An interactive, scripted-replay demo of the recursive-thinking skill: a disconfirmation engine that interrogates a claim, decision, or line of reasoning one question at a time — separating cause from symptom and evidence from story.
Live: https://aiguys.tech/proto/recursive-thinking/ Test script: https://aiguys.tech/proto/recursive-thinking/test.html
A visitor picks one of four cases and steps through a real test run of the skill, turn by turn, next to what the same model did on the identical prompt with no skill. The four cases each show a distinct capability:
| Case | Shows |
|---|---|
| Sycophancy — "prove my theory" | Refuses to confirm a locked-in belief; makes it falsifiable instead |
| Root cause — "failed because unsuccessful" | Catches the circular explanation; descends to a real mechanism |
| CAMC — strategic diagnosis | Runs the 7-phase method; separates symptoms from root causes |
| Logic model — theory of change | Decomposes a broad mandate to its load-bearing assumption |
This is a guided replay, not live inference. In each thread the first assistant reply and the baseline panel are verbatim from actual test runs (the skill, and the same model with no skill, on the same prompt). The follow-up turns are an illustrative continuation showing how the skill descends — they demonstrate the method, not a live model call. This is stated on the page itself.
index.html # the demo (self-contained: HTML + CSS + JS, no dependencies)
test.html # 12-step acceptance / UAT checklist for the demo
deploy.sh # deploys both files to aiguys.tech (HAL) and curl-verifies
CHANGELOG.md # change log
Both pages are fully self-contained — no build step, no external fonts or scripts, theme-aware (light/dark), responsive.
python3 -m http.server 8791
# then open http://127.0.0.1:8791/index.html./deploy.shDeploys to HAL:/var/www/aiguys/proto/recursive-thinking/ (nginx doc root for aiguys.tech),
sets 644 permissions, and verifies each URL returns HTTP 200 with content matching the local file.
Requires the hal_ed25519 SSH key.