target-mcp: Skill for Claude Code

TARGET checklist — parallel fan-out orchestration

You are the orchestrator. You do not score the paper yourself; you split the 39-leaf checklist across parallel subagents, each scoring a few leaves, then you merge their verdicts and submit them to the MCP server for validation and rendering. This cuts a ~20-minute single pass to ~2 minutes wall-clock.

The whole point is that all MCP calls stay with you (the orchestrator). The subagents are pure scorers: they receive the rubric and manuscript as plain text and return a verdict array. They do NOT need to reach the MCP tools themselves — keep it that way; it is simpler and avoids subagent-reachability questions.

Step 1 — Parse the manuscript (you, one MCP call)

Call parse_manuscript with the manuscript TEXT (or parse_pmcid for an open-access PMCID). Pass citation= with the paper's full APA reference so it appears on the render; on the PMCID path it is auto-built. Also pass supplements= when you have them — TTE methods (estimand, identifying assumptions) frequently live in a supplementary protocol table.

Keep the returned text_sha256 — you need it to submit.

Step 2 — Get the scaffold prompt (you, one MCP call)

Call assess_manuscript with mode="scaffold". It returns:

• system — the rules + the full per-leaf rubric for all applicable leaves,
• user_content — the manuscript as the server ingested it (the ONLY text evidence quotes may be copied from),
• leaf_ids — the ordered list of leaves to score (usually all 39).

Do NOT hand-score. Move to the fan-out.

Step 3 — Split the leaves into batches (keep 6x/7x pairs together)

Group leaf_ids into ~11 batches of 3–4 leaves. Rule that matters: keep each 6x specification leaf with its paired 7x emulation leaf (and lettered/roman siblings) in the SAME batch — judging the emulation of a component is more coherent when the same subagent just judged its specification. This is the one dependency in an otherwise-independent checklist.