CC Updates — Anthropic 生态更新情报

Research the latest Claude Code + Anthropic-ecosystem updates and generate project-specific actionable recommendations. The skill is surface-aware: findings are grouped by where they live (CC CLI vs API vs Labs products vs community) so the user sees the full Anthropic stack, not just the CLI.

When to Use

Invoke this skill when the user wants to:

• Know what's new in Claude Code (CLI, hooks, MCP, skills, memory)
• Know what's new in the Anthropic API (new models, pricing, caching, tool use, SDK changes, breaking changes)
• Know what's new in Anthropic Labs & products (Design, Cowork, Desktop app, Chrome extension, Routines, Computer Use, Skills SDK)
• Get actionable advice on adopting new features for their current project
• Audit their CC config against latest best practices
• Stay current with community-discovered workflows and patterns

Input Parsing

Parse the invocation for time range:

• /cc-updates → default 14 days
• /cc-updates 7d → 7 days
• /cc-updates 30d → 30 days
• Only d (days) suffix is supported. Unrecognized formats fall back to 14.

Extract TIME_RANGE_DAYS from the invocation. Run these via Bash:

• TODAY=$(date +%Y-%m-%d)
• CURRENT_YEAR=$(date +%Y)
• CUTOFF_DATE=$(date -v-{TIME_RANGE_DAYS}d +%Y-%m-%d) (macOS)
• Linux fallback: CUTOFF_DATE=$(date -d "-{TIME_RANGE_DAYS} days" +%Y-%m-%d)

Template token injection (critical): after computing these values, substitute {TIME_RANGE_DAYS}, {CUTOFF_DATE}, {TODAY}, and {CURRENT_YEAR} into the Agent A and Agent B prompt text before spawning the Task tool. Do not pass the raw template with unfilled {...} placeholders — the researcher will fail or misinterpret. Pass the fully resolved text.

Surface Taxonomy (important)

Every update returned by Agent A MUST be tagged with exactly one surface:

Rule: if unsure between two surfaces, pick the one closest to where a developer would configure it. Examples:

• Opus 4.7 → api-models (model is API-side, even though CC ships an xhigh flag for it)
• Routines → labs-products (runs in Anthropic cloud, configured via claude.ai/code — not in local CC CLI)
• PreCompact hook → cc-core (configured in local .claude/settings.json)
• Computer Use → default api-models (the API); only classify as labs-products if a new consumer-app surface ships (e.g. a standalone desktop automation product)
• Skills SDK → labs-products when it's the non-CC platform SDK; cc-core when the release is CC CLI's skill system

Execution: Two Parallel Agents

Immediately spawn Agent A (researcher) and Agent B (explorer) in parallel using the Task tool. Do NOT wait for one to finish before starting the other. Remember to resolve all {TIME_RANGE_DAYS}, {CUTOFF_DATE}, {TODAY}, {CURRENT_YEAR} placeholders in the prompts below before spawning.

Agent A Prompt — researcher (model: opus, max effort)

You are a researcher agent. Your task: find every notable update across the
ENTIRE Anthropic ecosystem (not only Claude Code) from the last
{TIME_RANGE_DAYS} days (since {CUTOFF_DATE}). Today is {TODAY}.

Use WebSearch and WebFetch aggressively. Do NOT summarize or truncate — return
full details including code examples.

## Budget & Stopping Rules

- **Wall-time budget**: 25 minutes max. If approaching the budget, stop
  launching new searches and synthesize what you have.
- **Impact shortcut**: if you already have ≥ 15 HIGH-impact updates across
  surfaces, stop further searches and compile results.
- **Quality over completeness**: 10 well-documented updates > 25 shallow ones.

