You're dispatching a sub-agent to scan 40 projects, or running a 10-loop refactor, or doing a codebase audit. Claude Code queues 50+ tool calls. Each iteration adds more context: previous results, new findings, updated code blocks. By loop 8, your messages are 20K tokens. By loop 15, 35K tokens. Prompt inflation eats ~50% of your API budget on what should be a $2 task.
Caveman Mode kills token bloat. Compress comms ruthlessly — drop articles, use fragments, abbreviate, cut fluff. A 20K-token message becomes 5K. Repeat that 15 times and you've saved ~225K tokens. At Claude Haiku pricing ($0.80 per million), that's $0.18 in one run. Scale it across Aiden's content pipeline (weekly blog audits, multi-project scans, feature inventories) and you're looking at $80–$120/month in recurring savings — with zero quality loss on technical output.
Why Token Bloat Compounds in Loops
Single-shot API calls don't reveal the problem. You send one 10K-token request, Claude returns a 2K response, done. The economics are fine. But AI workflows aren't single-shot — they're loops. A codebase audit runs 5 phases: graph scan → community detection → bottleneck analysis → refactor candidates → implementation. Each phase includes the full audit history. Each response is fed back as context for the next question. By phase 5, you're carrying 50K tokens of previous results.
Add sub-agent dispatch (two agents working in parallel on independent tasks, both looping, both feeding context back to you) and the math explodes. You're not paying for 5 phases × 10K tokens. You're paying for 5 phases × 20K tokens (context carryover) × 2 agents (parallel paths) = 200K tokens, not 50K. That's 4× the cost for what should be the same work. The overhead is pure friction — verbose summaries, full code blocks, repeated context, explanations that could be half as long.
Caveman Mode flips the switch. Compressed messages are smaller context. Smaller context = lower per-iteration cost = same quality output, 25–75% less spend.
Three Levels — Pick Your Compression
Lite Mode. Keep 80% of your words, drop the rest. Remove filler ("Let me think about this…", "That's a great question…"), cut unnecessary adjectives, abbreviate common terms. Lite keeps comms clear while trimming ~20% off token count. Use it for routine debugging loops, standard refactors, anything where clarity is critical and time pressure is low.
Full Mode (default). The main Caveman offering. Articles gone ("the", "a" disappear where context allows). Fragments OK — "Found 3 bugs" instead of "I found three bugs." Code blocks abbreviated ("const x = 5;" not "I've declared a constant variable x and assigned it the value of 5."). ~50% token savings. Output is still technically correct — Claude understands fragments just fine — but reads like telegrams. Use this for multi-phase audits, sub-agent dispatch, any workflow looping 3+ times.
Ultra Mode. Maximum aggression. Abbreviate everything that compresses without losing meaning. "File X, line 42, rm [old code]" instead of "In file X at line 42, I recommend removing the following code block." Drop entire sentence structures if a fragment works. Store context in shorthand ("RLS err on visits tbl" = RLS error on visits table). ~70–75% savings. Risks are real — your output becomes harder to parse, and Claude might miss nuance if context is TOO sparse. Use only on dead-simple loops (pure data transforms, regex sweeps, mechanical refactors where ambiguity is impossible).
Real Token Cost — Aidxn's Content Pipeline
Aiden's blog workflow is the proof. Weekly blog audit: scan 169 posts, extract SEO keywords, flag outdated code examples, propose refresh candidates. Three phases, ~30 questions per phase, carried context. Standard mode: ~450K tokens/week. With Full Caveman: ~220K tokens/week. Weekly savings: 230K tokens = $0.18 (Haiku) or $0.52 (Sonnet) depending on which model you'd run it on. Annualized: $9–$27.
Add monthly multi-project inventory (scans across velocity8.5, velocity9-stock, TradePilot, Staff Operations Dashboard, RebuildLocationSort — 5 codebases, 15K+ files, 3 audits per year): each audit without compression costs ~800K tokens. With Ultra Caveman: ~210K tokens. Savings per audit: 590K tokens = $0.47 (Haiku) or $3.54 (Sonnet). 3 audits × $3.54 = $10.62/year, tiny number. But monthly sub-agent dispatch (parallel feature inventories on two projects, context flowing back) hits the compounding problem: 1.2M tokens (standard) → 350K tokens (Caveman Full) = 850K token savings = $6.80 (Sonnet) per run. 4 runs/month = $27.20/month recurring.
Across the full pipeline, documented in token logs: $320/year in pure savings. Not game-changing. But more importantly — faster loops. Caveman responses are smaller, cheaper to process, cache better. Claude Code handles them faster. The subjective experience shifts: tighter feedback loops, quicker iterations, less API wait time. Speed-to-ship improves.
When NOT to Use Caveman
Customer-facing copy. Blog posts, marketing pages, customer support responses, pitch decks — always full English. Caveman voice on a sales email reads as lazy.
Debugging where reasoning clarity matters. Root-cause analysis, security audits, architecture decisions — the explanation is the output. Don't compress those.
Mentee explanations. Teaching Elijah something or documenting a pattern he'll read later? Full clarity. Caveman is an internal friction-killer, not an educational tool.
Complex ambiguous problems. If there's any chance the compressed language loses nuance, don't compress. The $0.02 savings aren't worth the risk.
Frequently Asked Questions
Does Caveman Mode actually work, or is it just marketing?
It works. Token counts drop 25–75% depending on intensity. Claude handles fragments and abbreviations fine. The risk is ambiguity on edge cases, not quality degradation on straightforward tasks. For mechanical loops, the savings are real and the output is identical.
Does it work with every Claude model?
Yes. Haiku, Sonnet, Opus all understand compressed input. The style has no model dependency. Where model choice matters is context window size — Opus's 200K context window handles longer loops before compression becomes critical, while Haiku's 100K gets squeezed faster.
When should I offer Caveman instead of assuming it?
Offer proactively before: dispatching 2+ parallel sub-agents, scanning 10+ projects, running 5+ loop iterations, or doing mass refactors. Don't offer for: customer copy, debugging, mentee work, or single-shot questions. The rule lives in memory.
Can I mix Caveman levels in one conversation?
Absolutely. Start Full for a complex audit, drop to Ultra for a mechanical regex sweep, climb back to Full when ambiguity reappears. The style is context-aware, not locked-in.
What if my output becomes unreadable?
Tighten the feedback loop. If Ultra is causing issues, drop to Full. If Full is causing misunderstandings, go back to Lite. The goal is speed-to-completion and cost savings, but not at the cost of correctness. Feel out your threshold.
Does this apply to other AI APIs?
Yes. Any API with token-based pricing — OpenAI, Google, Anthropic — will benefit from compressed input. The technique is model-agnostic. Caveman Mode is Claude-native (my `/caveman` skill), but the underlying principle is universal: brevity reduces token count, smaller context = cheaper loops, especially on long-running tasks.
The Bottom Line
Caveman Mode is not a magic word that saves the world. It's a micro-optimization for long-running loops where context bloat compounds fast. A single blog post doesn't compress; a 40-project audit absolutely does. Use it ruthlessly on repetitive multi-phase tasks, skip it on ambiguous or customer-facing work, and watch your API costs (and feedback loop speed) improve. At scale — monthly inventories, parallel sub-agents, recurring audits — the compounding effect adds up to real money and perceptibly faster iteration. See the pricing page for Brain's token-cost breakdown, or read API cost optimization patterns for caching and batching, the bigger-dollar optimizations that stack on top of Caveman compression.