The community responded with an explosion of tools. Here are all the optimization techniques I've found. Ranked from C-Tier (meme) to S-Tier (serious engineering).
⭐ C-Tier: Use /compact when you're at 60% of the conversation's max context. 60% is the sweet spot. Context tokens aren't all equal, and when you're near the limit, agents start drifting and follow instructions less reliably. Why meme-tier? Because nobody's going to do it manually. (It's done automatically by the S-Tier tool.)
⭐⭐ B-Tier: Compress prompts before they reach the model. In any prompt, roughly half the words carry no real information, and agents are perfectly capable of filling in the blanks. That's why there's a whole category of specialized models for prompt compression with minimal quality loss: LLMLingua or Bear-2, for example. Why B-Tier? It's only really suitable for deterministic workflows (like n8n or similar), not for highly variable agentic environments. But it's excellent in those specific cases.
⭐⭐ B-Tier: Always tell Claude to launch its exploration sub-agents with Haiku instead of Fable, Opus, or Sonnet. This can be up to 92% cheaper on exploratory searches. Haiku is great at digging, not at thinking.
⭐⭐⭐ A-Tier: /effort default to xhigh. All the time. But 80% of what you ask Claude doesn't actually need that level of token waste. Switch to medium or high for most tasks — it cuts the amount of thinking required by 2x or 4x to reach the result.
⭐⭐⭐⭐ S-Tier: Agent harness compaction tool. Why S-Tier? It's plug-and-play and the gains are immediate, no matter what you're doing with your agent. Tokenade, for example, automatically optimizes: the commands the model uses, file re-reads, code/web search, MCPs, the order of tool calls, etc. It currently ranks #1 on THOL (Token-Harness Optimizer Leaderboard), with -27% token cost on 200k+ token conversations and -40% on 400k+ token conversations (Opus & Fable).
Do you use any of these techniques ?
Tokenade : https://tokenade.net/
THOL : https://pi-infected.github.io/token-harness-optimizer-leaderboard/
LLMLingua : https://huggingface.co/spaces/microsoft/llmlingua-2
Bear-2 : https://thetokencompany.com/
Top comments (0)