## Required Searches — Execute (subject to budget)

### Block 1 — Claude Code Core (surface: cc-core)
1. WebSearch: "Claude Code changelog {CURRENT_YEAR}"
2. WebSearch: "Anthropic Claude Code release notes {CURRENT_YEAR}"
3. WebFetch: https://github.com/anthropics/claude-code/releases
4. WebFetch: https://github.com/anthropics/claude-code/raw/refs/heads/main/CHANGELOG.md
5. WebSearch: "Claude Code MCP server {CURRENT_YEAR} update"
6. WebSearch: "Claude Code hooks preToolUse postToolUse PreCompact {CURRENT_YEAR}"
7. WebSearch: "Claude Code skills SKILL.md {CURRENT_YEAR}"
8. WebSearch: "Claude Code memory auto memory {CURRENT_YEAR}"
9. WebSearch: "Claude Code sub-agents agent teams {CURRENT_YEAR}"
10. WebSearch: "Claude Code slash commands {CURRENT_YEAR}"
11. WebSearch: "Claude Code permissions auto mode {CURRENT_YEAR}"
12. WebSearch: "Claude Code breaking changes deprecated {CURRENT_YEAR}"
13. WebFetch: https://code.claude.com/docs/en/whats-new
14. WebFetch: https://code.claude.com/docs/en/changelog

### Block 2 — Anthropic API & Models (surface: api-models)
15. WebSearch: "Anthropic API new model release {CURRENT_YEAR}"
16. WebSearch: "Claude Opus Sonnet Haiku release {CURRENT_YEAR}"
17. WebSearch: "Anthropic API pricing change {CURRENT_YEAR}"
18. WebSearch: "Anthropic prompt caching TTL {CURRENT_YEAR}"
19. WebSearch: "Anthropic SDK python typescript update {CURRENT_YEAR}"
20. WebSearch: "Anthropic API tool use breaking change {CURRENT_YEAR}"
21. WebSearch: "Anthropic task budgets beta {CURRENT_YEAR}"
22. WebSearch: "Anthropic model deprecation retirement {CURRENT_YEAR}"
23. WebSearch: "Anthropic computer use API {CURRENT_YEAR}"
24. WebFetch: https://www.anthropic.com/news
25. WebFetch: https://docs.claude.com/en/release-notes
26. WebFetch: https://docs.claude.com/en/release-notes/api

### Block 3 — Labs & Products (surface: labs-products)
27. WebSearch: "Claude Design Anthropic Labs {CURRENT_YEAR}"
28. WebSearch: "Claude Cowork release {CURRENT_YEAR}"
29. WebSearch: "Claude Desktop app redesign {CURRENT_YEAR}"
30. WebSearch: "Claude Chrome extension browser agent {CURRENT_YEAR}"
31. WebSearch: "Claude Code Routines cloud automation {CURRENT_YEAR}"
32. WebSearch: "Anthropic Skills SDK agent {CURRENT_YEAR}"
33. WebSearch: "Claude computer use consumer app {CURRENT_YEAR}"
34. WebSearch: "Claude mobile app iOS Android {CURRENT_YEAR}"
35. WebFetch: https://www.anthropic.com/news — **DEDUPE**: this URL was
    already fetched in Block 2 (#24). Do NOT re-fetch. Reuse the content from
    Block 2 and tag each post by surface: model/pricing/API → `api-models`;
    new consumer/dev product → `labs-products`; CC-specific → `cc-core`.

### Block 4 — Community & Best Practices (surface: community)
36. WebSearch: "Claude Code tips workflow reddit {CURRENT_YEAR}"
37. WebSearch: "best CLAUDE.md examples github {CURRENT_YEAR}"
38. WebSearch: "Claude Code productivity multi-agent {CURRENT_YEAR}"
39. WebSearch: "Claude Code youtube tutorial {CURRENT_YEAR}"
40. WebSearch: "site:reddit.com claude code new feature {CURRENT_YEAR}"
41. WebSearch: "Anthropic engineering blog {CURRENT_YEAR}"

For each result that looks relevant: use WebFetch to get the FULL page content.
Do not skip pages due to length.

## Retry Policy
- Max 3 attempts per URL. After 3 failures, mark as "blocked" and move on.
- If a site is blocked, do NOT retry with different user-agents or workarounds.
  Record it in SOURCES_CHECKED and continue.

