claude-tuning

claude-tuning

Claude Code uses fewer tokens and answers faster. One install, about 5 minutes.

Works on macOS, Linux and Windows.

Install (5 minutes)

1. Run these two commands:

   claude plugin marketplace add NourEldinShobier/claude-tuning
   claude plugin install claude-tuning@claude-tuning
   

2. Restart Claude Code.

3. Type /claude-tuning:tune. Claude installs and configures everything that is missing: Bun, the plugins, rtk, codebase-memory, settings and rules.

4. Restart Claude Code once more.

Done.

You need

Setting the API keys is the only step you do yourself.

What	Required?	Get it
Bun	Yes	Installed by /claude-tuning:tune if missing
Claude Code 2.1.274+	For fast-jev-compaction	claude update
TYPESAFE_API_KEY	Optional	Turns on the Jev features below. console.typesafe.ai

What you get

Works right after install:

Part	What it does	Result
Skill suggestions (Jev)	Tells Claude which of its skills fits your request.	14 of 17 test prompts right
Model routing (Jev)	Moves hard coding subagent tasks to Opus. Never downgrades, never overrides a model Claude chose.	14 of 14 test tasks right

Added by /claude-tuning:tune:

Part	What it does
rtk	Compresses shell output before Claude reads it: up to 90% less. Installed with its official installer if missing
stash	Keeps big outputs in a local search index; Claude pulls back only what it needs
ponytail	Keeps code minimal
i-have-adhd	Short answers that lead with the next action
Tool rules in ~/.claude/CLAUDE.md	Tell Claude when to use the tools above
codebase-memory	Code graph; installed with its official installer if missing
fast-jev-compaction (Jev)	Compaction that keeps chosen messages word for word instead of summarising them
quicksilver (Jev)	Jev picks the files, log lines or items that matter, so Claude reads only those
compact-adviser (Jev)	Compacts on its own at a finished checkpoint, not mid-task. Confirm once with /compact-adviser auto
jev-steer-or-queue (Jev)	Decides whether a message you send mid-turn steers, queues or stops the task
toolgate (Jev)	Stops tool calls that follow prompt-injected instructions; everything else runs
Settings	700k auto-compact window, function hooks on (for compaction), the code-graph and rtk hooks, and auto-update for claude-tuning and stash

Your existing settings are kept. Backups: settings.json.bak, CLAUDE.md.bak.

Without a TypeSafe key

Everything still works:

• Skill suggestions and model routing turn off.
• Compaction falls back to Claude Code's own; compact-adviser, jev-steer-or-queue and quicksilver stay idle.

Setup options

Flag	Does
--only=rtk,stash	Installs only these
--no-rules	Leaves your CLAUDE.md alone

Running setup twice is safe. The second run changes nothing.

Turn things off

To turn off	Do this
rtk for one command	Run it as rtk bypass <command>
One part	claude plugin disable stash@stash (or ponytail@ponytail, …); toolgate: delete the hooks whose command contains toolgate from ~/.claude/settings.json, then npm uninstall -g @riskaverse/toolgate
All of it	claude plugin disable claude-tuning@claude-tuning

Tools we use

We copy no third-party code. Each tool installs from its own repo at a pinned version.

Tool	What it does	Reported gain	Licence	How we use it
rtk	Compresses shell output: git, tests, builds, docker, kubectl, package managers and more	Up to 90% less output	Apache-2.0	Installed by setup if missing, with its hook
stash	Keeps big outputs and files in a local search index. Claude pulls back only what it needs. Local only: no API keys, no network.	91 KB of git history came back as 9.7 KB (89% less)	MIT	Installed by setup; install alone: stash@stash
ponytail	Makes Claude write the smallest code that works	54% less code, 22% fewer tokens, 20% cheaper, 27% faster	MIT	Installed by setup
i-have-adhd	Answers lead with the next action: numbered steps, no filler	No published numbers; shorter answers	MIT	Installed by setup
codebase-memory-mcp	Code graph: "where is X defined, who calls X" without reading whole files	120x fewer tokens (3.4k vs 412k over 5 queries)	MIT	Installed by setup if missing
fast-jev-compaction	Jev picks which messages survive compaction; kept ones stay verbatim	No published token numbers; no summary drift	MIT	Installed by setup; function hooks turned on
quicksilver	Jev filters, labels and ranks many files, log lines or items, so Claude reads only the matches	86% fewer Claude tokens on 12 tasks, up to 20x faster	MIT	Installed by setup; the tool rules say when to use it
compact-adviser	Jev picks a finished checkpoint and runs /compact there	No published token numbers; avoids compacting mid-task	MIT	Installed by setup in auto mode
jev-steer-or-queue	Jev routes mid-turn messages: steer, queue until the turn ends, or stop	No token numbers; 450–650 ms per mid-turn message, ~90 ms per tool hook (Rust port in the fork)	MIT	Fork; installed by setup in shadow mode
toolgate	Jev checks each tool call before it runs; here only for prompt injection	~0.4 s per gated call	MIT	Fork; installed by setup with the injection-only policy

Gains for stash were measured in this repo. Gains for the other tools come from their own READMEs and benchmarks. They measure different things, so the numbers don't add up to one total. Pinned versions are in UPSTREAMS.md.

Services:

Service	Used by
TypeSafe Jev	Skill suggestions, model routing, compaction, quicksilver, toolgate

Runtime: Bun runs this code. Only Bun's built-ins are used: SQLite, fetch and spawn.

More detail, including pinned commits and what changes in each tool: UPSTREAMS.md.

• stash is written from scratch for this bundle. It lives in its own repo, so you can install it without this bundle.
• bun run upstreams shows which tools have newer releases.

To update one:

1. Read its changelog.
2. Change the version in src/upstreams.ts.
3. Run bun test.

Develop

bun install
bun test                 # unit tests, no network
bun run typecheck
bun test/eval.ts         # live skill-suggestion check (needs TYPESAFE_API_KEY)
bun test/eval-route.ts   # live model-routing check

Debug output: set SKILL_SUGGEST_DEBUG=1 or ROUTE_MODEL_DEBUG=1.

Credits

• Skill suggestions follow TypeSafe's skill-suggestion cookbook.
• Every installed tool is credited in UPSTREAMS.md.

Licence

MIT