clade-performance-tuning

A model-agnostic agent-skills platform.

审核状态：已审核Quality 72Security 78

复制安装命令

用 Codex 或 Claude 安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它先审查 Skill 页面再帮你安装。

复制前请先查看来源、License 和安全提示。

项目 README

来源文件：README.md

抓取于 2026年8月27日

Tons of Skills

A model-agnostic agent-skills platform. The canonical layer is harness-free by construction; Claude Code is currently the verified-native harness. Other harnesses remain engineering candidates until their native-path integration is verified; source research alone is never presented as public support.

Version semantics: the release badge is this marketplace's display version. npm packages, including the ccpi CLI and publishable plugins, retain their own package versions; they are intentionally not expected to equal the display version. The version-surface checker governs the display surfaces without rewriting package semver.

Install

Inside Claude Code, one command installs the whole marketplace:

/plugin marketplace add jeremylongshore/claude-code-plugins

Or use the CLI:

pnpm add -g @intentsolutionsio/ccpi
ccpi install devops-automation-pack

Browse the marketplace · Explore plugins · Download bundles

Killer Skill of the Week — no-ai-slop by Peter Yang

Strip AI slop from any draft — named-pattern edits that keep the writer's real voice

no-ai-slop does two jobs and refuses to fake a third. In Edit mode it makes the minimum effective edit — cutting throat-clearing, weak verbs, and abstract nouns while deliberately preserving the writer's cadence, bluntness, humor, and honest admissions, so a rough draft still sounds like the same person afterward. In Detect mode it names each AI-slop pattern it finds, quotes the offending line, and gives the fix in a few words — and pointedly does NOT score the draft or guess whether an AI wrote it. That restraint is the whole point: AI detectors guess; named patterns are evidence the reader can check. MIT-licensed, single focused skill, actively maintained by Peter Yang.

"AI detectors guess. Named patterns are evidence the user can check." — Peter Yang

Grade: A | Week of July 22, 2026 (W30) | View on GitHub

Previous picks: tonone, mnemos, databricks-pack, kobiton-automate, skyvern, code-cleanup, web-analytics, token-optimizer, executive-assistant-skills, skill-creator, cursor-pack, crypto-portfolio-tracker. See all at tonsofskills.com.

Scale, labeled

Every number below names the cohort it counts and the command that reproduces it — an unlabeled count is how a corpus ends up with five contradictory answers to "how many skills."

Count	Cohort	Reproduce with
442	catalog plugins (catalog-entry cohort)	`node scripts/generate-readme-toc.mjs` over `marketplace.extended.json`
3,067	marketplace-visible skills (distinct)	`node -e "import('./scripts/corpus-resolver.mjs').then(m=>console.log(m.resolveCorpus('marketplace-visible').length))"`
347	agent definitions in plugins	`git ls-files 'plugins/*' \| grep '/agents/.\.md'`
19	plugin categories	`ls -d plugins/*/`

📦 Live npm Downloads

Across 396 published packages in the claude-code-plugins namespace. Updated daily by GitHub Actions.

Window	All packages	Established (>30d)
Last 24 hours	962	962
Last 7 days	2,920	2,916
Last 30 days	12,868	12,779

_{"Established" excludes packages first published within the last 30 days, so a bulk-publish event doesn't dominate the headline.}

Top 10 by last 30 days:

#	Package	Last 30d
1	`@intentsolutionsio/openrouter-pack`	556
2	`@intentsolutionsio/groq-pack`	496
3	`@intentsolutionsio/databricks-pack`	274
4	`@intentsolutionsio/clickhouse-pack`	273
5	`@intentsolutionsio/wallet-security-auditor`	263
6	`@intentsolutionsio/notion-pack`	258
7	`@intentsolutionsio/elevenlabs-pack`	244
8	`@intentsolutionsio/freshie-inventory-manager`	214
9	`@intentsolutionsio/supabase-pack`	210
10	`@intentsolutionsio/agency-os`	204

_{Last refreshed 2026-08-19T03:03:05.709Z.}

Ways in

Five real questions, five doors — each resolves to a live, generated surface, never a hand-maintained list:

By category — the table below, regenerated from the catalog on every sync.
By plugin — tonsofskills.com/explore, the full browsable catalog.
By skill — tonsofskills.com/skills, searchable across the whole corpus.
By job to be done — tonsofskills.com/cowork, curated bundles as one-click downloads.
By certification tier — the Certification section below, rendered from the live report.

