Skip to content
Back to search

quicksilver

Claude Code skill: hand bulk judgment calls to Jev. 86% fewer Claude tokens on a 12-task benchmark, up to 20x faster. One-line npx install.

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/uditakhourii--quicksilver/badge.svg)](/listings/uditakhourii--quicksilver)

HTML

<a href="/listings/uditakhourii--quicksilver"><img src="/api/listings/uditakhourii--quicksilver/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

npx github:UditAkhourii/quicksilver

The installer copies the skill into `~/.claude/skills/quicksilver` and asks for your Jev key **once**. Get a key at [console.typesafe.ai](https://console.typesafe.ai). Then restart Claude Code. That's it. Claude uses the skill on its own whenever a task looks like "read a lot to decide a little".

<details> <summary>Other ways to install</summary>

# pass the key non-interactively (CI, dotfiles)
npx github:UditAkhourii/quicksilver install --key YOUR_JEV_KEY

# as a Claude Code plugin
/plugin marketplace add UditAkhourii/quicksilver
/plugin install quicksilver@quicksilver

# from a clone
git clone https://github.com/UditAkhourii/quicksilver && cd quicksilver && ./install.sh   # or .\install.ps1

`JEV_API_KEY` or `TYPESAFE_API_KEY` in your environment also works. Needs Node 18+. No npm dependencies. </details>

