| name | anthropic-ai-claude-code-skilld |
| description | ALWAYS use when writing code importing "@anthropic-ai/claude-code". Consult for debugging, best practices, or modifying @anthropic-ai/claude-code, anthropic-ai/claude-code, anthropic-ai claude-code, anthropic ai claude code, claude-code-2.1.88, claude code 2.1.88. |
| metadata | {"version":"2.1.201","generated_by":"Anthropic · Haiku 4.5","generated_at":"2026-07-04T00:00:00.000Z"} |
Exhen/claude-code-2.1.88 @anthropic-ai/claude-code@2.1.201
Tags: stable: 2.1.193, latest: 2.1.201, next: 2.1.201
References: package.json • README
Search
Use skilld search "query" -p @anthropic-ai/claude-code instead of grepping .skilld/ directories. Run skilld search --guide -p @anthropic-ai/claude-code for full syntax, filters, and operators.
API Changes
This section documents version-specific API changes in @anthropic-ai/claude-code v2.1.201 — prioritize these when using the CLI or SDK.
-
NEW: ClaudeDesignInput and ClaudeDesignOutput tools — perform Claude Design operations (generate, manage, sync designs); call with operation: "list" first to discover available actions source
-
BREAKING: ProjectsInput — parameter force renamed to present_to_user with inverted semantics: old force bypassed budget guards; new present_to_user marks the doc as the deliverable the user asked for source
-
BREAKING: ProjectsOutput (project_write) — knowledge object removed entirely, replaced with present_to_user?: boolean flag; code reading knowledge_size, max_knowledge_size, search_threshold, rag_active, or remaining_budget will crash source
-
BREAKING: ProjectsOutput (project_info) — knowledge.search_threshold, knowledge.rag_active, and knowledge.remaining_budget fields removed; only knowledge_size and max_knowledge_size remain source
-
NEW: FileReadOutput — added source?: "seeded" field to distinguish dedup matches from startup context (CLAUDE.md / nested memory) vs. prior tool results source
-
NEW: ReportFindingsInput — findings array items now accept category?: string (kebab-case slug like "correctness", "simplification", "efficiency", "test-coverage"); enables fine-grained finding classification for code review workflows source
-
REMOVED: PushNotificationOutput — idleSec?: number and hasFocus?: boolean fields no longer present; use sentAt timestamp and disabledReason enum for notification state tracking source
Best Practices — @anthropic-ai/claude-code@2.1.201
-
Define workflow meta as a pure literal object with no computed values or function calls — the schema enforces immutability; computed values cause parse failures and make resume semantics ambiguous source
-
Prefer pipeline() over parallel() in workflows when stages are logically sequential — it provides wall-clock speed (slowest single-item chain) without barrier latency while still running all items concurrently through the stages source
-
Set persistent: true on Monitor when watching session-length state (PR monitoring, log tails) — persistent ignores the timeout_ms default of 300s and runs until TaskStop or session end source
-
Use ScheduleWakeup with delaySeconds between 300–3600 (5–60 min) for idle polling loops rather than sleep commands — avoids blocking, respects prompt cache TTL, and keeps sessions responsive source
-
Spawn worktree-isolated agents for concurrent mutations using isolation: "worktree" — prevents file-lock collisions and allows parallel edits across a codebase without shared state hazards source
-
Pass run_in_background: false to Agent when you need its result before continuing — the default background mode is asynchronous; synchronous agents block until completion but return structured output immediately source
-
Use CronCreate with durable: true only when the user explicitly requests the task survive session restarts — the default durable: false is in-memory only and auto-deletes with the session source
-
Check agent() return value is non-null before using it in workflows — agents fail silently on terminal API errors, returning null instead of throwing; filter with .filter(Boolean) before downstream work source
-
Set effort levels on expensive verification phases — effort: "high" or "xhigh" on agents verifying code correctness or security ensures deeper reasoning; low effort is appropriate only for cheap mechanical stages source
-
Override agent model for specific task tiers — use model: "haiku" for mechanical stages (pagination, filtering, boilerplate) and model: "opus" for novel design/reasoning phases to optimize cost and latency source
-
Batch file reads using limit and offset parameters for large files — reading partial chunks avoids token overrun; many small reads in parallel are cheaper than one full read that truncates source
-
Set tool timeouts conservatively: Bash max 600s, Monitor max 3600s — long operations that timeout silently drop output; estimate wall-clock time and add buffer for network latency source
-
Resume workflows from runId after edits to cache unchanged agent() calls — the longest unchanged prefix of agent() calls returns cached results instantly; only edited/new calls re-run, preserving prior work source