Claude Code Flash — AI skill for Claude Code
Cut Claude Code token usage: hooks, Sonnet/Opus subagents, orchestration rules and /review + /task commands that keep big reads, boilerplate and diffs off the frontier model.
How to install Claude Code Flash
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open alexcding/claude-code-flash and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Claude Code Flash does
Cut Claude Code token usage: hooks, Sonnet/Opus subagents, orchestration rules and /review + /task commands that keep big reads, boilerplate and diffs off the frontier model. Self-hosted take on Spotify's shunt plugin.
Alternatives in AI
- System Prompts Leaks — Extracted system prompts from ChatGPT (GPT-5.4, GPT-5.3, Codex), Claude (Opus 4.6, Sonnet 4.6, Claude Code), G 38.6k ★
- Optimization — ai-research-skills AWQ, GPTQ, GGUF, Flash Attention, bitsandbytes 5.4k ★
- Launch An AI Feature — Workflow recipe — take an AI/LLM feature from a probabilistic-aware PRD to a launch-ready model card by chaini 1.3k ★
README
claude-code-flash
A local, self-hosted take on the idea in Spotify's ["Portal by Spotify cut my Claude Code token usage by 90%"](https://engineering.atspotify.com/2026/9/portal-by-spotify-cut-my-claude-code-token-usage-by-90): keep the frontier model from reading huge files or churning out boilerplate, and hand that work to a cheaper model instead.
Spotify's version routes files to workers on their internal Portal platform. This one uses nothing but Claude Code itself: four hooks, three subagents (two on Sonnet, one on Opus) and a rules file. No external service, no extra auth, no source code leaving your Anthropic account.
What it does
| Piece | Role |
|---|---|
Read hook |
Denies whole-file reads over FLASH_MIN_LINES (default 200) and tells Claude to delegate or read a targeted window instead. |
Bash read hook |
Denies cat / less / more / bat of large files when the output is not piped or redirected. cat big.log | grep ERROR, head and tail are still allowed. |
Bash diff hook |
Denies bare git diff, git show, and gh pr diff in the main session. Summary forms (--stat, --name-only, --oneline, ...) and piped forms pass. |
bulk-reader agent |
Sonnet subagent that reads files and returns structured bullets, never file dumps. |
code-writer agent |
Sonnet subagent that writes pattern-following boilerplate and reports back a file list instead of the code. |
reviewer agent |
Opus subagent that reviews a diff and returns verified findings as a table. |
SessionStart hook |
Injects `context/rules.md` into every main session (startup, resume, /clear, after compaction): response contract, roster, delegation, six-part briefs, task buckets, verification and reporting rules. Works like a global CLAUDE.md without touching yours. |
/review command |
Hands the working tree, a PR number, a branch or paths to the reviewer subagent and relays its findings table and ACCEPT / REWOR |
Related Skills
Sg Code
SG CODE — Claude Code-like AI coding assistant. Native desktop terminal with multi-provider support. ChatGPT (
Spread Your Claudes
Offload your coding agent's big-context reads to other AI CLIs (Gemini, Cursor, Codex, paid Gemini web app) so
AI Video Editor
AI video editor - turn raw footage into multiple professionally edited versions from prompts. Color grading, z
EchoFlow ClaudeCode
EchoFlow Code — a local-first desktop workspace for Claude Code and model providers. Run sessions, review diff
Santoryu Skills
Three blades, one mind. Opus (Claude Code) plans and reviews; two fast "blades" — a repo-aware Cursor agent an
Ag Local Bridge
Local OpenAI, Anthropic & Gemini API bridge for Antigravity — use Claude Sonnet/Opus 4.6, Gemini 3.7/3.6/3.5,
Related Agents
Deep Research Sonnet
Bulk-fan-out worker for multi-axis breadth-first research. Sonnet 5 tier (NOT frontier) — the canonical worker
Sonnet Worker
Sonnet 5 volume lane. Boilerplate, tests from a spec, mechanical migrations, docs, research reads, any task wi
Tp API Surface Tracker
PROACTIVELY use this when the user asks "what changed in our public API", "did we break anyone", "is this a br