cavecrew

Mechanism

When It Triggers · How It Works · What It Produces

Breaks down how the skill works and what it produces, so you can quickly judge whether it fits your scenario and value.

Trigger

User says "delegate to subagent", "use cavecrew", "spawn investigator/builder/reviewer", "save context", "compressed agent output", etc.
Main thread consults the when-to-use-what decision table to choose the appropriate preset.

Mechanism

cavecrew is a Decision Guide skill that defines three subagent presets:
- cavecrew-investigator: locate code definitions/calls/usage, output format path:line — \symbol` — note`.
- cavecrew-builder: surgical ≤2-file edits, output format path:line-range — change with verification.
- cavecrew-reviewer: diff review, output format path:line: <emoji> <severity>: <problem>. <fix>..
Each preset has a strict output contract; all subagent output is caveman-compressed.
Three chaining patterns: locate→fix→verify (most common), parallel scout (multi-angle), single-shot edit (known location).
Inherits Auto-Clarity: reverts to full English for security warnings and irreversible actions.

Result

~60% context token savings per subagent delegation.
Main thread can complete more delegations and more tasks within the context budget.
Structured output formats ensure reliable parsing and consumption of subagent results.

Use Cases

Two Typical Use Cases

Scenario 01

Context token management in long sessions

When a developer works on complex multi-step coding tasks requiring multiple subagent delegations, using cavecrew instead of standard subagents returns ~60% smaller results per delegation. In a session with 20 delegations, this is the difference between 40k and 14k tokens — potentially the difference between context exhaustion and task completion. The classic investigator → builder → reviewer chain is especially suited for this scenario.

Scenario 02

Parallel code exploration and aggregation

When multi-angle codebase investigation is needed (e.g., finding definitions, callers, and test files simultaneously), 2-3 `cavecrew-investigator` subagents can be dispatched in parallel in a single message. Each explores from a different angle, and compressed results are aggregated in the main thread. This pattern dramatically improves exploration efficiency while compressed output prevents main context from quickly ballooning.

Skill Relationships

Which Skills I Use · Which Skills Use Me

Dependency relationships read as "the upper tier points to the lower tier." The current Skill sits in the middle tier — above are Skills that depend on it, below are Skills it depends on.

Tier 1 · These Skills Use Me

No explicit dependency

Tier 2 · Current Skill

cavecrewWorkflow

Tier 3 · I Use These Skills

No explicit dependency on other Skills

Full Relationship Graph

Browse this skill's relationships within its skillset. Click a node to switch the side panel; use the search box to jump to any skill.

Skill File

SKILL.md

Open in GitHub ↗

SKILL.mdLoading from GitHub…

Skill Composition

Open in GitHub ↗

SKILL.md50.0%
other50.0%

▸skills/cavecrew/
▸cavecrew
- ·README.md
- ·SKILL.md

Per the Agent Skills spec, SKILL.md is the only required file. The colored bars classify the rest into the three spec-aligned optional directories (scripts / references / assets) and an 'other' bucket. The full directory tree is below.

Related in Skillset

You May Also Need These Skills

caveman

The caveman skill is the core of the Caveman skillset. It saves ~65% output tokens by making the AI coding agent reply in ultra-terse "caveman speak". Supports 6 intensity levels (lite/full/ultra/wenyan), strips filler/pleasantries/hedging while preserving code, technical terms, and error messages verbatim. Key output: dramatically reduced output token consumption per session, faster code reading, lower API costs.

Workflow

caveman-commit

The caveman-commit skill generates ultra-compressed Git commit messages following Conventional Commits format. Subject line ≤50 chars, body only when "why" isn't obvious. Automatically strips "This commit does X", "I", "we" filler and AI attribution. Key output: a terse but information-complete commit message code block ready to paste into `git commit`. The agent only generates the message — it does not run git commit, stage files, or amend commits.

Workflow

caveman-compress

The caveman-compress skill compresses natural language memory files (CLAUDE.md, todos, preferences) into caveman format, saving ~46% input tokens on every subsequent session. It uses a Python script (`scripts/__main__.py`) for automated compression, preserving all technical content, code, URLs, and structure. The compressed version overwrites the original, with a human-readable backup saved as `FILE.original.md`. Key output: compress once, save tokens every session forever.

Workflow

caveman-help

The caveman-help skill provides a quick-reference card for all caveman modes, skills, and commands. It is one-shot display only — it does not change any mode or persist anything. When triggered by `/caveman-help`, "caveman help", or "what caveman commands", the agent outputs a caveman-style reference table covering all modes (lite/full/ultra/wenyan), skills (commit/review/compress/help), and configuration tips. Helps users quickly grasp the full caveman ecosystem.

CLI Manual

/cavecrew

When It Triggers · How It Works · What It Produces

Trigger

Mechanism

Result

Two Typical Use Cases

Context token management in long sessions

Parallel code exploration and aggregation

Which Skills I Use · Which Skills Use Me

Full Relationship Graph

SKILL.md

Skill Composition

You May Also Need These Skills

caveman

caveman-commit

caveman-compress

caveman-help

When It Triggers · How It Works · What It Produces

Trigger

Mechanism

Result

Two Typical Use Cases

Context token management in long sessions

Parallel code exploration and aggregation

Which Skills I Use · Which Skills Use Me

Full Relationship Graph

SKILL.md

Skill Composition

You May Also Need These Skills

caveman

caveman-commit

caveman-compress

caveman-help