Use Caveman and Measurement to Cut Claude Code Token Waste

Quick answer

Caveman is a third-party skill/plugin that encourages compressed communication to reduce token use. Evaluate it with controlled tasks; concise prose can save context, but omitted constraints and weak handoffs can increase total cost through rework. Optimize verified outcomes per token, not brevity alone. Review its current claims and installation at the Caveman repository, and use Claude Code's /context plus usage tooling for your own measurements.

Separate four token costs

Startup: system guidance, CLAUDE.md, memory, skill descriptions, and tool metadata.
Input: user prompts, files, logs, command results, and retrieved docs.
Reasoning/action: the iterative agent loop and tool calls.
Output: explanations, plans, reports, and handoffs.

Caveman primarily targets communication verbosity. It cannot repair a giant CLAUDE.md, an unfiltered log dump, redundant MCP servers, or repeatedly failed implementation.

Benchmark, do not believe slogans

Create a fixed task set:

explain a code path;
diagnose a seeded bug;
implement a bounded feature;
review a security-sensitive diff;
produce a handoff after compaction.

Run each with no plugins, Caveman only, Ponytail only, and the intended plugin stack. Keep model, repository revision, prompt contract, and verification commands constant. Record:

Metric	Why it matters
Total tokens/cost	resource use
Time to verified pass	productivity
Corrections	instruction quality
Tests/checks passed	correctness
Defects found by blind review	hidden quality cost
Handoff completeness	continuity
Context at finish	sustainability

Run multiple trials because agent behavior varies.

Compress without losing the contract

Good concise handoff:

Goal: fix refresh-token race; API unchanged.
Cause: concurrent rotation updates same row without compare-and-swap (`session.ts:84`).
Change: conditional update + losing request reload; tests in `session-race.test.ts`.
Verified: focused test 50x, typecheck, build pass.
Open: PostgreSQL integration not run locally; CI owns it.
Next: reviewer examine retry semantics; do not deploy.

Bad concise handoff:

fixed auth. tests pass.

Compression should preserve decisions, evidence, and risk while removing repetition and ceremony.

Audit the whole plugin stack

Each enabled plugin may add descriptions, hooks, tools, automatic workflows, and output conventions. Two plugins can duplicate planning or completion checks. Use /context, /mcp, plugin inventory, and controlled runs to identify overlap. Keep a plugin only when the improvement exceeds context, latency, security, and maintenance cost.

Combine efficiency techniques

filter logs and diffs before ingestion;
use exact file ranges and task-directed search;
move long reference material to on-demand skills;
use subagents for isolated research with concise summaries;
start new sessions from structured handoffs;
request the shortest output that preserves evidence;
disable unused MCP servers and plugins;
use deterministic scripts for repeated parsing.

Mastery lab

Benchmark one real debugging task across four configurations. Have a blind reviewer score correctness and handoff quality. Calculate tokens per verified success, not just percent reduction. Adopt Caveman only if its net effect is positive for your workload and team comprehension.

FAQ

Does shorter output always reduce billed tokens?

It reduces output tokens, but missing context may increase later input, reasoning, and correction costs.

Can Caveman and Ponytail be combined?

They target different waste—communication versus implementation complexity—but test for conflicting instructions and lost evidence.

What is a safe optimization target?

Minimize total time and cost to a verified, reviewable outcome under an acceptable defect rate.