Browse by category

The 19 categories below link into the live marketplace. Plugin counts are the catalog-entry cohort — regenerated from marketplace.extended.json by this generator; the catalog itself lives on tonsofskills.com, never in this file (§ 6A of the platform blueprint).

	Category	Plugins
🤖	AI & Machine Learning	36
🎭	AI Agents & Agency	10
🔌	API Development	26
💼	Business Tools	6
👥	Community	21
₿	Crypto & Web3	27
💾	Database	26
🎨	Design	2
🔧	DevOps & Infrastructure	36
📚	Examples & Templates	5
🧩	MCP Servers	16
📦	Packages	5
⚡	Performance	25
✅	Productivity	30
🎁	SaaS Skill Packs	106
🔐	Security	27
✨	Skill Enhancers	9
🧪	Testing	28
📁	Analytics	1

What the classes mean

Four artifact classes live in this repository, distinguished on sight and never blurred — provenance is a truth requirement here, not a UX nicety:

Class	What it is	How the reader can tell
Canonical skill	First-party, harness-free, the source of truth	No `.source.json` in its plugin directory
Generated adapter	A thin, machine-produced harness projection	Lives under a generated path with a "generated — do not edit" header
First-party package	An Intent Solutions distribution (npm, cowork zip)	`@intentsolutionsio` scope, IS-authored license
Upstream mirror	Somebody else's work, hosted mirror-by-default	`.source.json` present — upstream author, license, and pinned commit recorded

Certification

Not yet certified. The certification program (tiers T0–T4 with retained, hash-matched evidence) is a later epic of the platform blueprint; until its report exists, no artifact on this surface claims a tier. This line is rendered from the absence of certification-report.json — honestly, not cosmetically.

Contribute

Start with the contribution guide, then the intake and review standards every submission passes through:

Contribution requirements — including the AI-assistance disclosure expectation.
Skill submission intake standard — the tiered document matrix new plugins ship with.
External-PR review standard — how maintainers triage and what gets a submission merged.

Governance

STANDARDS.md — the public spec posture and the canonical-documents index (one owner per fact class).
The platform master blueprint — CI gates, release, docs governance, and this README's own landing contract.
GOVERNANCE.md · SECURITY.md · LICENSE

Provenance

External plugins are hosted mirror-by-default: the contributor's repository stays the source of truth, every mirrored source is pinned in a content lockfile, and upstream credit — author, license, resolved commit — is recorded in the mirror itself. Improvements flow by upstreaming to the author's repository, never by silently editing the mirror. The full decision record is the external-sync model.

License

MIT for the repository scaffolding and first-party tooling; each plugin carries its own license in its manifest, and mirrored plugins keep their upstream license verbatim.

其他

中风险

来源需自行核对维护者身份。
未检测到明显脚本安装指令。
可能需要外部 token、网络权限或第三方服务。
未检测到高风险命令。
扫描发现：2 条。

Codex — Git Clone 安装

安装前请先查看来源仓库和风险报告。
克隆仓库：git clone https://github.com/jeremylongshore/tons-of-skills-marketplace.git
将 "skills/.curated/clade-performance-tuning" 文件夹复制到 Codex 的 skills 目录中。
重启 Codex 让新的 skill 生效。

