Agent-Skills.md

Agent Skills: Learning System

Self-improvement system for managing (finding/using, saving/adding, and pruning) solutions, patterns, lessons learned, and workarounds. Use this skill whenever a mode says "check learnings", "check memories", "prior solutions", "remember this", "save this pattern", or when you discover a non-obvious fix after 2+ failed attempts. Also trigger when encountering unexpected behavior that others might hit. Load with read_file on .kilocode/skills/learning/SKILL.md (ignore the absolute path in the location tag).

UncategorizedID: ScotterMonk/AgentAutoFlow/learning

Repository

ScotterMonk/AgentAutoFlow
ScotterMonkLicense: CC0-1.0
371

Install this agent skill to your local

pnpm dlx add-skill https://github.com/ScotterMonk/AgentAutoFlow/tree/HEAD/.kilocode/skills/learning

Skill Files

Browse the full folder contents for learning.

Download Skill

Loading file tree…

.kilocode/skills/learning/SKILL.md

Skill Metadata

Name
learning
Description
Self-improvement system for managing (finding/using, saving/adding, and pruning) solutions, patterns, lessons learned, and workarounds. Use this skill whenever a mode says "check learnings", "check memories", "prior solutions", "remember this", "save this pattern", or when you discover a non-obvious fix after 2+ failed attempts. Also trigger when encountering unexpected behavior that others might hit. Load with read_file on .kilocode/skills/learning/SKILL.md (ignore the absolute path in the location tag).

Learning System

Purpose: Knowledge base for solutions, patterns, lessons, and workarounds.

Definitions

learnings folder: Read skill-local guidance only when .kilocode/skills/learning/AGENTS.md is confirmed to exist; otherwise use root AGENTS.md for the project-specific learning folder and archive location. File format: {category}-{subcategory}.md (all lowercase, hyphen-separated).

Category examples: python, flask, javascript, sql, css, testing, sync, config, git, patterns. Subcategory examples: imports, routing, async, errors, performance, edge-cases. (Example: python-imports.md, flask-routing.md, testing-edge-cases.md)

Workflow

Entry routing:

  • Caller says "check learnings", "prior solutions", "patterns", "memories" → Start at Phase 1: Find.
  • Caller says "save this", "remember this", "lesson learned", or you discover a non-obvious insight → Start at Phase 2: Evaluate.

Phase 1: Find Learnings (when stuck or researching)

  1. Search the {learnings folder} using keyword search matching the current problem domain.
  2. Read only the matching file(s) or relevant sections.
  3. Apply any matching Fix/Correct entries to the current task.
  4. If nothing relevant found, continue without learnings — do not block progress.

Phase 2: Evaluate (Fast Fail)

Analyze the conversation for reusable, non-obvious insights. Assign a Tier:

  • Tier 1 (Global/Critical): Highly reusable, prevents major errors. (Save to learnings folder)
  • Tier 2 (Domain/Useful): Feature-specific, nice-to-know. (Save to learnings folder)
  • Tier 3 (Local/Obvious): One-off edge case, standard documentation. (Skip/Exit immediately)

If Tier 3, exit the learning skill immediately without prompting the user.

Phase 3: Draft Content

For any file writes, follow ## Formatting Standards in the coding-markdown skill.

Before drafting, check for duplicates:

  • Search existing learning files for similar entries.
  • Exact match found → Exit skill (already captured).
  • Overlapping/related found → Append to or update the existing entry instead of creating a new file.

Use telegraphic language (→ ✅ ❌ ⚠️). Hard limit: ≤30 lines per entry.

Unified Template (for bug fixes, errors, and gotchas):

## [Topic/Issue Name]
<!-- meta: date=YYYY-MM-DD, tier=[1|2], status=active -->
- **Trigger/Symptom**: [What happened or when to use]
- **Cause/Mistake**: [Why it failed or common wrong approach]
- **Fix/Correct**: [The solution or right approach]
- **Why**: [1-line reason]

(Optional: If the issue took 3+ attempts to solve, append a brief Timeline: and Blind Spots: section.)

For patterns and techniques (not errors), a prose format with a "Recommended pattern" section is acceptable — but still include the <!-- meta --> comment for date tracking.

Phase 4: Micro-Approval (BLOCKING)

Present a highly condensed summary to the user for approval:

💡 Proposed Learning: [Category]

  • Issue: [1-line description]

  • Fix: [1-line solution] Save to [File]? (y/n)

  • User approves → Save to file.

  • User declines → Discard silently. Do not re-propose the same learning.

  • User requests changes → Revise the draft and re-present for approval.

Phase 5: Just-in-Time (JIT) Maintenance

Agents do not run on a schedule. Maintenance happens during active use:

  • When reading an existing learning file to solve a current problem, check the date metadata.
  • If an entry is > 6 months old AND you determine it is obsolete/superseded, propose to the user as part of your current response:
    • Archive: Move the entry to the learning archive folder (unchanged filename).
    • Delete: Remove entirely (only if the entry is proven incorrect, not just outdated).