Prompt Mirror — Concept Lens

Your prompts are a mirror. They reveal which technical concepts you understand intuitively and which ones have gaps. This skill reads yesterday's session prompts, identifies one concept worth understanding, and delivers it as a morning lesson: everyday analogy first, engineer vocabulary second.

Goal

Talk like an engineer — without writing code. You don't read or write code, but you understand concepts, judge AI output quality, and participate in technical conversations with confidence.

How it works

1. Extract — Run scripts/extract_prompts.py YYYY-MM-DD to collect all user prompts from yesterday's sessions
2. Analyze — Read the extracted prompts and identify concept gaps:
   • Technical terms used inaccurately
   • Related but distinct concepts being confused
   • Moments where AI output quality couldn't be judged (or was caught)
   • Concepts that engineers would use in conversation, used implicitly
3. Select one — Pick the highest-impact concept. Check data/taught.jsonl to avoid repeats
4. Deliver — Write a concise lesson to the Journal using the template in templates/lesson.md

Analysis Guidelines — Concept Lens

• Hunt for concept gaps, not code mistakes — Not "you should use first_ts" but "file modification date and session start date are different concepts"
• Start with everyday examples — Like: "The 'last edited' date on a Google Doc is not the same as when it was created" — analogies anyone can grasp
• Add engineer vocabulary — After the analogy, one line: "Engineers call this mtime vs created_at"
• Anchor to yesterday's experience — Not abstract explanations, but "this happened in your session yesterday, and knowing this concept would have prevented it"
• Distinguish delegated prompts — See gotchas.md. Prompts sent via orchestrators (e.g., AGI Cockpit) to sub-agents may be AI-generated, not human-written
• Don't force it — If nothing surfaces, "No lesson today" is a valid output

What NOT to Teach

• How to write or read code (the user doesn't code)
• Tool-specific command syntax (that's what docs are for)
• Things already corrected within the session
• Typos or language slips

Config

Read config.json in this skill directory. If missing, use defaults:

• Journal dir: ~/Documents/Journal/ (or configure your own path)
• Projects dir: ~/.claude/projects
• Codex dir: ~/.codex

Gotchas

See gotchas.md — failure patterns discovered in practice. Read every run.

Memory

After delivering a lesson, append to data/taught.jsonl:

{"date":"YYYY-MM-DD","topic":"one-line topic","category":"concept|vocabulary|mental-model","engineer_term":"the technical term taught","prompt_examples":["truncated example"]}

On each run, read this log to:

• Avoid teaching the same topic twice
• Track which categories are most frequent
• Spot progression (are the concepts getting more advanced over time?)

Jaggedness Map

Inspired by Andrej Karpathy's observation that AI models are "simultaneously a brilliant PhD student and a 10-year-old" — humans have the same jaggedness in their prompt capabilities. This feature builds a personal strength/weakness map over time.

How it works

After 7+ entries in data/taught.jsonl, generate or update data/jaggedness-map.json:

{
  "updated": "YYYY-MM-DD",
  "total_lessons": 12,
  "strong": [
    {"area": "Architecture delegation", "evidence": "Consistently clear system design prompts"},
    {"area": "Requirements definition", "evidence": "Scope and constraints well-specified"}
  ],
  "weak": [
    {"area": "Implicit references", "evidence": "3 lessons on ambiguity patterns"},
    {"area": "Data pipeline concepts", "evidence": "Confused event time vs processing time, precision vs recall"}
  ],
  "improving": [
    {"area": "Security awareness", "evidence": "Secrets management lesson led to behavior change"}
  ],
  "no_lesson_days": 6,
  "no_lesson_trend": "increasing — prompt quality improving"
}

When to generate

• Weekly: After 7+ taught.jsonl entries, append a ## Prompt Mirror — Jaggedness Map section to the weekly journal entry
• On demand: When the user asks "show my prompt map" or "jaggedness map"

Analysis sources

1. data/taught.jsonl — What concepts were taught (= weak areas that needed teaching)
2. Prompt extraction — What areas the user prompts confidently in (= strong areas)
3. "No lesson" days — Increasing trend = overall improvement

Map rules

• Strong = areas where prompts are consistently clear, specific, and technically accurate. No lessons needed
• Weak = areas where multiple lessons clustered. 2+ lessons in same domain = pattern
• Improving = areas where a lesson was taught AND subsequent prompts show the concept being applied correctly
• Don't fabricate strengths. Only mark "strong" if there's positive evidence from prompt patterns