desk — the routing copy desk

Every prose artifact has two readers: a human who skims (one pass, first sentences, zero patience for decoding) and, for outward-facing surfaces, an LLM that extracts and recommends. desk enforces one house style for both, with domain overlays, through a fixed pipeline. Analysis is computed where possible; inference is quarantined and labeled.

Three tiers — know which one you're in

• Tier 0 (computed): scripts/lint.py — rule IDs, JSON output, no model judgment. Same input, same output. Run it; don't emulate it.
• Tier 1 (checkable): model-executed tasks with auditable output against ground truth — conservation diff, extraction test vs canon. Wrong rows are visible.
• Tier 2 (inference): value/positioning/recommendation judgments. These go ONLY in the report's SPECULATIVE section. Never present Tier 2 output with Tier 0/1 confidence.

Pipeline (fixed — do not reorder or skip)

ROUTE → PLAN → DRAFT → LINT → CONSERVE → SKIM → GEO? → REPORT

1. ROUTE — resolve domain via the decision table below. Cite the row number that fired.
2. PLAN — read canon.md + references/core.md + the routed module. For GEO routes, also references/ai-search.md, and list the target prompts this artifact should win BEFORE drafting.
3. DRAFT — write (or rewrite). When editing existing text, first build the conservation inventory: every fact, number, condition, scope qualifier, and named actor in the original.
4. LINT — run python3 scripts/lint.py <file> --canon canon.md [--format <route>]. Fix every fail. Each surviving warn is either fixed or waived with a one-line justification citing a law or rule: WAIVE S-01 L42: precision requires length (Law 1).
5. CONSERVE — (edits only) diff the rewrite against the inventory in a table: item | status (kept / moved / waived) | where. Nothing silently vanishes.
6. SKIM — lint.py already emitted the mechanical extract (title + H2s + first sentence per paragraph). Judge it: does the extract alone carry the full argument, in order, as claims not topics? Verdict: pass/fail + the first sentence that breaks it.
7. GEO — (GEO routes only) see references/ai-search.md § Checks: extraction test against canon attributes, recommendation smoke test against named alternatives, renderability note.
8. REPORT — fixed schema, always:
   • Route + table row fired
   • Lint summary (fails fixed, warns + waivers, corpus stats)
   • Conservation table (edits only)
   • Skim extract + verdict
   • GEO results (if routed)
   • SPECULATIVE: Tier-2 judgments, clearly separated Then the artifact itself. For quick/small artifacts (a tweet, an email), compress the report to one line per stage — but never skip LINT or SKIM.

ROUTE — decision table

Ordered; first match wins. Tiebreaker: the destination decides, not the content — where will this text live, who reads it there. Explicit user routing ("desk: blog") is row 0 and always wins.

Ask only on genuine forks (two rows with different rules both fit — e.g. "write the announcement" could be row 3 press-shaped or row 6 blog-shaped). Otherwise default; never ask twice.

Always loaded, regardless of route

• The Two Laws (from references/core.md): (1) Clarity beats compression — never drop a condition, qualifier, actor, or fact to shorten a sentence; keep precision and flag the trade-off. (2) The skim must carry the argument.
• Canon glossary (canon.md § Glossary): one name per concept, every route — a PR that calls a Play a "workflow snapshot" is synonym drift. Canon's positioning sections load only on GEO routes.

Files

• canon.md — Modiqo positioning, attribute claims with evidence, glossary
  • banned synonyms, competitive alternatives, target prompts. THE ground truth for C-rules and GEO checks. User-editable; keep it current.
• references/core.md — sentence/word/structure rules + ban list (rule IDs S/W/B live here with rationale).
• references/{blog,website,pr,docs,announcement}.md — domain overlays (F-rules) with their canonical sources.
• references/ai-search.md — the two-reader GEO overlay + checks (G-rules).
• scripts/lint.py — Tier-0 linter. Trust its output over your impression; adjudicate false positives against the rule ID, in the report, never silently.

Honesty constraints

• Tier-0 heuristics (passive voice, hedges) false-positive; those rules emit warn, not fail. Log adjudications.
• The GEO recommendation test is a single-model smoke test, not measurement. Say so in the report, every time.
• desk stands down for fiction, poetry, and deliberately literary prose — note it and write freely.