Codex — 手动复制安装

安装前请先查看来源仓库和风险报告。
从源仓库下载 SKILL.md 及相关文件。
在 Codex 的 skills 目录中创建新文件夹。
将所有 skill 文件复制到新文件夹中。
重启 Codex 让新的 skill 生效。

Claude Code — Git Clone 安装

安装前请先查看来源仓库和风险报告。
克隆仓库：git clone https://github.com/jeremylongshore/tons-of-skills-marketplace.git
将 "skills/.curated/clade-performance-tuning" 文件夹复制到 Claude Code 的 skills 目录中。
重启 Claude Code 让新的 skill 生效。

Claude Code — 手动复制安装

安装前请先查看来源仓库和风险报告。
从源仓库下载 SKILL.md 及相关文件。
在 Claude Code 的 skills 目录中创建新文件夹。
将所有 skill 文件复制到新文件夹中。
重启 Claude Code 让新的 skill 生效。

Cursor — Git Clone 安装

安装前请先查看来源仓库和风险报告。
克隆仓库：git clone https://github.com/jeremylongshore/tons-of-skills-marketplace.git
将 "skills/.curated/clade-performance-tuning" 文件夹复制到 Cursor 的 skills 目录中。
重启 Cursor 让新的 skill 生效。

Cursor — 手动复制安装

安装前请先查看来源仓库和风险报告。
从源仓库下载 SKILL.md 及相关文件。
在 Cursor 的 skills 目录中创建新文件夹。
将所有 skill 文件复制到新文件夹中。
重启 Cursor 让新的 skill 生效。

GitHub Copilot — Git Clone 安装

安装前请先查看来源仓库和风险报告。
克隆仓库：git clone https://github.com/jeremylongshore/tons-of-skills-marketplace.git
将 "skills/.curated/clade-performance-tuning" 文件夹复制到 GitHub Copilot 的 skills 目录中。
重启 GitHub Copilot 让新的 skill 生效。

GitHub Copilot — 手动复制安装

安装前请先查看来源仓库和风险报告。
从源仓库下载 SKILL.md 及相关文件。
在 GitHub Copilot 的 skills 目录中创建新文件夹。
将所有 skill 文件复制到新文件夹中。
重启 GitHub Copilot 让新的 skill 生效。

Windsurf — Git Clone 安装

安装前请先查看来源仓库和风险报告。
克隆仓库：git clone https://github.com/jeremylongshore/tons-of-skills-marketplace.git
将 "skills/.curated/clade-performance-tuning" 文件夹复制到 Windsurf 的 skills 目录中。
重启 Windsurf 让新的 skill 生效。

Windsurf — 手动复制安装

安装前请先查看来源仓库和风险报告。
从源仓库下载 SKILL.md 及相关文件。
在 Windsurf 的 skills 目录中创建新文件夹。
将所有 skill 文件复制到新文件夹中。
重启 Windsurf 让新的 skill 生效。

查看 SKILL.md 原文

name: clade-performance-tuning
description: "Optimize Anthropic API latency \u2014 streaming, prompt caching, model\
  \ selection,\nUse when working with performance-tuning patterns.\nconnection reuse,\
  \ and parallel requests.\nTrigger with \"anthropic slow\", \"claude latency\", \"\
  speed up anthropic\",\n\"anthropic performance\", \"claude response time\".\n"
allowed-tools: Read, Write, Edit
version: 1.0.0
license: MIT
author: Jeremy Longshore <jeremy@intentsolutions.io>
tags:
- saas
- anthropic
- claude
- performance
- latency
compatibility: Designed for Claude Code

Anthropic Performance Tuning

Overview

Claude latency has two components: time to first token (TTFT) and tokens per second (TPS). Different strategies target each.