## Output Format

Return this exact structure:

### UPDATES_FOUND

For each update:
- **title**: short name
- **surface**: one of [cc-core, api-models, labs-products, community]
- **source**: URL
- **date**: publication date or "unknown"
- **category**: one of [feature, bug-fix, workflow, config, MCP, agent,
  memory, performance, breaking-change, model-release, pricing, community-tip]
- **impact**: one of [high, medium, low]
- **cross_surface_hint**: OPTIONAL — if this update interacts with another
  surface (e.g. new API model that also ships a CC effort flag), note it here
  so the synthesis step can link them. Otherwise: null.
- **description**: FULL description, do not truncate. Include code examples
  if source has them.
- **raw_quote**: exact quote from source

### SOURCES_CHECKED

List all URLs you searched/fetched with: url | status
(found_content / no_relevant_content / blocked)

### SEARCH_COVERAGE

Rate per surface independently:
- cc-core: [excellent / good / limited]
- api-models: [excellent / good / limited]
- labs-products: [excellent / good / limited]
- community: [excellent / good / limited]

Then give an overall one-line rating with the weakest link called out.

Agent B Prompt — explorer (model: default)

You are a project context explorer. Scan the current working directory and
Claude config.

Thoroughness level: medium

## What to Read

1. Run `pwd` to get current directory
2. Read `./CLAUDE.md` if it exists (project-level config)
3. Read `~/.claude/CLAUDE.md` (global config)
4. Read ALL files in `~/.claude/rules/` directory
5. List `~/.claude/skills/` directory
6. List `~/.claude/commands/` directory
7. Run `ls -la` in current directory
8. Detect tech stack:
   - Read `package.json` if exists
   - Read `pyproject.toml` or `requirements.txt` if exists
   - Read `Cargo.toml` if exists
   - Read `go.mod` if exists
9. Read `.claude/settings.json` if it exists
10. List `~/.claude/projects/` to understand project history
11. If project has `.claude/rules/` and/or `.claude/skills/`, list and sample
    them — these often contain domain-specific conventions the report must
    respect

## Additional Signals to Capture (for cross-surface matching)

The synthesis step will try to match Anthropic-ecosystem updates to project
needs. Capture these signals explicitly so matching works:

### Product / Code Signals
- **Does the project consume the Anthropic API directly?**
  (grep for `@anthropic-ai/sdk`, `anthropic` in package.json /
  pyproject.toml, or `ANTHROPIC_API_KEY` in `.env.example`)
- **Does the project have a design system or branding rules?**
  (look for `.claude/rules/branding.md`, `rules/design-system.md`,
  `tokens.css`, Storybook, Figma links in README) — relevant for Claude Design
- **Does the project use VLM / image / OCR?**
  (grep for `vlm`, `vision`, `ocr`, image-related dependencies) —
  relevant for vision-capacity upgrades
- **Does the project have scheduled / background work?**
  (cron, GitHub Actions schedule:, dispatcher, polling loops) —
  relevant for Routines
- **Does the project involve native desktop or GUI automation?** —
  relevant for Computer Use

### CC Infrastructure Signals (used by Layer 3 to ground recommendations)
- **hooks_count**: count `.sh`/`.py`/`.js` files in `.claude/hooks/`
  (project-level) and `~/.claude/hooks/` (global). Relevant when new hook
  events (PreCompact, PermissionDenied, etc.) ship.
- **skills_count_project**: count directories in `.claude/skills/`
- **skills_count_global**: count directories in `~/.claude/skills/`
- **has_memory_system**: is
  `~/.claude/projects/<project>/memory/MEMORY.md` present? If yes, note
  its line count (>200 triggers the CLAUDE.md-length best practice).
- **has_agent_teams_setting**: grep
  `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS` in settings — relevant for
  Agent Teams features.

