Level Up
Run a 7-question adaptive assessment to map what the user knows and doesn't, relevant to the current project. The output is two files future agents rely on.
Files (repo-relative)
notes/learning/user-knowledge.md — verbatim Q&A pairs + ratings, one section per question, rounds appended.notes/learning/LEARNING-PLAN.md — one concise bullet per genuine gap found.
State-check first: read both files in full if they exist. If previous rounds exist, pick mostly-new territory and calibrate starting difficulty to the recorded level. If missing, create the folder and both files (plan starts as just a header).
Question rules
- 7 questions, strictly one at a time, plain text — never the questions UI.
- Start easy, adapt difficulty each answer: good answer → harder, weak answer → sideways or down.
- Orchestrator level only: systems, architecture, failure modes, security, data, scaling, product strategy, unit economics. NEVER syntax or code trivia — the user architects via AI agents, they don't write code.
- Anchor questions in the current project's real stack and features. When a question touches real code, read it and show the actual snippet when teaching.
- Cover different territory across rounds (e.g. round 1: request flow, DB, billing, moats; round 2: deploys, testing, incidents, data modeling, AI engineering, webhook security, cost engineering).
Feedback Principle
Stanford tested two ways of giving feedback to students. Group one got standard comments. Group two got the same comments with one added sentence: "I'm giving you these comments because I have very high expectations and I know you can reach them." Group two improved at four times the rate as group one, despite having the same correction.
After every single answer
- Rate honestly 1-10. No flattery — the user wants calibration, not comfort.
- Say concisely what was missed or wrong, and teach the correct concept in a few sentences.
- Immediately save the verbatim answer + rating + gap notes to
user-knowledge.md.
- If a genuine gap surfaced, append one concise bullet to
LEARNING-PLAN.md. Skip minor misses.
- If the user pushes back on a rating ("I knew that, just didn't say it"), bump only if genuinely deserved, and record the bump with its reason.
- When the user says they have since learned a plan item, mark its bullet: strikethrough +
✓ learned YYYY-MM-DD.
After question 7
Append a final summary to user-knowledge.md: per-question ratings, overall score, the recurring pattern across answers (e.g. "architecture instincts ahead of failure-mode instincts"), strengths to build on, and gaps added. Give the user the same summary in chat, concise.
1---2name: level-up3description: Gauge the user's technical + product knowledge through 7 adaptive questions, log verbatim answers with honest ratings, and grow a learning plan from the gaps found. Use when the user says "level up", "level-up session", "quiz me", "gauge my knowledge", or wants a new assessment round. Differentiator: this finds and maps gaps; the `teach` skill delivers lessons on them.4---56# Level Up78Run a 7-question adaptive assessment to map what the user knows and doesn't, relevant to the current project. The output is two files future agents rely on.910## Files (repo-relative)1112- `notes/learning/user-knowledge.md` — verbatim Q&A pairs + ratings, one section per question, rounds appended.13- `notes/learning/LEARNING-PLAN.md` — one concise bullet per genuine gap found.1415State-check first: read both files in full if they exist. If previous rounds exist, pick mostly-new territory and calibrate starting difficulty to the recorded level. If missing, create the folder and both files (plan starts as just a header).1617## Question rules1819- 7 questions, strictly one at a time, plain text — never the questions UI.20- Start easy, adapt difficulty each answer: good answer → harder, weak answer → sideways or down.21- Orchestrator level only: systems, architecture, failure modes, security, data, scaling, product strategy, unit economics. NEVER syntax or code trivia — the user architects via AI agents, they don't write code.22- Anchor questions in the current project's real stack and features. When a question touches real code, read it and show the actual snippet when teaching.23- Cover different territory across rounds (e.g. round 1: request flow, DB, billing, moats; round 2: deploys, testing, incidents, data modeling, AI engineering, webhook security, cost engineering).2425## Feedback Principle2627Stanford tested two ways of giving feedback to students. Group one got standard comments. Group two got the same comments with one added sentence: "I'm giving you these comments because I have very high expectations and I know you can reach them." Group two improved at four times the rate as group one, despite having the same correction.2829## After every single answer30311. Rate honestly 1-10. No flattery — the user wants calibration, not comfort.322. Say concisely what was missed or wrong, and teach the correct concept in a few sentences.333. Immediately save the verbatim answer + rating + gap notes to `user-knowledge.md`.344. If a genuine gap surfaced, append one concise bullet to `LEARNING-PLAN.md`. Skip minor misses.355. If the user pushes back on a rating ("I knew that, just didn't say it"), bump only if genuinely deserved, and record the bump with its reason.366. When the user says they have since learned a plan item, mark its bullet: strikethrough + `✓ learned YYYY-MM-DD`.3738## After question 73940Append a final summary to `user-knowledge.md`: per-question ratings, overall score, the recurring pattern across answers (e.g. "architecture instincts ahead of failure-mode instincts"), strengths to build on, and gaps added. Give the user the same summary in chat, concise.
David Ondrej