Poor quality and too much cost? Exactly why i created Godspeed. (MIT) GIT 2 Prompt Claude Instal
github.com/itsribbZ/Godspeed
march ($3,179) + april ($3,151) = $6,331 across 9.37B tokens. most of april was dev time on opus 4.7 building the plugin itself.
what it does: scores every prompt on an S0–S5 complexity scale (keyword + regex classifier, ~5–15 ms, zero API call) and routes to the cheapest model that will handle it well. S0–S2 → haiku. S3 → sonnet. S4–S5 → opus. for S3+ it decomposes into parallel sonnet workers
instead of running monolithic opus — ~10× cheaper per task with quality parity (anthropic MARS pattern).
classifier scores 69% exact on a 200-prompt held-out eval, +31.5 percentage points over the best naive baseline. repro: `python toke/automations/brain/eval/brain_vs_baselines.py --json out.json`.
what i actually use it for day to day:
• deep research / planning without burning main-session context (parallel sonnet workers do the search, opus synthesizes)
• long agentic sessions where the main session would otherwise compact (1M-token window support, vector-embedded memory for cross-session recall)
• file cleanup that won't blow up — sacred-rule gate requires explicit approval before any delete
• learning rabbit holes — every synthesis cites its sources so i can verify claims
honest tradeoff: classifier misses ~20% of S3+ edge cases. when it's uncertain it emits a stderr advisory and lets you override with /model manually. it's a cost-aware router, not a magic oracle — no guardrail can replace reading the output.
install (2 commands, in any claude code session):
/plugin marketplace add itsribbZ/Godspeed
/plugin install godspeed@itsribbZ-godspeed
once installed, `/godspeed:godspeed-info` renders the pipeline diagram and `/godspeed:godspeed-settings` dumps the routing manifest so you can see every tunable knob. MIT, no telemetry, no phone-home.
[留言]
为什么值得关注
能改变理解方式,而不只是重复常识;符合当前抓取需求;它提供了新的理解或解释,而不只是表面观点
来源:reddit,领域:tech,保留分:0.62
讨论总结
讨论量较低,暂无明显增量信息。