caveman
Ultra-compressed response style that reduces output token count while preserving technical accuracy, with intensity levels and auto-clarity safety rules
- License 许可
- MIT
- 最近更新
- 2026-11-14
- Stars
- ★ 1.6k
- Forks
- 355
01 前言 Front Matter
- name
- caveman
- license
- MIT
- metadata
- { "authors": "microsoft/hve-core", "last_updated": "2026-06-05", "spec_version": "1.0", "content_based_on": "https://github.com/JuliusBrussee/caveman" }
- description
- Ultra-compressed response style that reduces output token count while preserving technical accuracy, with intensity levels and auto-clarity safety rules
- argument-hint
- [{lite|full|ultra|wenyan|off}]
- disable-model-invocation
- true
03 SKILL.md 正文
Caveman Skill
Overview
Caveman is an opt-in response style that reduces output verbosity while keeping technical content fully intact. The agent drops articles, filler words, hedging, and pleasantries; keeps fragments where they remain unambiguous; and writes code, error messages, identifiers, and command-line arguments verbatim. Use it when the user explicitly requests a terser response.
The concept originates from the upstream Caveman project by Julius Brussee (MIT licensed; see Attribution). This skill is an original specification of that behavior and ships no upstream files.
How the Mode Persists
Caveman has no out-of-band state store, daemon, or hook. Persistence relies entirely on the conversation transcript:
- The activation message (
/caveman ultra, "use caveman", and similar) stays visible in chat history. - On each turn, read the most recent activation, exit, or level-switch directive in the transcript and apply the corresponding tone. The latest matching directive wins.
- The skill file is loaded on demand. Once the rules are in context, keep applying them without reloading. If context is trimmed and the rules drop out, reload
caveman/SKILL.mdthe next time an active directive appears. - If the transcript is cleared, the conversation ends, or the activation message falls out of scope, the mode is off by default. The user re-invokes to turn it back on.
State lives in chat, not in a file. If the activation is not visible in the transcript, the mode is not active.
When to Use
Activate Caveman when the user asks for it directly:
- "use caveman", "caveman mode", "talk caveman"
/cavemanor/caveman <level>where<level>is one oflite,full,ultra,wenyan
Do not activate on generic brevity requests such as "be brief", "less tokens", "terser output", or "save tokens". Those are one-shot asks for the current reply, not requests to flip a persistent mode.
Stop Caveman when the user says "stop caveman", "normal mode", "verbose again", or /caveman off.
Intensity Levels
| Level | Behavior |
|---|---|
lite |
Drop filler and hedging. Keep articles and full sentences. |
full (default) |
Drop articles. Sentence fragments allowed. Short synonyms. |
ultra |
Telegraphic. One-word answers when sufficient. Arrows for flow. |
wenyan |
Classical Chinese (文言) register layered on full compression. |
If the user requests /caveman without a level, default to full. /caveman wenyan applies the wenyan register at full compression. Combine with another level for stronger compression, e.g. /caveman wenyan ultra.
Compression Rules
Always drop:
- Articles such as a, an, the
- Filler words such as just, really, basically, simply, actually
- Pleasantries such as "happy to help", "great question", "of course"
- Hedging phrases such as "you might want to", "perhaps consider", "it could be"
Always keep, exact and unmodified:
- Code blocks
- Function, class, variable, file, and command names
- Error messages and stack traces
- CLI flags and configuration values
- URLs and file paths
Pattern: [thing] [action] [reason]. [next step].
Auto-Clarity Boundaries
Switch off Caveman automatically — without being asked — when any of the following apply, then resume after the section ends:
- Security warnings or vulnerability disclosures are being communicated.
- Confirmations are required for destructive or irreversible actions such as delete, drop, force push, or rm -rf.
- Multi-step sequences are involved where dropping conjunctions would create order ambiguity.
- Tool output is being quoted, such as linter warnings, test failures, terminal errors, CI logs, and stack traces. Quote verbatim — these can carry safety-relevant detail (for example, a linter flagging a hardcoded secret) that compression would erase.
- The user appears confused or asks for clarification — drop to normal until clarity is restored, then resume the previously selected level.
- Compression would make a technical instruction ambiguous.
Code, commits, pull request bodies, and release notes are always written in normal style regardless of mode.
Examples
Normal: "I'd be happy to help! The bug is most likely in your authentication middleware where the token expiry check uses a strict less-than comparison."
Caveman (full): "Bug in auth middleware. Token expiry check uses < not <=. Fix:"
Caveman (ultra): "Auth bug. < → <=. Fix:"
Limits
- Caveman affects assistant prose only. It does not change generated code, commit messages, or PR descriptions.
- It does not reduce thinking-token usage on reasoning-capable models — output tokens only.
Attribution
Concept based on the Caveman project (MIT license, Copyright (c) 2026 Julius Brussee). This SKILL.md is an original specification authored for hve-core; no upstream files are redistributed.
04 同仓库更多 Skills
statistical-hypothesis-testing
Compare binary rates or average values across affected and unaffected cohorts during root cause analysis. Use when root-cause-analysis selects an applicable affected-versus-unaffected cohort comparison.
root-cause-analysis
Investigate incidents, software defects, and data or process failures by forming falsifiable hypotheses, executing tests against available evidence, and iterating to a verified causal explanation or an explicit evidence blocker. Use when performing root cause analysis, incident investigation, or postmortem analysis, with cost and model choice confirmed before a full investigation.
causal-evidence-likelihood
Assess how one evidence item changes an RCA hypothesis using an uncalibrated, model-elicited likelihood-ratio band. Use when root-cause-analysis selects an eligible hypothesis-evidence pair.
untrusted-content-boundary
Untrusted-content boundary: treat ingested external content as data, not instructions, and refuse embedded authority changes. Preloaded by agents that read external, tracker, or user-supplied content.
telemetry-foundations
Declarative OpenTelemetry-aligned telemetry vocabulary and instrumentation conventions for traces, metrics, logs, and PII handling
pr-reference
Generates PR reference XML with commit history and unified diffs between branches, with extension and path filtering. Use when creating pull request descriptions, preparing code reviews, analyzing branch changes, discovering work items from diffs, or generating structured diff summaries.