§3.3 · Chat
Project Instructions — compress without losing intent
A 1,500-token instruction block can usually do the work of a 5,000-token one.
The Project Instructions on a Claude.ai Project (or the system prompt on an API call) are read on every turn. Bloat there compounds — every wasted word is a tax on every future message.
Most production instruction blocks compress to 30–60% of their original size without losing instructional fidelity. Here's the playbook.
Step 1 — strip the politeness
"You are a helpful assistant. Please be kind and thoughtful. Always try to give your best answer."
Cut. Claude is already helpful. Stating it adds tokens and accomplishes nothing.
Step 2 — collapse repetitions
Look for the same idea stated three different ways. "Be concise. Don't be verbose. Avoid unnecessary words." → "Be concise."
Step 3 — replace prose with structure
Long prose:
> When the user asks about pricing, you should respond with the current pricing tiers, but make sure to include the start date for each tier, the monthly cost, and the included features, and don't include any historical pricing that's no longer offered.
Compressed:
On pricing questions: list current tiers only.
For each tier include: start date, monthly cost, included features.
Don't include retired tiers.
Same instructions. ~50% fewer tokens.
Step 4 — extract reusable blocks
If you have a tone description that applies to multiple projects, save it as a Context Library @slug block and reference it by name instead of restating it in every project.
Step 5 — run the Prompt Critic
The Prompt Critic tool flags hedges ("perhaps", "kind of"), wordy phrases ("in order to" → "to"), and vague qualifiers ("very", "really") that compress without affecting Claude's behavior.
Step 6 — measure
Compare the old block vs the new block on a representative test set of prompts. If outputs are equivalent, ship the compressed version.
A real example
| Version | Tokens | Cost per 1k messages |
|---|---|---|
| Original | 4,200 | $12.60 |
| Compressed | 1,800 | $5.40 |
| Compressed + cached | 1,800 (cached read) | $0.54 |
Compression alone cuts your bill in half. Adding caching cuts it by ~95%.
The Project Instruction Compressor module does the compression pass for you and previews the diff.