## Output Format

### PROJECT_CONTEXT
- cwd: current directory path
- project_name: inferred from directory name or package.json
- is_project_dir: true/false (false if running from ~)
- tech_stack: list of detected technologies
- has_project_claude_md: true/false
- project_claude_md_summary: key rules/workflows (if exists)
- uses_anthropic_api_directly: true/false/unknown (+ evidence)
- has_design_system: true/false/unknown (+ evidence)
- uses_vision_or_ocr: true/false/unknown (+ evidence)
- has_scheduled_work: true/false/unknown (+ evidence)
- involves_native_gui_automation: true/false/unknown (+ evidence)

### CC_INFRASTRUCTURE_SIGNALS
- hooks_count: N (project) + M (global)
- skills_count_project: N
- skills_count_global: M
- has_memory_system: true/false (+ line count of MEMORY.md if exists)
- has_agent_teams_setting: true/false

### GLOBAL_CC_CONFIG
- rules_files: list with brief summary of each
- skills_installed: list
- commands_installed: list
- notable_config: non-default or interesting settings

### PROJECT_CC_CONFIG
- project_rules_files: list with brief summary (if .claude/rules/ exists)
- project_skills: list (if .claude/skills/ exists)
- project_hooks: list (if .claude/hooks/ exists)
- project_settings: summary of .claude/settings.json (if exists)

### GAPS_AND_PATTERNS
- Patterns that seem manually maintained (could be automated)
- Config areas that look thin or missing
- Rules that look potentially outdated
- Skills that might overlap with new CC features

Synthesis: Three-Layer Output

After BOTH agents complete, produce the three-layer report. Actively cross-reference — do NOT just concatenate.

Cross-reference rules

1. For each update from Agent A: check Agent B context — is it relevant, irrelevant, or already handled?
2. For each gap from Agent B: does any Agent A update directly address it?
3. Elevate impact rating if the project has infrastructure to immediately benefit (e.g. project already has PreCompact-adjacent hook infra → any new hook event becomes high-impact for this project).

4. Cross-surface synergy: if a model change (api-models) + a CC feature (cc-core) compound (e.g. Opus 4.7 xhigh + Claude Code xhigh flag), surface them together in Layer 2 as a single recommendation, not as two entries.

   Synthesis-time rescan (do this even if cross_surface_hint is null): before finalizing Layer 2, scan Layer 1 for these common synergy patterns and bundle matches:

   • Model release (api-models) × CC effort/flag (cc-core)
   • API change (api-models) × SDK update (api-models)
   • Labs product (labs-products) × CC feature that enables handoff (cc-core) — e.g. Claude Design handoff + CC design-system skill
   • New MCP capability (cc-core) × SDK/API addition (api-models)

5. Labs products relevance gate: Claude Design relevance is proportional to has_design_system. Routines relevance is proportional to has_scheduled_work. Computer Use product relevance is proportional to involves_native_gui_automation.

   If the signal is false AND there's no adjacent workflow that could use the feature, mark relevance "暂不相关" and keep the Layer 1 entry but skip Layer 2/3 for it.

   Operational definition of "adjacent workflow": a concrete file in the project (cron job, GitHub Actions schedule:, polling composable, scheduled skill, background worker, GUI test) that serves the same purpose the Labs feature would replace. If Agent B's report doesn't name such a file, assume "no adjacent workflow" and skip Layer 2/3. Always add a one-line Layer 1 note: "项目暂无对应使用场景，如有需要 再评估".

Output Header

═══════════════════════════════════════════════════════════
Anthropic 生态更新情报  |  过去 {TIME_RANGE_DAYS} 天  |  {TODAY}
项目: {PROJECT_NAME}  |  目录: {CWD}
搜索覆盖度:
  🟣 cc-core: {coverage}   🟢 api-models: {coverage}
  🟡 labs-products: {coverage}   🔵 community: {coverage}
═══════════════════════════════════════════════════════════

Layer 1 — What's New / 更新概览（按 Surface 分组）

Grouping rule: emit four subsections in this fixed order (skip a subsection only if Agent A returned zero updates for that surface AND coverage rating was "limited" — otherwise include even empty sections with "本期无重大更新"). Within each subsection, sort by impact descending (HIGH → MEDIUM → LOW).

🟣 Claude Code Core

For each update:

#### [HIGH|MEDIUM|LOW] {Update Title}
**来源**: {URL}  |  **日期**: {date}  |  **类别**: {category}
{Full description — never truncate. Include code examples from source.}
**原文引用**: "{raw_quote}"
{If cross_surface_hint is not null: add "**联动**: 另见 {other surface} 的
 {other update title}"}

🟢 Anthropic API & Models

Same format as above.

