call_human()

ai tutor · no code output

the AI that makes you do it yourself.

paste a coding problem. it hands you one step at a time and never writes the code. your exam doesn't have copilot.

one-time. no subscription. opens this week.

~/two_sum.pystep 2 / 4

> claude.solve(problem)

> plan() → 4 steps [hidden]

✓ call_human(step=1) → pass

> call_human(task="write the loop that checks each pair of numbers")

░ waiting for human... 00:00

how it works

the AI is the orchestrator. you're the tool it calls.

  1. 01 > paste(problem)

    paste the problem

    your assignment, a leetcode, last year's midterm. python, java, javascript or c.

  2. 02 > plan() → hidden

    it plans, you don't see the plan

    claude breaks it into 3 to 7 steps and keeps them to itself. you only ever see the next one.

  3. 03 > call_human(task="…")

    it calls you

    each step is a function call where you're the function. you write the code. it waits.

  4. 04 > grade() → line 3

    it points, it never writes

    wrong? it tells you which line and what to think about. hints go concept → location → plain english. never code.

example session

it tells you where you're wrong. it won't tell you what to type.

~/total.py · your codestep 2 / 3
1def total_of(nums):
2 for n in nums:
3 total = 0
4 total += n
5 return total

> call_human(task="add up every number in the list")

< returned 5 lines

> grade(step=2)

↺ retry · line 3

close. total = 0 sits inside the loop, so it resets on every pass. where should it live so it survives?

> hint(level=3)

1. start a running total before you look at any number. 2. add each number to it. 3. give it back once every number is counted.

quotes only your own code. points at line numbers. every response is scanned for code you didn't write, and thrown out if it has any.

why not just use a chatbot

every AI tutor can be talked into writing your code. this one can't.

the no-answer rule

chatbots: a prompt asking the model to hold back. a clever message can talk it out of it.

call_human(): enforced on the server. every reply is checked, and anything with code you didn't write is thrown out.

the plan

chatbots: shows you the whole approach up front.

call_human(): hidden. you get one step at a time.

when you're wrong

chatbots: rewrites your line "for clarity".

call_human(): points at the line number and asks the question.

proof you did it

chatbots: none.

call_human(): a receipt with every step, your paste count, and a keystroke replay.

the receipt

finish a problem, get proof a human solved it.

every keystroke is recorded while you work. when you finish, you get a public receipt: time taken, steps, hints used, how many times you pasted, and a replay anyone can scrub through.

send it to a TA, pin it on your github, drop it in the group chat before the midterm. it's the one thing an AI can't fake for you.

you choose whether the replay is public. stats-only is one click.

receipt · examplecall_human()

two sum

time
23:41
steps
4 / 4
hints
2
pasted
0 times
▶ replay 4x14:40 / 23:41

solved by a human.

founding access

$5, once. that's it.

  • ✓unlimited sessions when the app opens this week
  • ✓every language: python, java, javascript, c
  • ✓public receipts with keystroke replay
  • ✓one payment, no subscription
  • ·free tier: 3 sessions a day, for everyone, forever

$5

one-time founding price. tax added at checkout where it applies.

$5 founding access

faq

questions

will it ever show me code?

no. every reply is scanned before you see it. if it contains code you didn't write, it's thrown away and the tutor has to try again in plain english. it can quote your own code back to you, and that's it.

isn't an AI tutor just a nicer way to cheat?

most are. this one is built the other way around: it can't write your code, and the receipt shows exactly how much you typed versus pasted.

does it run my code?

no. it reads it the way a TA would and tells you what's off. run it yourself.

which languages?

python, java, javascript and c.

what if i paste something in?

you can. pastes over 40 characters are logged and counted on your receipt. nothing is hidden from you or from whoever reads it.

what do i get for $5?

unlimited sessions once the app opens this week, for a single payment. the free tier stays at 3 sessions a day.

who's building this?

a CS student at Dalhousie in Halifax, in public, this week.

> call_human(task="get good before the midterm")

your exam doesn't have copilot. practice like it.

$5 founding access