Fork-path desk writer (pipeline.44). One desk at a time, DEEP — charge assembled from the desk's slice, cheap propose→OK gate, bounded source-search with orchestrator reconcile, supplied citizen voices (pipeline.43), artifact to the desk's own corpus. Production ends at artifacts-on-disk; review is a separate decoupled flow. Run after /desk-slice.
Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Fork-path desk writer (pipeline.44). One desk at a time, DEEP — charge assembled from the desk's slice, cheap propose→OK gate, bounded source-search with orchestrator reconcile, supplied citizen voices (pipeline.43), artifact to the desk's own corpus. Production ends at artifacts-on-disk; review is a separate decoupled flow. Run after /desk-slice.
["docs/research/2026-07-11-desk-slice-fork.md — the one-doc (fork design + Task 1 charge-format rules)","docs/adr/0012-autonomous-deep-dispatch-write-edition.md — the decision (+ S289 side-fork addendum)","docs/media/charge_brief_template.md + docs/media/examples/charge_brief_c100_civic_exemplar.md — charge contract + the A-graded worked shape"]
/deep-dispatch — One Desk, Deep (fork path)
What this is
/deep-dispatch {desk} [{XX}] runs ONE desk deep against its slice from /desk-slice. It is the fork's writer stage (pipeline.44) — it does NOT hand back to /write-edition compile. The frozen edition path keeps the monolith; this path ends at per-desk artifacts on disk, and the decoupled review flow picks them up after.
v0.1 → v2.0 (S313): v0.1 was the edition-coupled pilot harness (read sift's slate, returned control to /write-edition Step 2/3). The fork uncouples it: slices in, artifacts out, review elsewhere. v0.1 in git history.
The principle
A packet prescribes subject + angle + "invent nothing" and produces cookie-cutter prose that misses the cycle's largest events (C100, traced). A charge points at raw signal + grants latitude, and the desk finds the story the packet walks past. The depth lever is dispatch-shape + why-context, not tool access alone (RESEARCH.md:1589). Two A-graded proofs (sports Tier-1; civic best-ever/#2) ran this shape. One desk holding its whole territory finds seams parallel desks walk past.
TWO LOCKED FORKS — re-read before every run (ADR-0012)
The orchestrator (this skill / Mags) spawns the source-search agents; the desk only writes. Desk agents keep Read/Glob/Grep/Write/Edit — never granted Agent capability. The catalog already enforces this; do not work around it.
The charge is pointers + tools, NEVER pre-assembled data. Name where the raw signal lives, do not transcribe it. A charge that pastes anomaly numbers / citizen rows IS the bone-filtered packet with more fields.
Hard caps (over-spawn guard, RESEARCH.md:1146)
≤3 source-search agents per desk run; ≤2 artifacts per desk per cycle.
Prerequisites
output/slices/c{XX}/{desk}.md — from /desk-slice. Missing → stop; do not improvise a slice from the world summary (curation is Mags-at-slice-time, not here).
output/engine_anomalies_c{XX}.json + output/world_summary_c{XX}.md present (the slice points into them).
Per-desk store writable: output/desks/{desk}/articles/. byline/desk ingest tags ride at post-publish (engine.46 substrate, built S311).
Step 1 — Assemble the charge
Scan the slice first: require('/root/GodWorld/lib/contextScan').scanFile(slicePath) — r.safe === false → abort, surface (blocks log to ).
r.matches
output/injection_blocks.log
Build the charge from [[../../../docs/media/charge_brief_template]], written to output/charges/c{XX}_{desk}.md:
PROJECT — the canonical block, verbatim-class (Task 1 held-rule): sheets are the world; real-world reasoning welcome, real-world facts the engine never stated are leaks; bug-is-event — "Cover it. Translate the event; strip the number. Do not scrub an anomaly as an 'artifact.' Do not soften a crime spike into a safety framework."
BEAT — the slice's LANE sentence, verbatim, + "You are {journalist}."
CYCLE — the slice's POINTERS block, verbatim. Pointers, never figures.
STORYLINES — the slice's STORYLINES block, verbatim, + "YOU pick which one carries your piece — the one that is most true about the city this cycle."
CANON-MAP — attached citizens per storyline + tools (MCP lookup_citizen / search_canon / search_world; dashboard API; Grep over output/) + rules (CANON_RULES.md; names from canon only; any vote lists all 9 with the math proven; ages = 2041 − BirthYear).
