Featured method

Caveman Compression

Caveman is a rule-based semantic compression style: write to the model in terse, telegram-like language that drops grammatical filler while keeping every load-bearing word. Think 'cave drawing,' not 'essay.'

Realistic savings

Realistic: 14–21% on real coding tasks across Claude Sonnet/Opus. Up to 39–45% fewer output tokens versus a plain 'be terse' instruction. Some directive tasks report ~60%.

Keep

  • Nouns & main verbs
  • Numbers & units
  • Names & technical terms
  • Meaningful adjectives
  • Uncertainty qualifiers (maybe, likely, ~)
  • Time/frequency markers (daily, by Friday)
  • Critical prepositions that change meaning

Drop

  • Articles (a, an, the)
  • Auxiliary verbs (is, are, will, have)
  • Pronouns (it, they, this)
  • Redundant prepositions
  • Pure intensifiers (very, really, quite)
  • Politeness padding & restated context

Before → after

Before

Could you please take a look at this function and let me know if there are any bugs in it, and if so, could you also explain what is causing them and how I might be able to fix them?

After

Review function. List bugs + root cause + fix.

Use it when

  • Interactive chat where prompt caching makes the economics favorable.
  • Agent pipelines where verbose intermediate output compounds across steps.
  • Directive tasks with clear inputs/outputs: code gen, data transforms, classification.

Skip it when

  • Nuanced reasoning or persuasion where grammar carries the argument.
  • Legal, medical, or financial precision where ambiguity is dangerous.
  • Creative writing where voice and rhythm matter.

Drop-in skill

caveman-skill.txt
# Caveman compression — paste as a system/style instruction
Compress my prompts and your reasoning to "caveman" style.
KEEP: nouns, verbs, numbers, names, technical terms, uncertainty, time markers.
DROP: articles, auxiliaries, pronouns, redundant prepositions, intensifiers, padding.
Preserve all facts and constraints. Stay unambiguous. Do not compress final user-facing prose unless asked.