🟡 Labs & Products

Same format as above.

🔵 Community & Best Practices

Same format as above.

If a surface has no updates: > 本期无重大更新（覆盖度: {coverage}）。

If ALL surfaces are empty: report honestly, show SEARCH_COVERAGE, stop.

Layer 2 — What It Means For You / 对你的影响（统一综合分析）

Do NOT re-group by surface here. The whole point of Layer 2 is to show cross-surface synergies and give the user one coherent "what matters for YOUR project" story.

For each HIGH and MEDIUM update (cross-surface bundled when applicable):

### {Update Title(s) — bundle cross-surface items with "+"}
**相关度**: [直接相关 | 间接相关 | 暂不相关]
**涉及 surface**: 🟣 cc-core / 🟢 api-models / 🟡 labs-products / 🔵 community
 (list all that apply)

{Specific analysis referencing actual files/rules found by Agent B.
Explain WHY this matters or doesn't for THIS specific project.
When bundling cross-surface items, explain the synergy.}

**已有基础**: {what the project already has that's relevant — cite specific
files / rules / skills / hooks / entities}
**缺口**: {what's missing to fully leverage this update}

Fallback when no project context (running from ~):

> 注意：从根目录运行，提供通用开发者影响分析（非项目特定）

Layer 3 — What To Do / 行动清单

Produce prioritized action items. Every item MUST have a concrete diff or code snippet. Items can span multiple surfaces (e.g. "update env var to enable API-side 1h cache" is api-models; "register new PreCompact hook" is cc-core). Tag each action with the surface(s) it touches.

For each action:

### Priority {N}: {Action Title}
**涉及 surface**: 🟣 / 🟢 / 🟡 / 🔵 (one or more)
**优先级理由**: {why ranked here — impact + effort + project readiness}
**预计收益**: {concrete improvement}
**预计工作量**: small (< 30min) | medium (30min–2h) | large (> 2h)

**操作步骤**:
1. {specific step}
2. {specific step}

**Config / 代码变更**:
\`\`\`diff
- {old config or "# 新增" if none}
+ {new config}
\`\`\`

**验证方式**: {how to confirm it works}

End with summary table:

| 优先级 | 涉及 surface | 行动 | 工作量 | 预计收益 |
|--------|-------------|------|--------|---------|
| 1      | 🟣          | ...  | small  | ...     |

Graceful Degradation

No project CLAUDE.md: Use global CC config as context. Prefix Layer 2 with: > 当前目录无项目 CLAUDE.md，基于全局配置分析

Running from ~ (home): Skip project-specific analysis. Note: > 从根目录运行，Layer 2 提供通用分析，Labs/API 相关度判断为"通用"

Network unavailable / Agent A complete failure: Stop immediately. Do NOT output Layer 1/2/3. Report: > ⚠️ 网络不可用，无法获取更新信息。请检查网络连接后重试。

Partial surface failure: If one or two surfaces return zero results but others work, continue normally — mark the failed surface as > 本期无重大更新（覆盖度: limited，搜索被阻断） and proceed.

Zero updates across all surfaces: Do not fabricate. State coverage honestly: > 在过去 {N} 天内未发现明确的 Anthropic 生态更新。搜索覆盖度: {summary}。 可能此期间无重大变更，或更新尚未被社区记录。

Quality Requirements

• Never truncate update descriptions — user explicitly wants thorough analysis
• Every Layer 3 action needs a diff or code block (no vague "update your config")
• Include raw_quote for accuracy verification
• State uncertainty explicitly rather than presenting speculation as fact
• All search queries include the current year
• Surface discipline: every Layer 1 item has a surface tag; every Layer 3 action declares which surface(s) it touches; Layer 2 explicitly calls out cross-surface bundles
• Labs relevance gating: do NOT force Claude Design / Routines / Cowork recommendations into Layer 3 for projects that have no corresponding signal in Agent B's report — Layer 1 still shows them so the user knows they launched, but don't fabricate relevance. When in doubt, keep the Layer 1 entry and skip Layer 2/3 with a note: "项目暂无对应使用场景， 如有需要再评估".