General-purpose escape hatch to offload any self-contained subtask to a cheap, isolated Haiku worker. Use when the primary session has a small, well-scoped chunk of work that does not need the main model or the main context — a quick classification, a one-off rewrite, a contained question — and you want it done out-of-band. Dispatches a Haiku worker with a free-form prompt (and optional system override) and returns its answer. Prefer the more specific AgentFlow skills (read, search, summarize, transform, review, gen) when one fits; this is the catch-all.
Generate boilerplate code, tests, documentation, or config files from a short spec, in an isolated context so the source/scaffolding work does not crowd the main session. Use when the user asks to scaffold or generate code ("write tests for this module", "generate a config for X", "create boilerplate for Y") and the generated code must integrate with existing definitions rather than be redefined. Dispatches a Sonnet worker that imports existing units, matches the project's runtime/framework, covers edge cases, and returns code only (optionally writing it to a path).
Read a single file and return ONLY the parts relevant to a query, instead of loading the whole file into the main session. Use when the user wants to check, find, or answer something about one specific file ("what does X do in foo.ts", "does this file handle Y", "show me the auth logic in handler.py") and the file is large enough that loading all of it would waste primary context. Dispatches a Haiku worker that reads the file in an isolated context and returns a focused, line-cited extract. For broad multi-file search use agentflow-search; for whole-file edits use the native Read tool directly.
Quick, opinionated code review focused on bugs, security, or style, returning terse structured findings (severity, line, issue, fix) instead of prose. Use when the user asks to review or check code for problems ("review auth.ts for bugs", "any security issues in this file", "check this diff for off-by-one errors"). Dispatches a Sonnet worker that reads the code in an isolated context and returns a fixed SEVERITY/LINE/ISSUE/FIX block per finding, with no filler. This is the strongest AgentFlow skill — native review is rarely this structured or this terse.
Find code across many files from a natural-language description, returning ranked locations instead of loading every file into the main session. Use when the user asks where something lives or how something is implemented across a directory or codebase ("where do we validate tokens", "find the retry logic", "which files touch the pricing config") and you want the search to happen in an isolated context. Dispatches a Haiku worker that greps/reads candidate files and returns ranked {file, line range, relevance, snippet} hits. The native Explore agent is a fine alternative; this skill is the cheaper Haiku-pinned path with a structured JSON result.
Condense long content — logs, stack traces, documentation, command output, conversation history — into a format-locked summary, in an isolated context so the raw bulk never loads into the main session. Use when the user wants something large boiled down ("summarize this 5k-line log", "give me the key errors from this trace", "condense these docs to 3 bullets") and especially when they specify an exact format or word budget that must be honored. Dispatches a Haiku worker that respects format/length/focus constraints exactly.
Convert or restructure data — JSON to CSV, extract fields, reformat, clean, reshape — in an isolated context so the input/output bulk stays out of the main session. Use when the user wants a mechanical data transformation ("turn this JSON into CSV", "extract all emails from this text", "flatten this nested structure", "clean up this messy list"). Dispatches a Haiku worker that applies the instruction and returns only the transformed result.