Your AI means well. Sometimes too well.
Tell it to "check the whole project" and it obediently reads everything in sight — burning tokens like a bonfire, most of them on code that has nothing to do with your problem.
Token PRO50 sits between your intent and your AI's enthusiasm. Before a single file is read, it runs a silent Token Risk Assessment. If an instruction drifts into wasteful territory, it pauses — asks exactly one sharp question — then steps aside.
We used to ship only the interception half. Honest truth: in real use it almost never spoke up. You're a careful user — you don't say "read the whole project" — so the guard stayed silent, and the daily waste went unnoticed.
v2 fixes that by doing the quiet work too. It now weighs every request, not just the reckless ones.
It doesn't make your AI smarter. It makes your AI respect your tokens — and your time.
Token PRO50 is a single Markdown skill file (no scripts, no dashboards) that teaches your agent restraint through two modes:
Mode A · Soft Optimization (every request). A lightweight scan catches the waste you don't notice:
- A verbose preamble → rewritten tighter (5C style), shown side-by-side so you can ignore it.
- A pasted wall of text → key points pulled out before it hits context.
- A chat that's run 20 rounds → compressed to "done / not done / next step."
- A trivial task ("translate this sentence") → nudged to a cheaper model or lower effort.
Mode B · Hard Interception (only the reckless). Four red lines your AI may not cross without checking with you first: global scan, ultra-long unfiltered read, unbounded tools, vague goal with no acceptance criteria. When one trips, it asks exactly one question and offers four escape hatches.
The default suggestion is A, because every senior engineer knows: you don't grep the whole repo on day one.
We all know the feeling. You paste a 6,000-character chat and say "summarize this" — and the model dutifully ingests all 6,000 words when a handful of points would do. You write a three-sentence task wrapped in 200 words of "hope this helps." None of it is reckless. All of it adds up.
v1 only caught recklessness. v2 also catches the drip-feed — and it never touches anything that could change your answer. If a compression might lose meaning, it asks first.
Soft Optimization only moves redundancy — never substance. Important deliverables (quotes, code, contract summaries) still get verified. In measured samples, prompt rewrites came out ~25% lighter and condensing a pasted wall or a long chat dropped 80%+ of the tokens. Your mileage depends on how much you paste; it is not a promise, it's a measured tendency.
- Read only what the task needs; skim a long doc's structure first.
- Use only the tools the current step needs; bound MCP/skills instead of disabling them.
- Simple tasks stay lightweight; decisions, risks, and final deliverables never skimp on review.
- Saving tokens never means skipping verification.
- For long chats: compress, don't pile on.
Drop this repo into your agent runtime's skills directory (the folder it scans for installed skills):
Windows:
<skills-dir>\token PRO50\
macOS / Linux:
<skills-dir>/token PRO50/
The folder needs SKILL.md (English, loadable) and SKILL.zh.md (Chinese source).
Works with any agent runtime that loads SKILL.md-formatted skills — Claude Code, Codex, Cursor, and similar.
SKILL.md— English (the loadable skill).SKILL.zh.md— Chinese source of the same skill.