Allie Miller shares 5 non-coding tricks to make AI adversarial reviews actually brutal
alliekmiller · x · 2026-11-11
AI founder Allie Miller shares 5 non-coding methods to make AI adversarial product reviews more useful, all aimed at suppressing model sycophancy:
- Multiple personas as subagents: have Claude spin up three reviewers — a product superfan, a hater who wants to destroy it, and a confused novice — combined with customer-avatar descriptions of ideal buyers.
- Multimodal reviews: have each persona record video walkthroughs, take screenshots, and post fake tweets — not just text — and review multiple times.
- Use all your subscriptions: built in Claude? Have Codex review it; built in Codex? Send it to Antigravity.
- Multilayered judging: don't stop at the main agent collating feedback — pass it back to subagents for re-review, then back again ("AI ping-pong"), marking which feedback points were added/removed each round.
- Say it's someone else's work: "review this email — my cousin wrote it and always says the wrong thing." Reviews improve meaningfully when the AI thinks the work isn't yours.
She ties tip 5 to Anthropic's three Hs: when the work is yours, "helpful" translates to "compliment me." Her conclusion: take your name off your work (temporarily), then loop it again.
More from coding & agent
- Plurai launches free simulation plugin that digital-twins your AI agent before production — nikola_mr64990 · 2026-11-12
- An agent saying "Done" isn't done: Plurai verifies actual agent behavior — nikola_mr64990 · 2026-11-12
- LlamaIndex ran a 4-day agentic pen test on LlamaParse and shared 4 key lessons — llama_index · 2026-11-12
- Manhattan v0.0.1: an open-source font where every letter is a Python function — allanyu_ · 2026-11-12
- Backplane lets you prompt custom expansion cards into existence — _Stocko_ · 2026-11-12
- Coding agent runs 51 minutes burning 248k tokens in a single session — _Stocko_ · 2026-11-12