Christian Arango
Side projects

Side project · AI agent

An AI agent for an initial IQ interview

I wanted to see how far a conversation can go as a first cognitive interview. This terminal agent runs an adaptive evaluation with Claude across reasoning, verbal, quantitative, and memory tasks. It gives an experimental estimate, never a clinical score.

Author
Christian Arango
Role
Product Owner and builder, with Claude Code
Code
chericito/ia-agent-cheri
Status
In development · October 2026
A real first turn: limits stated up front, then item 1. The scoring notes stay hidden.

The problem

A chat model wants to please you; an evaluator can’t

Left alone, a model flatters, hints at answers, and leans on what it already knows about you. The evaluator has to stay neutral, adjust difficulty honestly, and keep its scoring out of sight.

What I built

A neutral evaluator in the terminal

  • Prompt

    A neutral evaluator

    The system prompt rules out flattery, never says which ability an exercise tests, and keeps knowledge apart from reasoning.

  • Hidden notes

    Scoring you never see

    The evaluator writes its notes in tags the script strips out, so they persist between sittings without appearing on screen.

  • Memory

    Stimuli that disappear

    Memory tasks show for a few seconds, then the screen and the scrollback are wiped. Each answer silently carries its response time.

See how a memory stimulus is shown and wiped
agente.ps1powershell
# <registro> is the evaluator's hidden scoring log: kept in the session, never shown.
$text = [regex]::Replace($text, '(?s)<registro>.*?</registro>', '').Trim()
$m = [regex]::Match($text, '(?s)<memoria(?:\s+segundos="(\d+)")?\s*>(.*?)</memoria>')
# ...the stimulus shows for N seconds, then:
# Wipe the screen and the scrollback so the stimulus can't be scrolled back to.
Write-Host -NoNewline "$esc[3J"
Clear-Host
Hidden notes are stripped, and the stimulus can’t be scrolled back to.

Decisions

Run it on a subscription, and keep it isolated

  • No API key, no per-use cost

    Each answer runs through Claude Code on the person’s Claude subscription and resumes the same session.

  • Nothing else leaks in

    Safe mode with no tools keeps other projects, memories, and settings out of the evaluation.

  • Pause anytime

    Progress saves after every answer, and separate sessions let someone else take their own evaluation.

Where it stands

It works in the terminal; the result is an estimate

The full flow runs in Windows PowerShell, with a step-by-step guide in English and Spanish. The evaluation itself is in Spanish, and its result is framed as a conversational estimate, not a diagnosis.