giovanniluci

Pixel Crew — AI skill for Claude Code

AI community

Smart MOD for Claude Code.

How to install Pixel Crew

This entry records only its repository, not the path inside it, so there is no exact command to give. Open giovanniluci/pixel-crew and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Pixel Crew does

Smart MOD for Claude Code. Picks the cheapest model (Haiku, Sonnet or Opus) for every message and subagent to save tokens, with a pixel-art agent pane showing who's working, on which model and why, plus 5-hour and weekly usage bars. Zero-token graphics, cache-safe routing, customizable mascots.

Alternatives in AI

Pixel Agents — Compressed Reference — VS Code extension with embedded React webview: pixel art office where AI agents (Claude Code terminals) are an 6.5k ★
ClaudeCodeStatusLine — Custom status line for Claude Code showing model, tokens, rate limits, and git info in real-time 333 ★
STS2 AI Agent — https://github.com/user-attachments/assets/89353468-a299-4315-9516-e520bcbfbd4b English README: README.md STS2 191 ★

README

🦀 Pixel Crew

**Smart model routing for Claude Code. Every message goes to the cheapest model that can do the job, and a crew of three pixel workers shows you who is doing what.**

Haiku, Sonnet and Opus each get their own pixel mascot and their own progress bar in a side pane. You see which model picked up your request and why, how many tokens it used, what it cost, and how much you saved compared with running everything on Opus.

*🇮🇹 Italiano più sotto.*

![Pixel Crew pane](docs/Screenshot1.png)

![Pixel Crew pane](docs/Screenshot2.png)

Why

Most of what you ask a coding agent is light work: find a file, list some functions, rename a variable. Running all of it on the biggest model burns tokens and usage limits for nothing. Pixel Crew sends light work to Haiku, normal work to Sonnet and hard work to Opus, without you switching models by hand.

How it routes

Keyword rules (free). Each model has its own keywords, in English and Italian (list below). Keywords match whole words, so "bug" does not fire inside "debugger" and "error" does not fire after "no" ("no errors").
Haiku classifier (a few tokens). When the rules can't decide (no keyword, mixed signals, a long message or code), Haiku answers easy / medium / hard. If it takes more than 2.5 s or fails, the message goes to Sonnet.
Cache guard. The prompt cache belongs to one model, so stepping down in a long conversation re-reads everything at full price. Above 40k tokens of context the model never steps down; stepping up is always allowed.
Escalation. Reply "still not working" (or "non funziona ancora") and the next turn moves one tier up.
Short replies. "no", "yes", "ok", "thanks", "grazie", "ciao"… always go to Haiku. "go on", "continue", "vai", "procedi" keep the current model.
Subagents by type. Explorers and searchers → Haiku; planners, reviewers and architects → Opus, plugin agents included (feature-dev:code-explorer, pr-review-toolkit:code-reviewer…