CRAFT — (a) supplied-voices anchor: "your storyline's citizens have spoken — supplied lines arrive with your writing charge; use at least one, verbatim or trimmed, never paraphrase-then-attribute"; (b) REAL ASKS ONLY: claim an interview/question/"did not respond" ONLY for an ask that actually happened (a supplied line or a recorded no-response in your materials); otherwise state absence plainly — a performed query that never occurred is fabrication (Task 1 new veer class); (c) canonize the finding, not the supporting metrics.
FRESHNESS — reconcile before writing; two sources disagree → newer/primary wins AND verify vs world_summary/MCP; a contradicted scope claim is a HARD STOP. (Writer-side belt; the structural catch is Step 4.)
Step 2 — Propose → cheap OK (Mike-direct S313 gate)
Launch the {desk} agent with the charge and ONLY: "Read your charge. Return 1–2 lines: which storyline you're taking and the angle. Do not write yet."
Mags checks territory fit + collision with the cycle's other desk runs — not taste. OK → proceed. Redirect → one line back; desk re-proposes once. Pennies spent here catch subject collisions before the deep spend.
Spawn ≤3source-search agents (S326 — the Haiku-pinned retrieval agent; proven Sonnet-parity + 0 fabrications on the C100 eval, ~3-4x cheaper per hop; falls back to Explore only if the catalog lacks it) scoped to the picked storyline, each a distinct angle so they don't collide:
Sports: (a) truesource/season for the player(s) in play; (b) prior canon / the arc; (c) this cycle's events + roster state.
Culture/other: adapt — one raw-signal angle, one prior-coverage angle, one cycle-events angle.
Exactly one seat may be retrievalLane=prior-published-arc: the existing "prior canon / arc" or "prior coverage / initiative arc" seat. Raw-signal, truesource, roster, cycle-event, anomaly, vote-math, and current-state seats must use exact-current or cross-file-reconcile.
The prior-arc dispatch names the storyline and asks the agent to call node scripts/notebooklmCanonSearch.js. It must return source title, source ID, citation number, excerpt, and reconcileVerdict. For a citizen origin question, the orchestrator may explicitly authorize --source-class canon-reference; otherwise the default published scope holds. The return gate in Step 4 is the enforcement — the agent holds Read/Glob/Grep and has drifted off the wrapper in testing, so verify the return rather than trusting the lane.
Each subagent returns sourcing as text — what it found, each claim traced to its source artifact/tab. They reach raw data at runtime via the tools the charge granted; the orchestrator does NOT pre-package data.
Serialize under quota pressure (S231 G-S2). If quota is tight, spawn one, await, then the next — a heavy parallel dispatch can be session-limit-killed (signature: <total_tokens>0</total_tokens> + ~500ms + session-limit string). Do not write the sourcing yourself from the EIC seat; surface the infra gap and stop.
Step 4 — Reconcile (orchestrator — the structural freshness guard)
May be delegated to one source-search agent in reconcile mode (S326): hand it the conflicting returns + the ground-truth pointers; it declares the hard stop, rules, and cites. The orchestrator still owns the outcome — read the ruling before passing anything to the writer.
Before the writer sees anything:
Cross-check the ≤3 returns against each other. Where two disagree on a specific fact (vote count, program scope/districts, date, dollar figure), the newer / primary source wins — verify against world_summary_c{XX}.md or MCP (lookup_initiative, search_world) before it reaches the writer.
A contradicted scope claim is a HARD STOP — resolve now, not in the prose. (The exact C100 OARI miss: a stale return said OARI was a D1/D3/D5 pilot when INIT-002 was citywide since ~C97; the contradicting signal was in-hand and unreconciled.)
Person-name identity is a standing check, same weight as vote counts/dates/dollar figures (kimi P3, 2026-08-03). Any named person in a locked record (civic-voice packet, generated quote, project filing) that conflicts with published canon is ruled canon-wins unless a retcon is explicit; log the drift as an engine.43 instance. (The C102 catch: "Roberta" vs Bobby Chen-Ramirez — one packet outlier against 10+ cycles of canon.)
Hand the writer reconciled sourcing + a one-line note per resolved fact; log each reconcile to the run log. (Structural because charge prose alone did not catch the S272 error; this step is why it can't recur silently.)
Record for every seat: retrievalLane, resultStatus, selected/used source IDs, citation count, and reconcileVerdict. Do not record the NotebookLM answer body. A prior-published claim that conflicts with current primary state loses and is logged conflict-current-wins.
