, scripts/, or assets/ already exist - whether the skill appears internal-only or user-invocable

Success criteria: You know the exact skill root, target mode, current file shape, and likely scope of the rewrite.

Step 2: Load the Claude Code skill runtime anchors

Read references/claude-skill-runtime.md fully before making any recommendation.

Extract these constraints from the reference:

• Claude budgets skill discovery from frontmatter and short descriptions first.
• Full skill bodies are loaded on invocation, not for listing.
• paths: activates conditional skills when matching files are touched.
• context: fork runs the skill in a forked subagent and should be rare.
• allowed-tools should be the minimum needed permission surface.
• Skill rewrites must use skill-native frontmatter only, not agent-native headers.

If this repository contains claude-code-source/, use the exact source files named in the reference to verify edge cases before rewriting unusual frontmatter.

Success criteria: Every planned change can be tied to a specific Claude runtime behavior.

Step 3: Audit the current skill against Claude's runtime

Evaluate the target skill using this checklist:

1. Description and trigger quality

- Does description say what the skill does, when to use it, and example trigger phrases? - Is whentouse missing or weak?

1. Frontmatter completeness

- Would allowed-tools, arguments, argument-hint, user-invocable, paths, context, or agent improve runtime behavior? - Are there any agent-only headers that must be removed or explicitly avoided?

1. Body size and shape

- Is the body carrying reference material instead of core workflow? - Are there giant examples, duplicated checklists, or framework-specific sections that belong in references/? - Structural content audit: Does the skill contain routing tables (task→reference mappings), personality/expertise definitions, numbered rule sets, or gate classifications? These are high-signal inline content, NOT bulk examples. Tag each as PRESERVE-INLINE or EXTRACT-WITH-MANDATORY-LOAD.

1. Invocation type

- Is this a public slash command, an internal router, or a conditional path-scoped helper?

1. Execution model

- Would context: fork help, or would it fragment work and require mid-flow user interaction?

1. Determinism opportunities

- Are there repeated shell snippets or fragile transformations that belong in scripts/?

Classify each issue:

• ROUTING: metadata is too weak for auto-invocation
• PROMPT: body is too large or redundant
• SAFETY: tool scope or execution mode is too broad
• STRUCTURE: references or scripts should be extracted

Success criteria: You have a concrete, source-backed audit of what must change and why.

Step 4: Write the per-skill DAG plan

Create both of these artifacts:

• .ulpi/plans/skills/<skill-name>-normalize-for-claude.md
• .ulpi/plans/skills/<skill-name>-normalize-for-claude.json

The plan must include:

1. Current state

- line count - existing frontmatter - current folders - largest bloat areas

1. Claude runtime findings

- each finding mapped to source references from references/claude-skill-runtime.md

1. Target state

- final frontmatter shape - files to keep, rewrite, create, or split

1. DAG tasks

- frontmatter rewrite - body trim - reference extraction - optional script extraction - validation

1. Guardrails

- what must not be changed - what must not be over-automated

1. Validation

- concrete commands or checks to confirm the skill is structurally sound

Each DAG task should include:

• id
• title
• rationale
• filesToModify
• filesToCreate
• dependencies
• validation

Success criteria: The markdown and JSON plans describe the same work and can be executed safely.

Step 5: Rewrite only if requested

If mode is plan, stop after writing the DAG artifacts.

If mode is rewrite:

1. Rewrite the frontmatter first.
2. Keep the body focused on:

- inputs - goal - workflow steps - guardrails - output contract

1. Move bulky material into references/, but respect content classification:

- Bulk (move to references): giant examples, framework variants, edge-case catalogs, verbose checklists, before/after code samples - Structural (keep inline OR extract with mandatory load directive): routing tables (task→reference mappings), numbered guardrail rules in EXTREMELY-IMPORTANT, gate classifications (BLOCK/CONCERN/OBSERVATION), personality/expertise sections - Routing tables that map task types to reference files MUST stay inline — they are the skill's primary navigation surface. - Concrete guardrail lists (numbered rules, always/never) MUST either stay in the EXTREMELY-IMPORTANT block or get a mandatory load directive in the step that needs them — not a soft "see references" suggestion. - When extracting personality/expertise to a reference file, add a mandatory load directive at session start, not just a "When To Load References" entry.

1. Create scripts/ only when deterministic repeated logic is clearly justified.
2. Keep the resulting SKILL.md lean enough that future edits remain easy.

Preferred rewrite outcomes:

• description and whentouse become trigger-oriented
• internal router skills become user-invocable: false
• domain skills gain paths: only when file-touch activation clearly helps
• context: fork appears only for self-contained analysis/research workflows
• no agent-only headers are introduced into the rewritten skill

Success criteria: The rewritten skill is structurally smaller, clearer to route, and justified by Claude runtime behavior.

Step 6: Validate the result

After planning or rewriting:

1. Re-read the final SKILL.md.
2. Confirm the body does not duplicate reference files.
3. Confirm frontmatter fields are intentionally chosen, not cargo-culted.
4. Confirm no agent-only headers were introduced.
5. Confirm every references/ file is directly linked from SKILL.md.
6. Confirm no routing tables, numbered guardrail rules, or gate classifications were stripped without replacement.
7. Confirm the plan or rewrite summary names the Claude source anchors that drove the major decisions.

Success criteria: The output is usable immediately by a human or follow-on agent without guessing.

Output Contract

Always report:

1. Skill: target path and normalized root
2. Mode: plan or rewrite
3. Top runtime issues: the highest-value Claude mismatches
4. Artifacts: exact plan or rewritten file paths
5. Guardrails applied: 3 to 5 bullets tied to Claude runtime behavior

If rewrite mode was used, also report:

• which sections moved to references/
• which frontmatter fields were added, removed, or intentionally omitted