Months of token-efficiency benchmarking: ripgrep plus project maps double your tokens for free
EagleApprehensive · reddit · 2026-10-02
After months of deep-diving into token-saving techniques — benchmarking and auditing nearly every solution posted on Reddit/X — the author built tools that beat them all on their own benchmarks, released as intentic.dev.
Approach: assume you can't control the model, only everything around it: environment (OS, packages, filesystem, memory, ports, CPU/GPU), the human (workflow, cold-cache prevention, resuming, scheduling, context passing across accounts/models), inner harness (Claude Code/Codex natives are already good; changing them was mostly net negative), and outer harness (pre-tool hooks, system prompt).
Key finding: each DIY mechanism takes at least a day to get right for under 1% token savings. Only two easy wins are worth it for everyone:
- ripgrep, with a few words in AGENTS.md instructing its use
- project maps appended to the system prompt
Everything else isn't worth the effort. The full set ships as a Docker sandbox you can verify yourself.
More from coding & agent
- Google's Cogentic uses multi-agent orchestration to produce novel results on five open math problems — KyeGomezB · 2026-10-02
- AI Hedge Fund open-source project adds paper trading: build, backtest, deploy — virattt · 2026-10-02
- Vibecoding Fails Because We Sell Components as Products, Not Because Models Are Bad — wonderwomancode · 2026-10-02
- Circle brings USDC and Agent Wallets to Muse with one-line prompt setup — 4KTV · 2026-10-02
- Durable agents are heating up, says agent pioneer Yohei Nakajima — ritakozlov · 2026-10-02
- $100/month showdown: ChatGPT Pro vs Claude Max 5x for heavy coding and technical work — St-CB · 2026-10-02