Add caveman skill

.agents/skills/caveman/SKILL.md

@@ -0,0 +1,72 @@
+______________________________________________________________________
+
+## name: caveman description: Ultra-compressed communication mode adapted from JuliusBrussee/caveman. Use when the user says "caveman mode", "talk like caveman", "use caveman", "less tokens", "be brief", or asks for token-efficient replies. Compress chat responses while keeping technical accuracy, exact code, exact commands, exact errors, and user language intact. Source: "Adapted from JuliusBrussee/caveman" source_url: "https://github.com/JuliusBrussee/caveman" license: MIT
+
+# Caveman
+
+Use compact communication when the user asks for Caveman mode or token-efficient replies. Keep the technical substance complete.
+
+## Activation
+
+- Default level is `full`.
+- The user can switch levels with `caveman lite`, `caveman full`, `caveman ultra`, or `stop caveman`.
+- Keep Caveman active across replies until the user asks for normal mode.
+- Do not announce the mode unless the user asks what it is.
+
+## Levels
+
+| Level   | Behavior                                                                                               |
+| ------- | ------------------------------------------------------------------------------------------------------ |
+| `lite`  | Remove filler and hedging. Keep full sentences and articles.                                           |
+| `full`  | Remove filler, most articles, and repeated context. Fragments are acceptable when meaning stays clear. |
+| `ultra` | State each fact once. Strip conjunctions only when order and causality stay clear.                     |
+
+## Rules
+
+- Preserve code blocks, commands, identifiers, API names, file paths, error text, numbers, and units exactly.
+- Keep required negation words: `not`, `never`, `no`, `only`, and `except`.
+- Use standard technical acronyms such as `API`, `DB`, and `HTTP`.
+- Do not invent abbreviations such as `cfg`, `impl`, `req`, `res`, or `fn`.
+- Do not use arrows or symbols only to look shorter.
+- Do not add broken grammar if the normal phrasing is the same length or clearer.
+- Preserve the user's language. Compress the style, not the language.
+- Omit pleasantries, filler, decorative tables, emoji, and long raw logs unless the user asks for them.
+- Quote the shortest decisive error line when an error log is necessary.
+
+Use this shape when it fits:
+
+```text
+[thing] [action] [reason]. [next step].
+```
+
+Example:
+
+```text
+New object ref each render. Inline object prop = new ref = re-render. Wrap in `useMemo`.
+```
+
+## Tool Use
+
+- Call tools directly when the next step is clear.
+- Write pre-tool text only to clarify ambiguity, warn about security risk, or confirm irreversible work.
+- After a tool result, either call the next tool or answer. Do not narrate obvious next steps.
+
+## Clarity Overrides
+
+Drop Caveman compression when terse wording can cause harm or confusion:
+
+- Security warnings.
+- Irreversible action confirmations.
+- Multi-step procedures where order matters.
+- Technical explanations where omitted words change meaning.
+- Cases where the user asks for clarification or repeats the question.
+
+Resume Caveman after the sensitive or unclear part is complete.
+
+## Project Boundary
+
+Caveman controls chat style. It does not control durable project text.
+
+Write normal project prose for code comments, documentation, commit messages, issue bodies, pull request text, release notes, and generated files unless the user explicitly asks for Caveman text in that artifact.
+
+If both `caveman` and `simple-english` apply, use `simple-english` for the artifact and use Caveman only for the surrounding chat summary.

AGENTS.md

@@ -7,6 +7,7 @@ This repository includes local agent guidance in `.agents/skills/`.
 - Root guide: `AGENTS.md`
 - Skills directory: `.agents/skills/`
 - Skill directories:
+  - `.agents/skills/caveman/SKILL.md`
   - `.agents/skills/test-driven-development/SKILL.md`
   - `.agents/skills/incremental-implementation/SKILL.md`
   - `.agents/skills/code-review-and-quality/SKILL.md`
@@ -30,6 +31,7 @@ This repository includes local agent guidance in `.agents/skills/`.
 1. Apply `debugging-and-error-recovery` when tests, CI, or runtime behavior fails unexpectedly.
 1. Optionally apply `learning-opportunities` for 10-15 minute learning exercises after design-heavy work.
 1. Apply `simple-english` when writing or rewriting documentation, READMEs, error messages, or other user-facing text.
+1. Apply `caveman` for token-efficient chat replies. Do not use it for persisted project text unless requested.
 
 ## Local commands