In a niche but intriguing intersection of high modernist literature and prompt engineering, developer herve76 has released 'finnegans-wake-skill,' a Claude skill designed to optimize cause-and-effect reasoning. Posted to Hacker News on November 20, 2026, the tool claims to reduce output token usage by up to 37% by borrowing specific compression devices from James Joyce's notoriously dense novel, Finnegans Wake. The skill targets questions like why did this happen? or does X cause Y?, forcing the model to deliver verdicts, rival explanations, and next steps in five short lines or fewer.
Literary Compression as Token Optimization
The core hypothesis of the skill is that Joyceβs structural techniques can be mapped directly to LLM output constraints. The repository details a direct correspondence between Wake devices and prompt rules: the 'riverrun' concept (where the last sentence runs into the first) is translated to placing the verdict in the first line; 'sigla' (Joyce's one-glyph codes) becomes a rule to tag facts once and never restate evidence; and the 'Shem and Shaun' duality is used to limit rival explanations to the single strongest contender and the one fact that defeats it. This approach aims to eliminate the verbose preamble that often bloats Claude's responses to analytical queries.
Benchmarking Against Baseline Prompts
The developer tested the skill using Claude Opus 5.5 at high effort across three datasets, including a blind holdout set written by an agent unaware of the skill's existence. Results showed a 37% reduction in output tokens on the blind holdout, 28% on the validation set, and 34% on fresh generated cases, with 100% correct verdicts in all scenarios. However, the source material admits a crucial nuance: a simple prompt instruction to 'Be concise and lead with the answer' achieved a 26% reduction in a later run. This suggests that while the Wake-inspired structure is superior, the majority of the efficiency gain stems from the basic directive of brevity rather than the literary mapping itself.
Input Costs and Practical Limitations
The skill is not a free lunch; it shifts the token burden from output to input. The system prompt adds approximately 400 tokens per call, and loading it into Claude Code incurs an additional 720 tokens per session. The developer notes that the skill only helps when analysis makes up a large share of the output and is ineffective for short replies, JSON extraction, or agent loops dominated by tool calls. For API users, the skill must be cached to offset the input cost, or it will only pay off on exceptionally long answers where the output savings exceed the fixed input overhead.
Key Takeaways
- The skill reduces output tokens by 28-37% across various test sets while maintaining 100% verdict accuracy.
- A simple 'be concise' prompt achieves 26% savings, indicating most gains come from brevity instructions rather than complex literary rules.
- Input tokens increase by ~400 per call, requiring caching or long outputs to break even on cost.
- The tool is specifically optimized for messy, multi-source cause-and-effect analysis, not general chat or data extraction.
The Bottom Line
This is a clever proof-of-concept that proves structure matters, but the data suggests you might get most of the benefit from just telling Claude to shut up and start with the answer. Itβs a fun read for Joyce fans, but a marginal win for engineers watching token bills.