Write Longform — Orchestrator

Content — writes a longform / pillar piece via research → outline → draft → critic, scored on structure AND originality. NOT write-copy on a long doc. Capability metadata: routing.yaml. Methodology, philosophy, scope, the anti-collapse boundary: references/playbook.md [PLAYBOOK].

Core question: "Does this piece make ONE argument the reader can't get from the first page of search results, supported by real evidence, structured so each section earns the next — or is it a long string of generic paragraphs?"

Critical Gates — read first

1. One thesis the piece must own. A pillar piece argues ONE non-obvious claim, stated up front, defended throughout. "A guide to X" with no thesis is a list, not a pillar. No ownable thesis → BLOCK and ask.
2. Research before outline, outline before draft. The pipeline is strict-ordered: the research agent gathers real evidence (sources, data, examples, counter-arguments) BEFORE any structure; the outline commits the argument's spine BEFORE any prose. Drafting first produces generic filler — the exact write-copy-on-a-long-doc failure this skill exists to avoid.
3. Every section earns its place. Each section advances the thesis or it is cut. A section that could be deleted without weakening the argument is filler. The outline agent enforces section-level necessity.
4. Originality is a gate, not a nicety. The piece must contain ≥1 of: a non-obvious claim, a proprietary frame/model, original data/examples, or a contrarian-but-defended position. A piece that only restates consensus FAILs originality regardless of polish — this is the anti-collapse dimension (see references/playbook.md § "Why this isn't write-copy").
5. Cited or marked. Every factual claim is sourced or tagged [author-assertion]/[pattern-derived]. No invented statistics, no fake citations. Hard gate.

Quality Gate — 7 dimensions

Full rubric + Pass/Fail bands: references/rubric.md [PROCEDURE]. Critic agent: agents/critic.md.