AI doesn't know about your school.

Ask Claude: what is “Mars Pro”? You'll get an answer about a planet or some software. Why doesn't AI know about your school? Because Mars IT School is niche/local. In this module — how to describe a task so AI doesn't guess, and how to understand where it's guessing by default.

Hook — “Mars Pro”? AI thinks about a planet.AI o'qituvchining maktabini bilmaydi

Mental model — what AI knows and how to tell it.AI nimani biladi va unga qanday tushuntirish

Two layers: Knowledge — what the model has in its head from training. Description — how you package the task. Each is a separate subsection. Verbatim quotes from Anthropic + translation. If the language gets fuzzy, trust the original next to it.

1. Knowledge spectrum — frequently · recently · consistently

The practical question isn't "does AI know this?", but "how well was it represented in what it read?". A four-zone scale:

2. Three levels of Description — product · process · performance

3. Description ↔ Discernment loop — the course's engine

Antipattern #1 in Mars: a mentor writes one long prompt → gets the wrong thing → blames the model. Pattern: short prompt → you look at what the AI did → you understand which layer was missing → you add to it. After 2-3 rounds — a working template that you then reuse.

Six techniques prompting — from Anthropic Fluency.Oltita texnika — prompting uchun

Verbatim from AI Fluency Framework lesson 9. Each line is a technique + where it maps to product/process/performance. Remember these six, and you'll cover 90% of cases.

Anthropic calls it a “secret weapon”: “Ask the AI itself to help improve your prompt.” Finished the prompt — ask Claude: “what in this request is unclear to you?” Often a whole layer you hadn't considered emerges from that.

Demo — four iterations of one prompt.Bitta prompt — to'rtta variant

Zuhra, a Python mentor from Yunusobod. Monday morning, she needs to prepare material for a group of 12-year-olds. The same basic request. Each version adds one Description layer. Mock-output shows what changes in quality.

V1 · bare — no layers

$ claude

› come up with 5 Python assignments for 12-year-olds

⠋ Done:

1. Sum of squares from 1 to n
2. Check if a number is prime
3. FizzBuzz
4. Guess the number (random)
5. Count vowels in a string

What's wrong: technically valid, but it's a generic Hour-of-Code set. For which 12-year-olds exactly? Just started? Already know conditions? What's the topic — loops, strings, dictionaries? It's completely unclear how to evaluate these. Discernment signal: the prompt hit mainstream AI knowledge and pulled the most average from the training data.

V2 · +product — what I want as output

› come up with 5 Python assignments, topic for/while,
  for 12-year-olds, who already know if/else but have just started loops.
  Format: 1 sentence condition + 1 example input/output.

⠋ Done:

1. Print numbers from 1 to 10, each on a new line.
   Example: input — none, output — 1\n2\n...\n10
2. Calculate the sum of all even numbers from 1 to 20.
   Example: input — none, output — 110
3. Print the multiplication table for 5 (up to 5×10).
4. Until the user types "stop", repeat the greeting.
5. Find the first number greater than 1 whose square is greater than 100.

Better: now on topic (loops), now considers the level (they know if/else). But it's still abstract, like school. "Sum of evens" for a 12-year-old — why? What's a typical mistake? What's the pedagogical approach? The AI doesn't know because we haven't told it how to think.

V3 · +process — how to think

› the same 5 assignments on for/while for 12-year-olds (they know if/else).
  For each assignment, think in steps:
    1. concept (what specific loop skill are we mastering)
    2. real-life example for a child (not "sum of evens" — something about
       their world: WhatsApp, TikTok, homework, recess)
    3. most common expected mistake (forgot increment? off-by-one?)
    4. assignment

⠋ Done:

1. CONCEPT: for with range
   REAL-LIFE: scrolling through a TikTok playlist — each video is 15 seconds
   MISTAKE: they'll forget that range(5) gives 0..4, not 1..5
   ASSIGNMENT: "Print 'video N (15 sec)' for 8 videos."

2. CONCEPT: while with an exit condition
   REAL-LIFE: waiting for a friend's reply on WhatsApp — until it arrives
   MISTAKE: infinite loop — they'll forget to change the variable
   ASSIGNMENT: "Until you get 'ok' — keep printing
                'waiting...' with a pause." (show time.sleep)

3-5. [similarly — push-up counter, recess countdown,
        checking if all friends have replied]

Much better: The AI started giving pedagogically meaningful material. "Expected mistake" is almost a lesson plan. But we still haven't told it how you'll know if an assignment is strong. What if the AI gives a weak "real-life example"? It has no criteria to discard it.

V4 · +performance — how you'll know what's good

› [V3 prompt in full] +
  CRITERION FOR A GOOD ASSIGNMENT:
  "A 12-year-old child can rephrase the assignment's condition
   in their own words in 30 seconds." If an assignment requires
   explaining the concept for longer — it's weak, discard it.

  First, create 8 candidates, then discard them
  based on this criterion, return 5 final ones
  + show which 3 you discarded and why.

⠋ Thinking…
⠋ Generated 8 candidates.
⠋ Running each through the 30-second test…

DISCARDED 3:
  × "Caesar cipher using a loop" — requires explaining encryption,
     30 seconds isn't enough, it's a loaded concept for a 12-year-old
  × "Bubble sort" — the concept of "sorting with an algorithm"
     itself requires explanation, plus while inside for
  × "Prime numbers" — needs explaining primality, 30+ seconds

5 FINAL (passed 30-sec test):
1. TikTok Playlist [as in V3]
2. WhatsApp Wait [as in V3]
3. Push-up Counter [as in V3]
4. Countdown to End of Break [new — better than push-ups]
5. Group Chat Friend Check [new]

The main thing that happened: AI started to filter out weak options itself. You didn't have to argue with it about every task. The performance criterion did the work inside the AI that a human mentor would do with their eyes without it.

Now multiply that by Mars context. Zukhra's ~/zukhra/CLAUDE.md (from M1) says: "I'm a Python mentor for 11-13 year olds at Mars IT School, my kids are from Yunusobod, native Uzbek, Russian as a second language, instructions and examples in two languages, 70% have basic Scratch". This file is loaded automatically. V4 with Mars context no longer just gives 5 tasks — it gives them bilingually, with the note "these 2 assume the child has seen Scratch blocks — for the 30% who haven't, explain in 2 minutes." What would have taken a V5 iteration without CLAUDE.md comes immediately with CLAUDE.md.

Edge of knowledge — where AI honestly doesn't know

Zuhra tries: "explain recursion using an Uzbek folk tale — for 12-year-olds". The AI cheerfully responds with "like Ivan Tsarevich..." — Uzbek tales are rare/local in its training. If not stopped, the material will go to children with Russian names.

Three places to "put" your context:

Exercise — 3 probes and a template-snippet.Uchta probe va shablon

Adaptation of "Bad Prompt Makeover" (AI Fluency lesson 8) + "Outsider Test" (AI Capabilities lesson 6). Do all three probes manually — without this, your next prompt will be a guessing game like everyone else's.

Quiz — check yourself in 3 minutes.O'zingni sina — 3 daqiqada

Three apply-questions. If you made a mistake in at least one, go back to the corresponding section. Not a grade, but an insurance before the next prompt.

What's next — Module 4.Keyingisi — Modul 4