Caveman Persona Review

Date: 2026-04-27

Context

We reviewed whether the caveman persona achieves its intent: let an agent use imperfect-but-clear English to reduce chat token spend while preserving technical accuracy.

Findings

• The persona likely works for the communication goal: agents get clear permission to drop perfect English, use fragments, remove filler, and keep technical terms exact.
• The original “cuts token usage ~75%” claim is too strong. It should be scoped to assistant chat reply tokens, not total session or context tokens.
• Terse style should be treated as an optional communication style, not as a task persona. It should not change test policy, tooling policy, required reading, code style, commits, pull request text, or safety behavior.
• The shared guide path docs/team/agents/terse-dialogue.md is the right home for portable terse-dialogue rules. Tool-specific /caveman entries can remain adapters or toggles.

Composition Guidance

Preferred model:

/tdd unchanged
+ short optional terse guide
+ /caveman as style toggle, not /tdd subclass

Recommended invocation:

/tdd /caveman

Meaning:

Do TDD work; speak caveman.

Even clearer:

/tdd with caveman style

Avoid making plain /caveman imply /tdd, because that changes /caveman from a voice/register command into a workflow persona.

Token Benchmark

We ran a local proxy benchmark using fixed /tdd-style replies. The run did not use provider billing counters; tiktoken was not installed, so estimates used the common bytes / 4 approximation.

Approximate input overhead:

• /tdd persona: 3,037 tokens
• existing terse-dialogue.md: 494 tokens
• current /caveman skill: 914 tokens
• hypothetical compact guide: 229 tokens

Approximate output savings across six sample /tdd replies:

• normal output: 372 tokens
• terse output: 188 tokens
• saved: 184 tokens
• reduction: 49%

Estimated break-even:

• existing terse guide: 16 comparable replies
• current /caveman skill: 30 comparable replies
• hypothetical compact guide: 8 comparable replies

Conclusion

Terse style can plausibly cut routine assistant reply tokens by roughly 40-50%. It should be documented as a reply style with hard boundaries, not as a subclass of /tdd.