Turn AI into a personal tutor: honest diagnosis, a week-by-week plan, spaced repetition and measurable tracking of your progress.
Open this lesson in KodokonA senior developer learning a new field almost always makes the same mistake: overestimating their transferable skills and underestimating their blind spots. Asking an AI directly for a learning plan makes the problem worse, because the model hands you a generic program calibrated on an average profile that doesn't exist. The right sequence is therefore diagnosis first, plan second. The diagnosis must be adaptive - each question adjusts to your previous answer, exactly like a human examiner who digs in wherever you falter. This is precisely what a conversational model is good at, provided you constrain it: without explicit rules, it drifts back to its default behavior, which is to teach you instead of assessing you.
You are a senior technical tutor tasked with establishing my skills diagnosis before building a learning program.
Context: I'm a developer with [X] years of experience in [YOUR CURRENT FIELDS] and I want to progress in [TARGET FIELD, e.g. Rust, distributed systems].
Mandatory flow:
1. Ask me 10 questions of increasing difficulty, ONE AT A TIME. Wait for my answer before the next one.
2. Do not comment on my answers during the test: no corrections, no praise.
3. Adapt each question to my previous answer: harder if I succeed, a lateral question on the same topic if I fail.
4. Include at least 2 questions where I have to write code from memory and 2 questions on edge cases.
Only at the end, deliver an assessment in this format:
- Solid strengths (with the evidence taken from my answers)
- Weaknesses (with the exact mistake I made)
- Likely gaps not tested
- The top 3 learning priorities, each justified in one sentence.
Forbidden: do not teach me anything during the diagnosis.Once you have the assessment in hand, make it the raw material for the plan. A good learning plan follows three principles validated by cognitive science research: deliberate practice (exercises just above your level, never comfortable), interleaving (mixing topics rather than treating them in sealed blocks) and spaced repetition (reviewing a concept just before you forget it, at increasing intervals). Also demand a verifiable success criterion for each week: "understand lifetimes" isn't verifiable, "get this program to compile without a superfluous clone()" is.
You are an instructional engineer specialized in training experienced developers.
Based on the diagnostic assessment below, build a 6-week learning plan to reach: [MEASURABLE GOAL, e.g. being able to write a concurrent HTTP service in Rust without assistance].
[PASTE THE DIAGNOSTIC ASSESSMENT HERE]
Constraints:
- 45 minutes a day, 5 days a week, no more.
- Each week: one verifiable goal, one mini-project, and 2 review sessions covering previous weeks (spaced repetition).
- 70% practice, 30% theory maximum.
- No content I could copy-paste: only exercises where I produce the work myself.
- Topics must be interleaved, not treated as isolated blocks.
Expected format: a week-by-week table with the columns Goal / Exercises / Reviews / Success criterion.
End with the exact question you will ask me at the end of week 1 to check that the goal has been reached.The weak link in any self-directed learning is review. The forgetting curve is brutal: without recall, you lose most of a concept within a few days. The proven countermeasure is active recall - retrieving information from memory, not rereading it - at increasing intervals (1, 3, 7, 14, 30 days). AI excels here, on one condition: give it a learning journal. Keep a text file where each session adds a line: date, concept, mistakes made. This journal becomes the persistent state the model lacks between two conversations.
You are my spaced repetition system. Here is my learning journal (concepts studied, dates, mistakes made):
[PASTE THE JOURNAL]
Your mission:
1. Select the 5 most urgent concepts to review based on their last-recall date and my past failures (absolute priority to concepts where I've already made mistakes).
2. Ask me one active-recall question per concept: never multiple choice, only open questions or code to write from memory.
3. One question at a time. After my answer: correct me, then tell me in how many days to review this concept (1, 3, 7, 14 or 30) based on the quality of my answer.
Final format: a list "concept - next review - justification" that I'll copy as-is into my journal.
Forbidden: never re-explain a concept until I've attempted an answer.Then there's tracking. A plan that never recalibrates becomes a fiction within two weeks. Schedule a light re-diagnosis every two weeks: five targeted questions on past goals, generated by the AI from your journal. Two signals mean it's time to revise the plan: you get everything right on the first try (the plan is too easy, you're no longer in deliberate practice) or you fail recalls on concepts seen more than two weeks ago (the review pace is too loose). In both cases, give the assessment back to the AI and ask for a corrected plan - never a fresh plan, always a justified diff.