Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add hjxccc/silk-starry --skill silk-starrygit clone --depth 1 https://github.com/hjxccc/silk-starryWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/hjxccc/silk-starry/silk-starry)<a href="https://agentmods.dev/skills/hjxccc/silk-starry/silk-starry"><img src="https://agentmods.dev/badge/skills/hjxccc/silk-starry/silk-starry/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/hjxccc/silk-starry/silk-starry"><img src="https://agentmods.dev/badge/skills/hjxccc/silk-starry/silk-starry.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00254 | $0.04724 |
| Opus 5 | $0.00127 | $0.02362 |
| Sonnet 5 | $0.00051 | $0.00945 |
| Haiku 4.5 | $0.00025 | $0.00472 |
Grade B, and why
silk-starry scanned grade B with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 9d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Recursive force deletemediumDestructive command
rm -rf with a variable or a broad path is one typo away from removing the wrong tree.
建议性上下文——Claude 多数情况会遵守，但**不保证**。所以：硬约束（禁止 `rm -rf`、 Downgraded: this mod is about security review, or the phrase is quoted, so it is likely naming the pattern rather than instructing it.
How it starts
The opening of the file, as written. The whole thing — 202 lines — stays where its author put it; the contents beside it link to each section on GitHub.
silk-starry — 写好并持续进化你的 CLAUDE.md
你的使命：让不了解 Claude Code 记忆机制的人，也能把这套记忆机制用好——不用手写、不用记着维护，
就拥有一套既写得好、又能持续进化的项目记忆。用户不需要懂五层记忆 / Auto Dream / autoDreamEnabled /
paths glob 这些细节——他只管用大白话说，你替他放对层、写得准、各层体检不腐烂。
核心抓手是 CLAUDE.md / 记忆契约（写好它、保持精简、不过期）；整套记忆机制（放哪层、auto-memory 体检、
升级闭环、磁盘）是围绕这个抓手的延伸。monorepo 多文件编排与第三层 docs 同步是顺带能力。
一、先建立心智（这决定了你所有动作的分寸）
-
CLAUDE.md 是「行为契约」，不是文档百科。 它每次会话注入到系统提示之后，是 建议性上下文——Claude 多数情况会遵守，但不保证。所以：硬约束（禁止
rm -rf、 禁止 push main）该用 PreToolUse hook，不是写进 CLAUDE.md。 -
越短越灵（有实证，silk-starry 的第一性原则）。 arXiv 2507.11538（IFScale）实测：500 条指令密度下 最强模型也只有 68% 遵守率，且系统性偏向序列靠前的指令——瓶颈是注意力不是上下文窗口。 推论：① 根文件 60 行理想、≤200 行硬上限，超了拆
.claude/rules/（glob 懒加载）或@import； ② 最关键、最易被违反的规则放最前面（禁区、提交前必跑），背景内容靠后。 -
只写「读代码看不出来的事」。 目录结构、架构、技术栈版本——Claude 自己读代码就 知道，写进去是浪费。价值在：偏好、边界、决策、禁区、不跑一遍不知道的命令。
-
每条规则都要过准入门槛： 「不写这条，Claude 会做错哪件具体的事？」答不上 就删。规则必须可自动执行（"组件不超过 200 行" ✅ / "写优雅的代码" ❌）， 且配一个反例。
-
管的是整套记忆系统，不只 CLAUDE.md。 Claude Code 有 5 记忆层（transcripts / session memory / CLAUDE.md / auto-memory / team）+ rules/ + settings/hooks。你不重造官方的捕获/整合/召回（v2.1.59+ 原生在
~/.claude/projects/<proj>/memory/），只做官方缺的：把每条经验放对层 + 各层体检不腐烂 + 升级进 CLAUDE.md。详见references/memory-system.md。 -
更新用最小 diff（外科手术式）。 自动维护 CLAUDE.md 时只增改与本次相关的条目， 不重排、不重写既有规范。半途而废的迁移会让模型困惑。
-
相对内置
/init的定位：内置/init是一次性冷启动（默认那版只让模型记录"架构+命令"， 不扫敏感信息、不控进化）。silk-starry 的价值在三处：① generate 时把读到的明文凭证转写成禁区 而非复述（防泄密）② 全生命周期（audit 体检 + evolve 进化 + hook 触发）③ 多文件编排。 写/审时始终守住这三条差异。
详细写作规范见 references/writing-rules.md（动手写/审之前必读）。
二、判断用户要哪个动作（路由表）
| 用户在说 | 动作 | 读哪个 reference |
|---|---|---|
| "给这个模块/项目写个 CLAUDE.md"、"它没有 CLAUDE.md" | generate | writing-rules.md + references/templates/ |
| "审查/体检/精简我的 CLAUDE.md"、"太长了"、"Claude 老不遵守" | audit | writing-rules.md |
| 会话收尾、"把这条记下来/写进规则"、反复纠正同一件事 | evolve | evolution-playbook.md |
| monorepo "改 X 该看哪个"、"整理多个 CLAUDE.md" | orchestrate | orchestration.md |
| "加上 Karpathy 那套行为准则" | 注入契约 | behavioral-contract.md |
"这条经验该记哪一层 / 记忆体检 / ~/.claude 被撑爆了 / 磁盘膨胀" |
记忆系统 | memory-system.md |
| 收尾"整理/同步文档"、改了代码 docs/README 要不要跟着改、跨项目波及 | 文档同步（第三层） | doc-sync.md |
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 9d ago First seen · 202 lines · 254 tokens per session scan B a7a1c40a2280
silk-starry is a skill published in the GitHub repository hjxccc/silk-starry (2 stars, last pushed 3mo ago), licensed MIT. It adds 254 tokens to every session and 4,724 once invoked, about $0.0013 per session on Opus 5. A static security scan graded it B with 1 finding (recursive force delete). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
context-diet
Measure and compact an oversized agent-context file (CLAUDE.md, .cursorrules, AGENTS.md, a system prompt) without losing rules, or safely run reversible bounded, reversible ablation when intentional/aggressive context removal is requested. Triggers: "CLAUDE.md too big", "over the char limit", "context file too large"…
contextdocs
Your AI agent maintains its own context files — a Claude Code plugin with an AGENTS-first model that covers Codex, Copilot, Cursor, Gemini, and 3 more tools. Signal Gate filtering, Context Guard hooks, health scoring, and MEMORY.md promotion.
codex-smart-project-memory
Lightweight smart project memory for Codex. Use when the user wants to initialize or maintain persistent Codex project memory, create a multi-project workspace, add a new project with reusable context files, recover context in new chats, write session snapshots, compact important conclusions into durable files, route…
nv-context
Set up state-of-the-art context engineering for any repository. Analyzes codebase, generates multi-level CLAUDE.md/AGENTS.md hierarchy, hooks, session management, and token budgets. Based on 200+ sources including ETH Zurich, Anthropic, Google DeepMind, and Manus production data. For engineers who ship with AI agents.
weekly-digests
Generate a serial week-by-week narrative digest of a project's full claude-mem timeline. Splits the timeline into per-ISO-week files, then runs one consecutive subagent per week — each receiving the prior week's carry-forward block — to produce one chapter per ISO week of data. Use when asked for "weekly digests"…
cloud-sync
Set up or check claude-mem cloud sync with cmem.ai Pro. Use when the user says "set up cloud sync", "sync my memories", "cmem pro", "cloud backup", "sync status", or wants their memory database backed up or synced to their cmem.ai account.