rlaope/ultraprompt

31 stars · Last commit 2026-09-15

Fable 5.1 solution tracing. Prompting skills so Opus, Sonnet, Haiku, Grok, Gemini and open-weight models work like Fable 5.1

README preview

# ultraprompt

Portable strategy skills for coding agents, to be distilled from the reasoning traces of a frontier coding model. (v0.2: baseline drafts, trace evidence pending — see [Status](#status) for what has actually landed.)

The pipeline: run a broad set of coding cases — frontend state machines, high-traffic backends, ML training loops, kernel optimization and quantization, agentic harnesses, low-level Linux — on **Claude Fable 5.1** via Claude Code; extract from the session transcripts how the model actually solves problems: what it explores first, how many hypotheses it keeps alive, what it accepts as evidence of "done", when it abandons an approach; distill those recurring strategies into English skill prompts that make a weaker agent — Claude Opus 5, Sonnet 5, Haiku 4.5, a GPT-class model, an open-weight model, anything that reads a system prompt — reason more like the stronger one. The output is not code. It is a set of carefully written prompts, to be grounded in trace evidence as case runs land.

## How it works

```
 (1) curate            (2) run              (3) extract           (4) distill          (5) publish
 ~16 case domains  →   Fable 5.1 via    →   full reasoning    →   recurring        →   skills/<axis>/
 frontend, servers,    Claude Code          trace: thinking       strategies along     SKILL.md
 ML, kernels,          sessions             blocks, tool-call     orthogonal axes      CASES.md
 agents, syscalls,                          sequences, self-      (8 core + 4 draft)
 storage, ...                               corrections
```

1. **Curate cases.** A catalog of ~16 domains (frontend, UI design, high-traffic servers, distributed systems, ML training, kernel/quantization, agentic harnesses, CLI/Linux, compilers, storage engines, networking, security, data engineering, testing/legacy, concurrency, games) with per-case difficulty and a note on which reasoning mode each case is designed to provoke. Design-type, debugging-type, and optimization-type cases within the same domain stress different strategies on purpose.
2. **Run on Fable 5.1.** Each case is executed as a real Claude Code session, not a one-shot completion, so the model plans, calls tools, hits failures, and recovers.
3. **Extract the trace.** From the session transcript we keep the full reasoning surface: thinking blocks, the exact tool-call sequence, dead ends, and self-corrections — not just the final diff.

View full repository on GitHub →