I built Perspectives because I got tired of ChatGPT agreeing with everything I said.

Ask any LLM to "consider multiple perspectives" and you get hedged consensus. The model acknowledges trade-offs exist, then settles on a moderate position that offends nobody. Useful for summaries. Useless for decision making.

Perspectives forces disagreement. 8 personas with fundamentally incompatible frameworks debate your question through a structured protocol, then vote using Single Transferable Vote to surface where they actually land. The output is a PDF report synthesising all of it.

How it works

Blind Proposals: Each persona generates a position without seeing the others. This prevents the "anchoring problem" where early responses shape later ones, bypassing the default sycophancy of LLMs.

Interrogation of Blind Proposals: Proposals face structured challenges from 3 opposing personas. A "high-empathy" persona (e.g., The Idealist) will be challenged by a "low-empathy" cluster (e.g., The Pragmatist). This reveals exactly where arguments buckle under pressure.

Discussion & Voting: Personas can debate (optional) before ranking preferences via STV. This highlights first-choice winners and preference flows rather than simple majority rule.

Analysis/Prediction Report: The final PDF structures recommendations first, followed by supporting analysis (factual background, risk assessment, evidence quality).

Two Operational Modes

Analysis Mode ("What should we do?"): Evaluates options and surfaces trade-offs. Output is qualitative judgment.

Prediction Mode ("What will happen?"): Generates probability estimates with resolution criteria.

Feedback Loops

Most AI agent projects have no way to measure whether their outputs are actually good. Users provide subjective feedback, which is noisy and unreliable. The system optimises for seeming useful rather than being useful.

Prediction Mode creates an objective feedback loop. When a prediction resolves, I can measure accuracy.

I'm integrating Polymarket as the verification source. Run a question through Perspectives, record the predictions, compare against actual outcomes when they resolve. Over time, this builds calibration data showing which methodologies perform best for different question types.

Persona Sets

Different decisions need different analytical lenses. Four built-in sets:

Philosophical (Default): Best for ethical dilemmas and strategic decisions.

Business-Focused: Best for commercial decisions.

Product-Focused: Best for product development.

Forecaster: Optimised for Prediction Mode.

Technical Details

LLM Support: Supports any OpenAI/Anthropic compatible API (Claude, OpenRouter, Ollama, Grok, etc.).

Web Search: Optional integration for grounding debates in recent events.

Output: Single PDF report per query.

What I'm Looking For

I've been building this solo and could use external feedback on a few things:

1. Does the blind proposal mechanism actually produce better disagreement?

2. Is the interrogation protocol overkill or useful? The structured challenge/response/verdict cycle generates rich data, but adds latency (dependant on concurrency settings).

3. What decisions would you run through this?

4. Do you use ChatGPT or similar systems to make decisions?

5. Do you find "chain of thought" output useful for tracking reasoning?

Links

Perspectives: https://getperspectives.app

Dev blog: https://blog.jmatthews.uk

Example Analysis Report (Is it viable to run a nation where all laws expire after 10 years and must be re-passed?): https://drive.google.com/file/d/1hsJOWsQDAtVOqOKF6_a_Q1jYOlB...

Example Prediction Report (Will Kraken IPO by 31st March 2026?): https://drive.google.com/file/d/1m3RedFtv8lKgFqf1_rvzl8W6cTs...

Happy to answer any questions in this thread.

Reflections on Re-Reading Dostoevsky's the Idiot (1976)

How to Steal Any React Component

Show HN: Perspectives – I wanted AI to challenge my thinking, not validate it

Sometimes Dropbox is just FTP: building a link shortener

Drowning in information but still missing everything?

AI boom could falter without wider adoption, Microsoft chief Satya Nadella warns

I Gave Claude Code 9.5 Years of Health Data to Help Manage My Thyroid Disease

Will Google Become Our AI-Powered Central Planner?

More than 200k UK workers switch to four-day week since 2019 (2025)

UK holds off joining Trump's Board of Peace over Putin concerns

I replaced my ChatGPT subscription with a 12GB GPU

More sovereign US bases? I dont think so

Google Antigravity is currently down for many users

The necessary post-Trump consensus

AliSQL – MySQL with DuckDB storage engine from Alibaba

The stable marriage problem. Or, ask for things

Prototypes vs. classes was: Re: Sun's HotSpot (1998)

Further scarcity: World leaders allegedly reduce NAND flash production

Apps for boycotting American products surge to the top of the Danish App Store

Garbage Collection Can Be Faster Than Stack Allocation (1986) [pdf]

From Human Ergonomics to Agent Ergonomics

Trump Proposes "Great Healthcare Plan"

Show HN: BetterTTS – Offline Lifelike TTS Web Reader for Windows

Show HN: Crackable- a Cryptography CTF Platform

The importance of humidity control in textile processing

Sources: The SGLang project becomes RadixArk with a valuation of US$400M

Fix coverage documentation to clarify when data is saved

OpenAI aims to ship its first device in 2026, and it could be earbuds

Researchers Identify Urban Atmosphere as Primary Reservoir of Microplastics

OCIO, Display Transforms and Misconceptions

Show HN: A Free Online Podcast Transcription Tool