| name | opt-caveman |
| description | Token optimization — activate the vendored caveman skill to compress inter-agent prompts and subagent results, cutting output tokens without weakening any quality gate. Use when: compressing inter-agent prompts or subagent results, a long-running run is burning tokens, or trimming handoff notes without losing meaning. |
One-Liner
Compress the words between agents, never the rigor — cut tokens in prompts/handoffs/results while every quality gate stays fully intact.
Wraps ../../vendor/caveman (MIT) and applies the policy in
../../resources/token-budget.md.
When to activate
- On by default in every workflow —
workflow-intake/bootstrap/feature/team/bugfix apply
this skill to every inter-agent prompt and subagent result as a standing rule (see
../../resources/token-budget.md). The caveman-activate hook
records caveman_mode: on at session start.
- This is the default target; it is where most wasted "politeness" tokens live.
- The
caveman-budget hook only escalates (suggests a stronger level) when token usage runs high —
it is not the trigger; caveman is already on.
When NOT to activate
- User-facing output — never compress what the user reads, unless they explicitly opt in
(
/caveman or "talk like caveman"). Especially never for Non-Technical users
(../../resources/persona-mode.md).
- Reasoning/thinking — caveman shrinks the mouth, not the brain; it does not touch analysis quality.
Levels
lite — drop filler, keep full sentences (safe default for inter-agent prose).full — terse telegraphic style; for bulk mechanical handoffs.ultra — maximum compression; only for high-volume low-ambiguity data passing.
Which level for which traffic is set by the Level selection matrix in
../../resources/token-budget.md: instructions → lite, routine
handoffs/plans → full, bulk data dumps → ultra, and judgment-critical content (review/security/
design/requirements) → no caveman, full prose. When unsure: lite for prose, ultra only for
pure data, none where a gate depends on it.
Hard rule
If compression would risk a weaker review, test, or design result, don't — use full prose and a
strong model tier (../../resources/agent-tool-mapping.md).
Quality outranks cost, always. Pair with artifact compaction (write long outputs to files, pass
summaries) for the biggest savings.
Example (lite, inter-agent handoff)
Before (42 tokens):
Could you please go ahead and carefully implement the user login
endpoint, and make sure to handle all of the error cases appropriately?
After (9 tokens):
Implement user login endpoint. Handle all error cases.
Meaning is preserved; only filler is dropped. The same content going into a review or security gate
would stay full prose.
Gate: the target is inter-agent / handoff / data traffic — never user-facing output, reasoning,
or judgment-critical content (review, security, design, requirements); the chosen level matches the
Level selection matrix; and the compressed text carries the same meaning as the original. If any
of these fails, use full prose.
