Auto Skill

Auto-Skill Creation

When you complete a complex task with 8+ tool calls, or recognise a repeating pattern, draft a skill candidate. Drafts stage in skills/drafts/ until approved — they are never auto-loaded into live sessions.

Step 1: Post-Task Detection Check

After completing any task, run this self-check:

1. Tool call count: Did this task require 8+ distinct tool calls for a coherent workflow?
2. Recurrence: Does your daily memory show this same task type appearing 3+ times across different dates?
3. Existing skill: Does a skill for this already exist in .claude/skills/? If yes, stop — consider proposing a patch instead.
4. Repeatability: Is this task type likely to recur, or was it a one-off?

If yes to 1 or 2, AND no to 3, AND yes to 4 → create a draft.

Do NOT draft for:

• Routine heartbeat operations
• One-off research tasks
• Tasks already covered by existing skills
• Simple single-step operations

Step 2: Draft the Skill

Create the draft at skills/drafts/[skill-name]/SKILL.md:

mkdir -p skills/drafts/[skill-name]

Use this template:

---
name: skill-name-here
description: One sentence, max 100 chars. What this skill does and when to use it.
created: YYYY-MM-DD
created_by: auto
trigger: "Natural language description of when this skill fires"
source_task_id: TASK-ID-THAT-GENERATED-THIS
version: 1
status: draft
---

## Purpose

[What this skill does. 2-3 sentences max.]

## When to Use

[Specific conditions that should trigger this skill. Be concrete.]

## Inputs Required

- `$PARAM_1` — description
- `$PARAM_2` — description (optional)

## Steps

1. [Step 1 — specific, actionable]
2. [Step 2]
3. ...

## Output

[What gets created or sent when this skill completes.]

## Approval Gate

[Does any step require user approval? Specify exactly which step.]

## Notes / Edge Cases

[Known failure modes, dedup considerations, platform quirks.]

Step 3: Self-Review Before Saving

Before writing the file, check:

• description is under 100 characters
• No skill with this name exists in .claude/skills/
• All steps are actionable — no vague instructions like "handle errors appropriately"
• Any step requiring external actions (email, deploy, post, delete) has an explicit approval gate documented
• source_task_id is filled in for traceability

If any check fails, revise before saving.

Step 4: Log and Notify

After saving the draft:

# Log the creation
cortextos bus log-event action skill_draft_created info --meta "{\"skill\":\"[skill-name]\",\"source_task\":\"[task_id]\",\"agent\":\"$CTX_AGENT_NAME\"}"

# Notify orchestrator — it will surface in the next morning digest
cortextos bus send-message $CTX_ORCHESTRATOR_AGENT normal "Skill candidate drafted: [skill-name] — source task [task_id]. Check skills/drafts/[skill-name]/SKILL.md. Awaiting digest review."

The orchestrator includes it in the next morning briefing for the user. The user replies approve [skill-name], reject [skill-name] [reason], or revise [skill-name] [feedback].

Step 5: Handling Approval Responses

When you receive an inbox message with a skill decision:

Approved

# Move from draft to active
mv skills/drafts/[skill-name]/ .claude/skills/[skill-name]/

# Update status field
sed -i '' 's/status: draft/status: active/' .claude/skills/[skill-name]/SKILL.md

# Log activation
cortextos bus log-event action skill_activated info --meta "{\"skill\":\"[skill-name]\",\"agent\":\"$CTX_AGENT_NAME\"}"

# Notify the user
cortextos bus send-telegram $CTX_TELEGRAM_CHAT_ID "Skill activated: [skill-name] is now live and will be used in future sessions."

# ACK the inbox message
cortextos bus ack-inbox [msg_id]

Rejected

# Move to archive with reason recorded
mkdir -p skills/archive/[skill-name]
mv skills/drafts/[skill-name]/SKILL.md skills/archive/[skill-name]/SKILL.md

# Append rejection reason to archived skill
echo "\n## Rejection\nReason: [reason]\nDate: $(date -u +%Y-%m-%d)" >> skills/archive/[skill-name]/SKILL.md

# Log
cortextos bus log-event action skill_rejected info --meta "{\"skill\":\"[skill-name]\",\"reason\":\"[reason]\"}"

cortextos bus ack-inbox [msg_id]

Revise

# Apply feedback and re-save the draft
# Then re-notify orchestrator for another digest cycle
cortextos bus send-message $CTX_ORCHESTRATOR_AGENT normal "Skill candidate [skill-name] revised per feedback. Ready for re-review."
cortextos bus ack-inbox [msg_id]

Draft Lifecycle

Drafts older than 14 days with no action: move to skills/archive/ and log skill_expired.

Critical Rules

1. No automated commits — skill files are never committed without explicit user approval
2. Draft means invisible — agents do NOT load or execute skills from skills/drafts/
3. No self-modification — skills never modify other skills (v2 feature)
4. No external calls without gates — any skill step with external side effects must document its own approval requirement