Detailed Analysis
A developer known as SijuEC has released an open-source "skill file" for Claude that adapts the speech patterns of Rocky — the alien engineer character from Andy Weir's science fiction novel *Project Hail Mary* — into a structured prompting system capable of reducing Claude's output token count by up to 83%. The project, hosted at github.com/SijuEC/eridani-speak and installable via a single Claude Code command, emerged from a practical frustration: the developer was consistently hitting Claude Pro's usage limits during normal sessions. By encoding Rocky's distinctive grammar — which strips articles, drops copulas, uses hyphenated compounds, repeats words for emphasis, and expresses emotion as bare fact — into a reusable system instruction file, the developer compressed a 335-token response down to 56 tokens while retaining the substantive content of the answer. The project ships with two operational modes: "Rocky," which preserves the character's warm, dense personality and activates via `#rockyon`, and "Signal," a personality-free technical notation variant activated via `#signalon`, both togglable mid-conversation.
The token reduction achieved here is not merely cosmetic. Output tokens are the primary driver of both response latency and, in rate-limited subscription tiers like Claude Pro, session consumption. Because Claude Pro and similar subscription tiers cap usage based on token throughput rather than query count alone, any prompt engineering that systematically compresses outputs without degrading informational density directly extends the practical utility of a fixed-cost subscription. The developer's own documentation notes a key finding from the build process: explicit grammatical rules produce weaker stylistic adherence than worked examples, which is why the skill file relies heavily on demonstration rather than instruction. This insight aligns with established understanding of how large language models internalize style — few-shot exemplars outperform zero-shot directives for constraining output register.
The project fits into a small but growing category of community-driven prompt engineering that borrows compression heuristics from unexpected sources. Research context reveals that similar efforts, including "caveman speak" prompting on platforms like Hacker News, achieve roughly 70% token reduction but sacrifice personality. Rocky's grammar closes that gap — achieving comparable or superior compression while preserving a coherent communicative voice that users find preferable for extended chat sessions. The alien engineer's syntax, designed by Weir as a linguistic artifact of a species that communicates via pressure waves and has no concept of verbal waste, turns out to be a remarkably natural fit for token-efficient AI prompting: it is information-dense, context-reliant, and hostile to filler.
Broader adoption of this approach raises questions about the emergent tension between natural-language usability and computational efficiency in AI interfaces. Claude and similar models are trained to produce fluent, verbose, human-readable prose — a stylistic default that serves general audiences but penalizes power users operating under token constraints. The eridani-speak project demonstrates that Claude can be steered away from that default reliably through system-level instruction without degrading factual accuracy or reasoning quality. Independent implementations noted in the research — including a Japanese-language adaptation on Zenn and a multi-tier slash-command system in the Claurst CLI — suggest the pattern is generalizable across languages and deployment contexts, not just an English-language curiosity.
The release also points toward a maturing ecosystem of user-built Claude extensions that treat the model's behavior as configurable infrastructure rather than fixed product. The Claude Code one-liner install path (`npx skills add SijuEC/eridani-speak`) treats a personality and compression profile as a portable, version-controlled artifact, which reflects how developer-oriented Claude users are beginning to modularize prompt logic the way software engineers modularize library dependencies. As Claude's context windows grow and token pricing remains a constraint for high-volume users, community-developed compression layers like eridani-speak represent an increasingly practical class of tooling — one that sits at the intersection of literary inspiration, linguistic analysis, and systems-level thinking about AI resource consumption.
Read original article →