最近面试AI Engineer,面试官问我怎么看 Prompt Engineering。
我直言:它没有大家想象中那么重要。
说完已经做好不被录取的准备。😂
后来看到 Claude Code 作者 Boris Cherny 的观点:
很多人习惯一步步命令AI“先做1、再做2、再做3”,但这不是现代模型的最佳使用方式。
更好的思路:
描述任务 + 设定约束 + 定义完成标准。
剩下交给模型自己发挥。
“Then just go let the model cook.”
🚨 Claude Code is eating your context window alive.
Most people try to save tokens by writing shorter prompts.
That's not where the real waste is.
Your terminal output, repo context, MCP tools, logs, and Claude's own verbosity can burn 10-50x more tokens than necessary.
I found 10 open-source tools built specifically to fix this:
1. Caveman Claude
Makes Claude communicate in brutally compressed language.
Less explanation.
Less repetition.
Same task completion.
Claims up to 75% fewer output tokens.
https://t.co/jtkpyRMcRy
2. RTK — Rust Token Killer
Your agent runs:
git diff
cargo test
npm test
docker logs
…and dumps thousands of useless lines into context.
RTK intercepts that output and strips the noise before Claude ever sees it.
Reported savings: 60-90%.
https://t.co/JTh4cWi2sL
3. Code Review Graph
Instead of throwing your entire repo at Claude...
it builds a Tree-sitter dependency graph and retrieves only the code relevant to the change.
On large repos, the difference can be ridiculous.
Claims up to 49x less context.
https://t.co/EpLyHsY2B9
4. Context Mode
One of the most interesting ones.
Instead of dumping huge tool responses directly into Claude's context...
it stores the raw data outside the conversation and lets the model query only what it needs.
Logs.
GitHub output.
Search results.
Large MCP responses.
Claims up to 98% context reduction on some workloads.
https://t.co/cloMBeLy7B
5. Claude Token Optimizer
Your CLAUDE.md can quietly become thousands of tokens of instructions Claude rereads constantly.
This tool compresses project instructions without throwing away the useful parts.
Example claimed reduction:
11K tokens → 1.3K
https://t.co/K634w1Enmx
6. Token Optimizer
Think of this as a context leak detector.
It looks for the hidden stuff slowly destroying your usable context:
→ redundant instructions
→ repeated tool output
→ unnecessary state
→ oversized context
Then helps clean it up.
https://t.co/zc9q1oCe78
7. Token Optimizer MCP
MCP servers are insanely useful.
They're also capable of dumping enormous JSON responses into your context.
This adds caching + compression between your MCP tools and Claude.
Claims 95%+ token reduction in some workflows.
https://t.co/P4Ctrdefbg
8. Claude Context
Built by Zilliz.
Instead of stuffing the whole codebase into context, it gives Claude hybrid/vector search over your repository.
Claude searches for relevant code.
Then reads only that.
Much better architecture than:
"here are 70 files, good luck."
https://t.co/qwiebGgR7r
9. Claude Token Efficient
Probably the easiest one here.
Drop a CLAUDE.md into your project.
It instructs Claude to:
→ stop repeating itself
→ avoid giant explanations
→ minimize unnecessary output
→ use tools more efficiently
No proxy.
No infrastructure.
https://t.co/AXoi8VM41I
10. Token Savior
Instead of navigating code by opening giant files...
it navigates through symbols.
Functions.
Classes.
References.
Definitions.
So Claude fetches the exact piece of code it needs instead of another 2,000-line file.
Claims up to 97% less context for navigation.
https://t.co/iNTMSbLcBW
The stack I'd actually use:
Huge monorepo?
Code Review Graph + Token Savior
Terminal output destroying context?
RTK
MCP / GitHub / logs dumping thousands of tokens?
Context Mode
Claude talks way too much?
Caveman + Claude Token Efficient
Codebase search sucks?
Claude Context
But here's the bigger lesson:
The future of AI coding isn't just better models.
It's context engineering.
The best developers won't dump 200K tokens into every agent and hope intelligence fixes it.
They'll give the model exactly what it needs.
Nothing more.
Smaller context.
Cheaper models.
Better routing.
More work shipped per dollar.
Before buying more tokens, run /context.
You might discover you never needed most of them.
More in this article
when using GLM-5.2 and GLM-5.3, use zcode rather than claude code
with zcode, you get to know numbers that no other IDE coding would ever show you:
it is your cache hit rate in real-time, which tells whether you're using your quota as expected or burning it up
also, it shows you the breakdown of your context tokens into messages, tools, skills, and system prompt, which remains a black box in claude code
the zcode also gets you an extra 150% quota on off-peak timings from https://t.co/hlsJYStOds. just stay away from 14:00 to 19:00 Beijing time, which costs you double
it's is a more suitable harness for any https://t.co/hlsJYStOds model.