← ClaudeAtlas

expert-analysislisted

Dispatches `forge-expert` subagents in parallel — one per chosen domain — to produce focused analyses of a feature against the codebase before a plan is drafted. Each expert covers one domain (architecture, performance, data/state, UI/UX, security, testing, build/tooling — pick from the role catalog) and returns a structured report citing `file:line` evidence.
V-Songbird/claude-plugins · ★ 0 · Data & Documents · score 78
Install: claude install-skill V-Songbird/claude-plugins
# Expert Analysis Dispatch Dispatch N domain experts in parallel via the `forge-expert` subagent. The agent's system prompt owns the role archetype, citation discipline, return format, and read-only constraints — this skill provides the per-dispatch parameters and orchestrates the parallel call. ## Required Inputs - The feature requirements as the user expressed them (verbatim, not paraphrased). - A short orchestrator summary of the codebase area the feature touches (≤ 10 bullet points or file paths). - 3–5 anchor files (`path` or `path:line`) the experts should start from. Experts walk the code themselves; the anchor list seeds, it does not bound. - The list of domains to cover, picked from `references/expert-roles.md`. ## Picking experts Use as many experts as the feature genuinely needs, capped at **5**. The selection is driven by the Step-2 search — every distinct architectural concern surfaced by the search becomes one expert dispatch. - **Cap: 5 experts.** Past 5, reports start to repeat each other and `/forge:master-plan` spends its budget reconciling overlap. - **Merge near-duplicates.** If two domains' analyses would overlap by > 50% (e.g. both reasoning about the same dialog state machine), dispatch one combined expert with both lenses in the role line, not two separate experts. - **Honor explicit user picks.** If the user names domains ("get a security review and a performance review"), dispatch exactly that list and skip the heuristic. There is no hard floo