skill-create

Instructions

You are converting behavioral instincts into an executable GSD skill. The resulting SKILL.md must be a real, useful skill — not a placeholder. It must follow the same format as /learn, /learn-eval, and /evolve.

Argument Parsing

Parse the argument passed to this skill:

• If the argument starts with CLU-, it is a cluster ID.
• If the argument starts with INS-, it is a single instinct ID.
• If neither, report "Unrecognized argument format — expected CLU-XXXXXXXX-XXXX or INS-XXXXXXXX-XXXX" and stop.

Step 1: Load Source

From a CLU- cluster ID:*

Locate the cluster file:

find "$HOME/.claude/instincts/global/clusters/" -name "<cluster-id>.md" 2>/dev/null

Read the file. Parse the frontmatter to extract:

• id, name, member_ids, avg_confidence, trigger_family, skill_hypothesis
• status

Read each member instinct file listed in member_ids:

find "$HOME/.claude/instincts/" -name "<ins-id>.md" 2>/dev/null

Parse each member instinct's frontmatter and ## Operational Rule body. Collect:

• id, name, status, confidence, description
• signals.trigger_patterns
• The full rule text

From a single INS- instinct ID:*

Locate the instinct file:

find "$HOME/.claude/instincts/" -name "<instinct-id>.md" 2>/dev/null

Read the file. Set source_cluster: null. Use the instinct itself as the sole member.

Step 2: Validate Members

All member instincts must have status == "proven" or status == "active".

Skip any member with status == "candidate" or status == "deprecated" — warn in the report but do not abort.

If zero members pass validation (all were candidates or deprecated), output:

Cannot create skill: no proven or active member instincts found.
Members found: N (all candidates or deprecated)
Run /learn-eval first to mature these instincts, then retry.

Then stop.

Step 3: Derive Skill Name

From cluster: Convert the cluster's name field to kebab-case. Drop any CLU/INS prefix. Use only lowercase letters, numbers, and hyphens.

From single instinct: Convert the instinct's name field to kebab-case the same way.

Examples:

• "Slack Unread Detection" → slack-unread-detection
• "validate-roster-before-outbound-comms" → already kebab, use as-is
• "Weekly Digest -- Concise Format" → weekly-digest-concise-format

Collision check:

[ -d "$HOME/.claude/skills/<skill-name>" ] && echo "EXISTS" || echo "FREE"

If the directory already exists, append a numeric suffix: <skill-name>-2, <skill-name>-3, etc. Check until a free name is found.

Step 4: Generate the SKILL.md

Create the directory and write the skill:

mkdir -p "$HOME/.claude/skills/<skill-name>"

Frontmatter — use this exact format:

---
name: <skill-name>
description: "<one-line from skill_hypothesis (cluster) or instinct description (single)>"
argument-hint: ""
allowed-tools:
  - Read
  - Write
  - Bash
  - Grep
  - Glob
---

Body — generate a real, runnable skill following the /learn style:

1. Open with an <objective> section:
   • For clusters: derive from skill_hypothesis
   • For single instincts: derive from the instinct's description
2. Write a ## Instructions section that synthesizes all member operational rules into a coherent, non-redundant process:
   • Read all member ## Operational Rule bodies
   • Identify the common behavioral theme
   • Produce step-by-step instructions that enforce the combined behavioral rules
   • Write prescriptive imperatives ("Always X when Y", "Never X unless Y")
   • Where members address different sub-concerns, organize them into numbered steps
   • Keep all trigger patterns and anti-patterns from members
3. Close with a ## Important Constraints section that enumerates all the "never" / "always" / boundary conditions from the member rules.

The generated skill must be actionable — an agent reading it must know exactly what to do. It must not say "follow best practices" or "be careful". Every statement must be a concrete instruction.