Latency Benchmarks (approximate)

Model	TTFT (p50)	TTFT (p95)	Output TPS
Claude Haiku 4.5	200ms	600ms	~150
Claude Sonnet 4	400ms	1.2s	~90
Claude Opus 4	800ms	2.5s	~40

Optimization Strategies

Instructions

Step 1: Always Stream

// Streaming delivers the first token ASAP — user sees response instantly
// instead of waiting for the full response to generate

const stream = client.messages.stream({
  model: 'claude-sonnet-4-20250514',
  max_tokens: 1024,
  messages,
});

// First token arrives in ~400ms (Sonnet)
// Full response may take 5-10s, but user sees progress immediately
for await (const event of stream) {
  if (event.type === 'content_block_delta') {
    yield event.delta.text;
  }
}

Step 2: Prompt Caching — Faster TTFT

// Cached prompts skip re-processing — dramatically lower TTFT for large system prompts
const message = await client.messages.create({
  model: 'claude-sonnet-4-20250514',
  max_tokens: 1024,
  system: [{
    type: 'text',
    text: largeSystemPrompt, // 10K+ tokens
    cache_control: { type: 'ephemeral' },
  }],
  messages,
}, {
  headers: { 'claude-beta': 'prompt-caching-2024-07-31' },
});
// TTFT drops from ~2s to ~500ms on cache hit with large prompts

Step 3: Use Haiku for Speed-Critical Paths

// Haiku is 2-4x faster than Sonnet with 80% quality for many tasks
// Use for: classification, extraction, simple Q&A, routing decisions

const route = await client.messages.create({
  model: 'claude-haiku-4-5-20251001', // 200ms TTFT
  max_tokens: 10,
  system: 'Classify the intent. Reply with exactly one word: search, create, update, delete.',
  messages: [{ role: 'user', content: userInput }],
});

// Then use Sonnet/Opus for the actual task

Step 4: Reuse Client Instance

// BAD — creates new connection pool per request
app.get('/api/chat', async (req, res) => {
  const client = new Anthropic(); // DON'T
  // ...
});

// GOOD — single client shared across requests
const client = new Anthropic(); // Module-level singleton

app.get('/api/chat', async (req, res) => {
  const message = await client.messages.create({ ... });
  // ...
});

Step 5: Parallel Requests

// When you need multiple independent Claude calls, fire them in parallel
const [summary, sentiment, entities] = await Promise.all([
  client.messages.create({ model: 'claude-haiku-4-5-20251001', max_tokens: 200,
    messages: [{ role: 'user', content: `Summarize: ${text}` }] }),
  client.messages.create({ model: 'claude-haiku-4-5-20251001', max_tokens: 20,
    messages: [{ role: 'user', content: `Sentiment (positive/negative/neutral): ${text}` }] }),
  client.messages.create({ model: 'claude-haiku-4-5-20251001', max_tokens: 200,
    messages: [{ role: 'user', content: `Extract named entities from: ${text}` }] }),
]);

Step 6: Minimize Output Tokens

// Fewer output tokens = faster response
system: 'Be extremely concise. Use bullet points, not paragraphs.',

// Set tight max_tokens
max_tokens: 256, // Don't use 4096 for short answers

Output

Streaming enabled for all user-facing responses (first token in ~400ms with Sonnet)
Prompt caching reducing TTFT for large system prompts
Model routing to Haiku for speed-critical classification/routing tasks
Client instance reused across requests (no per-request connection overhead)
Parallel requests firing independent Claude calls concurrently

Error Handling

Issue	Cause	Fix
TTFT > 3s	Large uncached prompt	Enable prompt caching
Slow output	Using Opus for simple tasks	Downgrade to Haiku/Sonnet
Timeouts	Long generation + default timeout	`new Anthropic({ timeout: 120_000 })`
529 overloaded	API capacity	SDK auto-retries; add fallback model

Examples

See Latency Benchmarks table and six numbered strategy sections above, each with complete TypeScript code examples.

Resources

Next Steps

See clade-deploy-integration for production deployment patterns.

Prerequisites

Completed clade-install-auth
User-facing application where latency matters
Understanding of streaming and async patterns

发现问题？提交给管理员复核

暂无评论，成为第一个评论者吧！