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

.claude-plugin/marketplace.json · Manifest · 16 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. {
  2. "name": "quicksilver",
  3. "owner": {
  4. "name": "Udit Akhouri"
  5. },
  6. "metadata": {
  7. "description": "Quicksilver: offload Claude's bulk judgment calls to Jev."
  8. },
  9. "plugins": [
  10. {
  11. "name": "quicksilver",
  12. "source": "./",
  13. "description": "Offload Claude's bulk judgment calls to Jev. Fewer tokens, lower latency, more headroom on your limits."
  14. }
  15. ]
  16. }
README.md · README · 177 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. <p align="center">
  2. <img src="assets/banner.svg" alt="quicksilver" width="100%">
  3. </p>
  4. <p align="center">
  5. <a href="#install"><img src="https://img.shields.io/badge/install-npx%20one--liner-5fd4ff?style=flat-square" alt="install"></a>
  6. <img src="https://img.shields.io/badge/Claude%20Code-skill%20%2B%20plugin-d97757?style=flat-square" alt="Claude Code skill">
  7. <img src="https://img.shields.io/badge/tokens-%E2%88%9286%25-8ee6a6?style=flat-square" alt="-86% tokens">
  8. <img src="https://img.shields.io/badge/deps-zero-lightgrey?style=flat-square" alt="zero deps">
  9. <img src="https://img.shields.io/badge/license-MIT-blue?style=flat-square" alt="MIT">
  10. </p>
  11. <h3 align="center">Stop paying Claude to skim.</h3>
  12. A big share of every Claude Code session is spent reading things only to decide
  13. whether they matter. *Which of these 187 files handle auth? Which of these 3,000
  14. log lines are real failures? Which of these 200 tickets are refund requests?*
  15. Claude reads it all, pays for it all, and the context fills up with noise.
  16. **Quicksilver** is a Claude Code skill that hands those calls to
  17. [Jev](https://docs.typesafe.ai), TypeSafe's System One model. Jev returns typed
  18. verdicts (yes/no, a label, a score) in about a second, for $0.042 per million
  19. tokens. Claude gets back a shortlist and spends its tokens on the thinking
  20. only it can do.
  21. ```
  22. you ─▶ Claude ──"which files handle auth?"──▶ quicksilver ──▶ Jev (187 files, parallel)
  23. │
  24. Claude ◀──── 4 file paths + confidence ─────┘ ~2.4k tokens instead of ~26k
  25. ```
  26. ## Install
  27. ```bash
  28. npx github:UditAkhourii/quicksilver
  29. ```
  30. The installer copies the skill into `~/.claude/skills/quicksilver` and asks for
  31. your Jev key **once**. Get a key at [console.typesafe.ai](https://console.typesafe.ai).
  32. Then restart Claude Code. That's it. Claude uses the skill on its own whenever
  33. a task looks like "read a lot to decide a little".
  34. <details>
  35. <summary>Other ways to install</summary>
  36. ```bash
  37. # pass the key non-interactively (CI, dotfiles)
  38. npx github:UditAkhourii/quicksilver install --key YOUR_JEV_KEY
  39. # as a Claude Code plugin
  40. /plugin marketplace add UditAkhourii/quicksilver
  41. /plugin install quicksilver@quicksilver
  42. # from a clone
  43. git clone https://github.com/UditAkhourii/quicksilver && cd quicksilver && ./install.sh # or .\install.ps1
  44. ```
  45. `JEV_API_KEY` or `TYPESAFE_API_KEY` in your environment also works. Needs Node 18+.
  46. No npm dependencies.
  47. </details>
  48. ## The benchmark
  49. Twelve tasks a coding agent really runs into. Eight use real public data (a
  50. supercomputer log, Banking77, UCI SMS Spam, SST-2, the Hono codebase and its git
  51. history, lodash). Each task was solved by **a Claude Code subagent working the
  52. normal way** (Read, Grep, Glob) and by **Claude + Quicksilver**, and both were
  53. scored against hidden ground truth.
  54. <p align="center"><img src="assets/benchmark.svg" alt="benchmark" width="100%"></p>
  55. | # | Situation | Metric | Claude alone | + Quicksilver | Claude tokens | Token cut | Time | Jev cost |
  56. |---|---|---|---|---|---|---|---|---|
  57. | 1 | Real log triage (BGL, 2k lines) | F1 | 54% | 23% | 84.4k → 12.5k | **−85%** | 105s → 44s | $0.033 |
  58. | 2 | Needles in a noisy service log (3k lines) | F1 | 100% | **100%** | 53.5k → 2.5k | **−95%** | 56s → 64s | $0.044 |
  59. | 3 | Support ticket routing (Banking77, 8 intents) | acc | 100% | **99%** | 15.0k → 2.7k | **−82%** | 60s → 6s | $0.003 |
  60. | 4 | Spam filtering (UCI SMS) | F1 | 97% | 92% | 20.4k → 4.3k | **−79%** | 61s → 8s | $0.004 |
  61. | 5 | Review sentiment (SST-2) | acc | 97% | **96%** | 19.5k → 3.0k | **−85%** | 85s → 6s | $0.003 |
  62. | 6 | Codebase discovery (Hono, 187 files) | F1 | 100% | **100%** | 26.4k → 2.4k | **−91%** | 43s → 6s | $0.013 |
  63. | 7 | Security review shortlist (40 files) | F1 | 100% | 89% | 9.5k → 2.5k | **−74%** | 48s → 3s | $0.001 |
  64. | 8 | CI failure triage (80 logs) | acc | 100% | **100%** | 9.1k → 2.5k | **−72%** | 30s → 3s | $0.002 |
  65. | 9 | Semantic search in lodash.js (17k lines) | hit@5 | 100% | **100%** | 9.0k → 4.1k | **−55%** | 39s → 39s | $0.213 |
  66. | 10 | "Where is X?" ranking over a repo | hit@3 | 100% | **100%** | 39.4k → 3.4k | **−91%** | 60s → 56s | $0.134 |
