Serge Shima / Work and projects
ZPL-80
Zip Prompt Language - compress heavy system prompts by ≥ 80 % token reduction
From the project documentation
ZPL-80 — Zip Prompt Language v1.1.2
Compact Query Language for Token Efficiency
Why ZPL-80 Exists
Large prompts burn tokens, time, and cash.
ZPL-80 compresses instructions by \~80 % while staying readable to any modern LLM.
Version 1.1 keeps the good parts of v1.0, drops the baggage, and builds in flexible CoT, format flags, and model wrappers.
Core Design Rules
Rule | What it means |
------------------------ | --------------------------------------------------------------------------- |
**Zero dead tokens** | Every character must add meaning for the model |
**Atomic blocks** | Prompt = sequence of self-describing blocks; omit what you don't need |
**Short, stable labels** | `CTX`, `Q`, `A`, `Fmt`, `Thought`, etc. One- or two-word labels only |
**System first** | Global rules live in the API's system role (or `[INST]…` wrapper for Llama) |
**Model aware** | Add the wrapper tokens the target model expects — nothing more |
**Optional CoT** | Fire chain-of-thought only for hard tasks via a single 🧠 trigger |
**Token caps** | Limit verbose sections with inline guards: `Thought(TH[INST] … [/INST]` around the whole prompt. |
**Plain completions** | No wrapper, just the ZPL text. |
Compliance Checklist
* [ ] Only needed blocks present
* [ ] No decorative lines or duplicated labels
* [ ] All macros defined before first use
* [ ] Symbols are single code-points
* [ ] Format / language flags set if required
* [ ] CoT token cap added for heavy tasks
* [ ] Wrapper matches target model
Roadmap
* v1.2 will add an auto-compression helper script and built-in diff-mode for interactive editors (Cursor, WindSurf)
* Community feedback welcome — open an issue or PR with failing edge cases
About the author
Serge Shima is co-founder and CTO of Majento in Tashkent and founder of Aimasters. He builds AI products and open-source tools for agents.