prompt-riser — a reasoning debugger that shows you how AI models think

The Problem

I’ve been working with reasoning models a lot recently — GPT-5.6 Sol, Claude Fable, DeepSeek R1, Kimi K3. They’re incredible at complex tasks, but there’s a fundamental problem: you only see the final answer, not the thinking.

MIT/Stanford published a paper this month showing that self-correction during reasoning matters more than chain length. Models that catch and fix their own errors outperform longer-but-uncorrected chains by a wide margin. But how do you know if your model is self-correcting or just powering through confidently wrong?

There was no tool to answer that question. So I built one.

Enter `prompt-riser`

prompt-riser is a CLI reasoning debugger. You give it a prompt, and it runs that prompt through 5 different reasoning strategies against any LLM, then shows you the thinking paths side-by-side.

The strategies

Strategy	What it does
Chain-of-Thought	Numbered step-by-step reasoning
Self-Correcting	Generate → Critique → Improve, multiple passes
Step-Back	Derive broader principle, then drill into specifics
Constraint-Based	Identify constraints, reason within them
Tree-of-Thoughts	Explore multiple branches, evaluate, pick best

How it works

prompt-riser run "What's the best way to deploy on a budget?"

The tool dispatches the prompt through each strategy as a separate LLM call (or sequence of calls for self-correcting), captures the full reasoning trace from each, and renders them as a visual tree:

┌─ Self-Correcting ────────────────────────┐
│  Pass 1: Use ECS Fargate...             │
│  Critique: Too expensive for budget      │
│  Pass 2: Lambda + EFS...                │
│  Critique: Cold starts might be issue   │
│  Pass 3: Lambda + provisioned concurrency│
│  ✅ Final answer                         │
└──────────────────────────────────────────┘

You can output as a terminal tree (Rich), JSON, or Mermaid flowchart.

The MIT/Stanford connection

The self-correcting strategy directly implements the paper’s finding — it does multiple generate→critique→improve passes. The tool doesn’t just show you the final answer; it shows you where the model corrected itself, which is the signal that matters most.

Tech stack

Python (click, rich, httpx, pyyaml)
OpenAI-compatible API (works with any provider)
AST-based strategy definitions — easy to add new strategies
14 tests, all passing

Try it

pip install -e .
prompt-riser run "Your prompt here" --output tree

GitHub →