  67. | 11 | Commit classification (181 real commits) | acc | 83% | 76% | 17.5k → 3.2k | **−82%** | 106s → 5s | $0.003 |
  68. | 12 | Numeric threshold stress test | F1 | 100% | **100%** | 10.2k → 2.4k | **−76%** | 19s → 6s | $0.003 |
  69. **Bold** = within 2 points of Claude alone.
  70. > **The claim:** on bulk judgment work, Quicksilver cuts the tokens Claude spends by
  71. > **86%** (median 82%). It matches Claude's accuracy on **8 of 12** real-world tasks
  72. > and runs **up to 20× faster**, for a median of **$0.004 of Jev per task**.
  73. Numbers are measured, not estimated, and the benchmark is fully reproducible:
  74. see [`bench/`](bench/README.md). Claude-side tokens subtract the fixed
  75. per-agent overhead, measured with a control task. Quicksilver is charged for
  76. loading its SKILL.md on every task, plus every command and every byte of output
  77. Claude reads back.
  78. ## Where it shines, and where it doesn't
  79. **Use it for, and trust it on:**
  80. - **Finding needles in big inputs.** Real failures in a 3,000-line log, the 4
  81. auth files among 187, the right function in a 17k-line file. It hit 100% on all
  82. of them with 91–95% fewer tokens.
  83. - **Bulk routing with clear labels.** Support intents, CI failure causes, sentiment.
  84. Accuracy was 96–100%, about 10–15× faster than Claude reading each item.
  85. - **"Where is X?" across a codebase.** 10 of 10 top-3 hits with 91% fewer tokens.
  86. **Use it as a shortlist, and let Claude check the `?` items:**
  87. - **Look-alike code.** On the security task it caught every vulnerable file, but
  88. it also flagged safe twins at p≈0.5–0.65. Those are exactly the items
  89. Quicksilver marks as borderline for Claude to review.
  90. - **Subjective or house-style labels.** Commit types (`perf` vs `refactor`)
  91. scored 76% against Claude's 83%.
  92. - **Labels that encode unwritten policy.** The BGL supercomputer log's
  93. "alert" labels mark many FATAL lines as normal. Claude alone only reached 54% F1
  94. there, and Quicksilver 23%.
  95. **Don't use it for:** writing, editing, multi-step reasoning, or anything `grep`
  96. answers exactly. On huge scans (thousands of Jev calls) wall-clock time is about
  97. the same as Claude's. The win there is tokens and context, not speed. Add
  98. `--fast` to pack items and go about 10× faster on obvious needles.
  99. ## What Claude can do with it
  100. ```bash
  101. qs filter "Does this file handle user sessions?" src # which files matter
  102. qs filter "Does this line report a failure?" app.log --lines # log triage, repeats collapsed
  103. qs classify --labels "bug,feature,question" --items issues.jsonl # bulk routing
  104. qs rank "where do we issue refunds?" src --top 5 # relevance ranking
  105. qs find "the retry backoff logic" huge_module.py # locate lines in huge files
  106. qs ask "Does this contract allow termination without notice?" --state @contract.txt
  107. qs status # key check + lifetime tokens saved
  108. ```
  109. The output is built for an LLM to read: one line per hit, repeated log patterns
  110. collapsed into line-number ranges, classify results as id lists, and a `?` on
  111. anything borderline. Every run ends with a receipt:
  112. ```
  113. — 3000 scanned · 6 matched · 0 borderline · 64.2s · jev 1.0M tok ($0.0436) · ~56k Claude tokens not read
  114. ```
  115. ## Better prompting, for free
  116. Quicksilver quietly changes how Claude prompts. Instead of "read all of this and
  117. tell me what matters", Claude has to state **one narrow, typed judgment**: a
  118. yes/no condition, a closed set of labels, or a rubric. That's the same
  119. discipline that makes any LLM prompt reliable. The skill teaches it explicitly:
  120. one judgment per question, exact boundary cases, a catch-all label, and no
  121. arithmetic or dates. The question becomes a reusable, testable unit rather than
  122. a vibe.
  123. ## Safety
  124. - It never sends `.env*`, private keys, certificates, or credentials files. It
  125. respects `.gitignore`, and skips binaries and files over 2 MB.
  126. - Content goes to TypeSafe's API (`api.typesafe.ai`). TypeSafe states that Jev
  127. is not trained on customer data. Don't point it at anything you can't send to
  128. a third party.
  129. - The key is stored in `~/.quicksilver/config.json` with user-only permissions.
  130. `npx github:UditAkhourii/quicksilver setup --remove` deletes it.
  131. ## FAQ
  132. **Does this replace Claude?** No. Jev can't write, reason, or edit. Quicksilver
  133. makes Claude cheaper by keeping skimming out of its context.
  134. **What does Jev cost?** $0.042 per million input tokens, and output is free. The
  135. whole 12-task benchmark cost $0.45 of Jev.
  136. **Is this official?** No. It's an independent open-source project, not affiliated
  137. with Anthropic or TypeSafe AI.
  138. ## License
  139. MIT © Udit Akhouri

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.