auto-update-skills

name: auto-update-skills description: After correction/feedback: propose new skill or update existing to capture reusable pattern. Covers knowledge, tools, policies, preferred style. Trigger immediately for critical issues, on repetition for trivial patterns.

Goal

Capture lessons from corrections by proposing a new skill that preserves existing knowledge, avoids repeated mistakes, and aligns with preferred behavior.

Also ensure existing skills are respected by extending them when they should have applied but lacked a trigger or instruction.

When to Use

• The user corrects a prior proposal, implementation, communication style or different chat pattern.
• The correction reveals a reusable rule, tool, policy, or preferred pattern.

Preconditions

• A correction or review has occurred in the current conversation.
• The correction is relevant beyond the single instance.
• A quick scan of existing skills shows no match, or a near match that can be fixed by updating the existing skill.

Definitions

• Critical issue: A correction that affects UI/UX, API, DB scheme, architecture, safety, compliance, legal risk, financial impact, or major business outcomes. Propose a new skill immediately on first occurrence.
• Trivial pattern: A small or low‑impact correction (formatting, minor naming, minor style preferences). Propose a new skill only after repetition.

Illustrative Categories (Non‑exhaustive)

Critical examples

• Programming: API changes, database schema changes, UI modifications, programming patterns.
• Engineering: Security or compliance rule (e.g., use approved validation utilities, avoid direct credential handling).
• Product: Regulatory constraints or roadmap policy that changes what can be shipped.
• Business: Pricing or contractual commitments that must be enforced.
• HR: Hiring or privacy rules that must be followed.
• Accounting: Tax or audit requirements that change reporting behavior.
• Legal: Contract clauses or jurisdiction requirements that alter guidance.

Trivial examples

• Engineering: Preferred code style or naming convention.
• Product: Template wording or formatting of a status update.
• Business: Slide layout preference.
• HR: Preferred phrasing in internal comms.
• Accounting: Report section ordering.
• Legal: Minor citation format preference.

Rules

1. If the correction is critical, propose a new skill immediately.
2. If the correction is trivial, propose a new skill only after the pattern repeats.
3. If the correction changes the assistant’s communication style, propose a new skill immediately if it is a strong preference or policy, otherwise wait for repetition.
4. Always propose a positive, goal‑oriented skill name (e.g., “remember‑existing‑knowledge”).
5. The skill proposal must include a one‑sentence Why now justification.
6. Before proposing a new skill, check existing skills by name and description to avoid duplicates.
7. If an existing skill should have triggered but did not, update that skill instead of proposing a new one.
8. SKILL.md under 500 lines. Split detailed content into reference files. Load only what's needed for the current workflow.
9. Ensure YAML frontmatter is valid, the name matches the directory and uses lowercase hyphens, and the description states what it does and when to use it.
10. SKILL.md is always loaded on activation, so keep essential principles and guardrails in SKILL.md and move detailed procedures to workflows or references.
11. If a skill grows complex, use a router pattern with workflows, references, templates, and scripts instead of expanding SKILL.md. Keep the rule max 500 lines for SKILL.md files.

Procedure

1. Identify the correction and classify it as critical or trivial.
2. Scan existing skills for a matching or near matching description.
3. If a matching skill exists, update its SKILL.md to include the missing trigger, condition, or instruction.
4. When updating or creating a skill, verify frontmatter and placement rules from this skill.
5. If no matching skill exists, decide whether to propose a new skill based on the Rules.
6. Draft a proposed skill (name + description + key rules).
7. Ask the user to confirm creation of the new skill.

Output Format

Use the language of the last messages; if mixed, use the most recent message.

Creating new skill proposal:

Proposal: Create a new skill
Name: <proposed-skill-name>
Description: <1–2 sentence description>
Key rules:
- <rule 1>
- <rule 2>
Why now: <one sentence>

Create this skill now? (yes/no)

Updating existing skill proposal:

Proposal: Update existing skill
Name: <updated-skill-name>
Description: <1–2 sentence description>
Key updates:
- <update 1>
- <update 2>
Why now: <one sentence>

Update this skill now? (yes/no)

Examples

• After a code review reveals an existing validation utility must be used, propose a skill that prioritizes approved shared utilities.
• After a PM corrects a roadmap constraint that changes scope decisions, propose a skill capturing that constraint.
• After a style correction that repeats across messages, propose a skill capturing the preferred tone or format.

Avoid

• Creating a skill for a one‑off correction with no reuse value.
• Proposing a skill that merely restates the user’s prompt without a reusable rule.
• Ignoring the critical vs. trivial distinction.
• Using a negative or prohibitive skill name.
• Skipping the user confirmation step.
• Duplicating an existing skill when a small update would solve the issue.