Balanced vs Aggressive vs Structured vs Caveman: Choosing a Prompt Optimisation Style
Last updated: Published 25 August 2026
Token Optimiser has four optimisation styles — Balanced, Aggressive, Caveman, and Structured. They don't run different algorithms; they run the exact same 17-step pipeline with different settings for how aggressively to compress and whether to force a structural rewrite.
In short: Balanced trims obvious filler while keeping prose natural (the free-tier default); Aggressive removes more — hedge phrases, reminder framing, filler sentences — while staying grammatical; Caveman compresses to the minimum viable representation, including single-line label-free forms; Structured reorganises a multi-part prompt into labeled Task/Requirements/ Constraints/Output sections, sometimes at a token-neutral or slightly higher cost, because clarity — not compression — is the point of choosing it.
Same prompt, four styles
Here's one real input prompt, run through all four styles with no model-specific target selected:
Original — 103 tokens
I would like you to please carefully write a professional follow-up email to a client. It's important that you remember to keep it concise and to the point. Please make sure the tone stays polite and friendly throughout. The email should mention that the project deadline has been moved to March 15th, and you should ask them to confirm they've received this update. Thank you so much for your help with this.
Balanced — 83 tokens (19% shorter)
The default, free-tier style. Conservative: trims filler and pleasantries but never drops a whole hedge sentence, and keeps prose readable.
Carefully write a professional follow-up email to a client. It's important that you remember to keep it concise and to the point. Make sure the tone stays polite and friendly throughout. The email should mention that the project deadline has been moved to March 15th, and you should ask them to confirm they've received this update.
What changed: the opening pleasantry and closing thank-you were removed. The hedge sentence (“It's important that you remember to…”) survives whole — Balanced is explicitly excluded from intensifier/hedge-sentence dropping, by design.
Aggressive — 77 tokens (25% shorter)
A paid-tier style. Same starting point as Balanced, plus a lower intent-collapse threshold and an exclusive consolidation pass that drops hedge adverbs and reminder framing outright.
Carefully write a professional follow-up email to a client. It's important that you keep it concise and to the point. Ensure the tone stays polite and friendly throughout. The email should mention that the project deadline has been moved to March 15th, and ask them to confirm they've received this update.
What changed beyond Balanced: “remember to” is dropped from the hedge sentence instead of kept whole, and “you should ask” is tightened to “ask” — still a complete, grammatical sentence, just denser.
Caveman — 74 tokens (28% shorter)
Maximum compression. Lowest intent-collapse threshold of the four, plus a compact-representation step that can convert a multi-part prompt into a semicolon-joined, label-free single line when that beats the alternative.
Carefully write a professional follow-up email to a client; friendly throughout; The email should mention that the project deadline has been moved to March 15th; ask them to confirm they've received this update; It's important that you keep it concise; to the point. Ensure the tone stays polite.
Every clause from the original survives — nothing is invented — but sentence boundaries are gone in favour of semicolon-joined fragments. This is the style where readability is deliberately traded for density; it's not what Structured (below) does.
Structured — 91 tokens (12% shorter)
Same phrase-level compression budget as Balanced, but always attempts to reorganise a multi-clause prompt into labeled sections — even when that doesn't reduce token count much, or would slightly increase it. Structured is the one style explicitly exempted from the token-safety gate for this reason: choosing it is itself a signal that structure is wanted.
Task: carefully write a professional follow-up email to a client Requirements: * The email should mention that the project deadline has been moved to March 15th, and you should ask them to confirm they've received this update. Output: * It's important that you remember to keep it concise and to the point. Make sure the tone stays polite and friendly throughout
Notice this saved fewer tokens than Aggressive or Caveman — that's expected. Structured optimises for scanability, not minimum length.
Which style to pick
- Balanced — default choice for most prompts; available on every plan.
- Aggressive — verbose prompts you send often, where you want more cut without losing grammar.
- Caveman — maximum compression, when density matters more than a polished read (e.g. internal tooling, not something you'd paste into an email yourself).
- Structured — multi-part requests with several distinct requirements or constraints, where a model reading a clearly labeled prompt matters more than shaving tokens.
Aggressive, Caveman, and Structured are Pro Max features; Balanced is available on every plan, including Free. All four run the identical hard-constraint validation — none of them can drop a number, link, or format requirement any more than the others can.
See our guide to reducing prompt tokens for the manual version of this same judgment call, and the full V1 benchmark for how these styles perform across 1,400 test executions, not just one example prompt.
Try all four styles on your own prompt.
See exactly which changes each style makes, side by side.