Minimum body length: 300 words. If the source instincts produce fewer than 300 words of synthesized instruction, expand with:

• Explicit trigger examples
• Anti-pattern examples (what NOT to do)
• Edge case handling from the instinct's trigger_patterns / anti_patterns

Step 5: Write SKL-*.md Reference

Ensure the skills directory exists:

mkdir -p "$HOME/.claude/instincts/global/skills"

Generate a unique SKL ID:

SUFFIX=$(python3 -c "import secrets; print(secrets.token_hex(2))")
SKL_ID="SKL-$(date +%Y%m%d)-${SUFFIX}"

Write ~/.claude/instincts/global/skills/${SKL_ID}.md:

---
id: <SKL_ID>
skill_name: "<skill-name>"
source_cluster: <CLU-ID | null>
source_instincts: [INS-..., ...]
created_at: "<ISO-8601 UTC now>"
---
## Skill Reference

**Generated from**: <CLU-XXXXXXXX-XXXX "<cluster name>" | INS-XXXXXXXX-XXXX "<instinct name>">
**Skill path**: ~/.claude/skills/<skill-name>/SKILL.md
**Members**: <N> instinct(s)

### Source Summary
<One paragraph summarizing what behavioral patterns were synthesized into this skill.>

Step 6: Update Lineage

Cluster file (if source was a cluster):

Add skill_id: <skill-name> to the cluster's YAML frontmatter. Use the confidence library's update helper if available:

python3 -c "
import sys
sys.path.insert(0, '$HOME/.claude/instincts/lib')
from confidence import update_instinct_file
update_instinct_file('<CLU_FILE_PATH>', {
    'skill_id': '<skill-name>',
    'updated_at': '<ISO_NOW>',
})
"

If the skill_id field does not exist in the cluster frontmatter yet, append it before the closing --- using the Edit tool.

Member instinct files:

For each member instinct that passed validation, add skill_created: <skill-name> to their lineage block:

python3 -c "
import sys
sys.path.insert(0, '$HOME/.claude/instincts/lib')
from confidence import update_instinct_file
update_instinct_file('<INSTINCT_PATH>', {
    'lineage.skill_created': '<skill-name>',
    'updated_at': '<ISO_NOW>',
})
"

If the skill_created sub-key does not exist yet inside lineage:, append it there using the Edit tool.

Step 7: Update INSTINCTS.md

Read ~/.claude/instincts/global/INSTINCTS.md. Find the ## Skills section. If it does not exist, append it at the end of the file.

Add one entry to the ## Skills section:

- <SKL_ID> | "<skill-name>" | from: <CLU-ID | INS-ID> | created <YYYY-MM-DD>

Replace (none yet) if that placeholder is present.

Step 8: Report

Output a structured report:

## Skill Creation Report

**Source**: <CLU-XXX "<cluster name>" | INS-XXX "<instinct name>">
**Members**: N instinct(s) (<N> valid, <N> skipped)

### Generated Skill
- Name:      <skill-name>
- Path:      ~/.claude/skills/<skill-name>/SKILL.md
- Reference: <SKL-ID>

### Lineage Updated
<For each source — list the CLU or INS IDs and what field was added>
- CLU-XXX: skill_id = <skill-name>
- INS-XXX: lineage.skill_created = <skill-name>
- INS-YYY: lineage.skill_created = <skill-name>

### Skipped Members (N)
<Only present if any members were skipped>
- INS-ZZZ | "<name>" | status: candidate — not mature enough

If the skill directory already existed and a suffix was appended, note it:

Note: ~/.claude/skills/<original-name>/ already existed — created as <skill-name> instead.

Important Constraints

• Never overwrite an existing skill directory. If ~/.claude/skills/<skill-name>/ exists, append a numeric suffix (-2, -3, ...). Never delete existing skills.
• Ensure directories exist before writing any files (mkdir -p for both the skill dir and ~/.claude/instincts/global/skills/).
• The generated skill must be real and actionable. It must read like /learn or /evolve — concrete steps, prescriptive instructions, no vague guidance. A reader should be able to execute it without ambiguity.
• Preserve all member lineage. Update every valid member's skill_created field. Do not skip members silently.
• Do not create new instincts during this process. This skill only synthesizes existing ones.
• Skip candidates and deprecated instincts from synthesis. Include them in the "Skipped" report section but never let them contribute to the skill body — immature instincts corrupt skill quality.
• Batch Python invocations where possible. Do not spawn a separate process for each instinct update.