Prior-arc return gate (S334 — the agent drifts, so the orchestrator checks). A prior-published-arc return is admissible only if all three hold:
first line is retrievalLane: prior-published-arc;
every prior-arc claim carries a NotebookLM source ID (uuid) + citation number + excerpt — a claim sourced to a file path (output/pdfs/*.pdf, an edition .txt, an archive file) means the agent read the artifacts directly and escaped the reviewed source scope;
the last line is a bare reconcileVerdict token from the enum, with no prose after it.
Fail any of the three → re-dispatch that seat ONCE, naming the specific violation. Fail twice → log the seat resultStatus=no_result, drop its material entirely, and continue the run without prior-arc sourcing. Never launder a file-path return into the writer's sourcing — it looks like a good return and is not. Two of two test dispatches failed this gate at wiring time; assume it fires.
Step 5 — Voice the citizens (pipeline.43)
Batch for the picked storyline's attached citizens — ask from the storyline context ("The Tribune is writing about . What's your honest take, as someone living it?" + the citizen's slice context line), record: true:
Entries {pop, name, quote, disp, recorded, fallback} — recording writes the speaking (PRESS page doc + gated intake; dials stay behind the cycle drain). Fallbacks are explicit entries, never silent gaps. Log per-call + total tokens to the run log.
Test-run mode (kimi P2, 2026-08-03): a pilot run — new model at the desk, harness change — may set record: false to exercise the full chain with no external writes (no PRESS page, no gated intake). The deviation MUST be named in the run's production-log row. Live runs stay record: true.
Step 6 — The deep write (LOCKED: desk writes, no Agent grant)
Launch the {desk} agent (routing per .claude/agents/REPORTER_DESK_INDEX.md) with the charge + reconciled sourcing + supplied lines:
You are {journalist}. Deep-dispatch mode. Your proposal was approved: {the OK'd pick+angle}.
{the charge}
ASSEMBLED SOURCING (reconciled — facts marked [verified] are cross-checked; trust them):
{the source-search returns, reconciled}
SUPPLIED CITIZEN LINES (their own words, fetched from their real state):
{POP-NNNNN | Name | "quote"} ← non-fallback entries for your storyline
Trace every specific fact to the sourcing above or reach it with your own
Read/Glob/Grep. Use at least one supplied line — verbatim or trimmed from
front/back, never paraphrase-then-attribute. For citizens without a supplied
line: quote only what the sourcing carries verbatim; otherwise paraphrase the
action. REAL ASKS ONLY — never claim a query or "did not respond" for an ask
that didn't happen. Write to output/desks/{desk}/articles/c{XX}_{slug}.md.
Emit: headline, "By {journalist} | Bay Tribune {Desk}", body.
Do NOT spawn agents. Do NOT read other desks' files. Write.
Supplied-line errors (kimi P1 + Mike's policy 2026-08-03 — two separate dynamics, never merged):
A supplied line is the citizen's own words, but it is not canon-insulated. The hard line is mutating a tracked entity: a wrong name for a tracked person (the "Watkins"/Ashford and "Roberta"/Bobby class), wrong district, wrong figure, or invented words in a real POPID's mouth. Do not print that segment — trim around it with ellipsis (front/back cuts only) and log the error to the production log. A named error pattern in the voice model is engine.43 signal. The guard is mechanical, not vigilance-based: Rhea's C102 gate is the enforcement behind the writer.
Social-world color is allowed. A first-name relationship reference inside testimony ("my friend Noor told me") is the citizen's real social world — essence, not citation. It needs no canon lookup and is never trimmed for being unverifiable. Only tracked-entity mutation is gated.
If a story ever genuinely needs a citizen who doesn't exist: never mint straight into the ledger — birth into Generic_Citizens (nothing is free), and only as last-case strict-necessity, not a pipeline. Businesses run the opposite dynamic: a new business named honestly in a story is the sanctioned way businesses become tracked (no generic-birth pool exists; ~90 tracked vs 386k population). The real-world-entity gate (canon substitutes) guards against real-world business/institution influx — it is not a citizen rationale.
Step 7 — Close the run
Append to output/production_log_c{XX}.md## /deep-dispatch one row per artifact: | Desk | Journalist | Storyline | Artifact path | File exists | Words | Searches | Voices ok/fallback | Reconciles | Status |. A run that produced no file is DROPPED with the reason — never a silent gap (G-W61).
Skill ends here. The artifact is unreviewed and unpublished. The decoupled review flow (EIC read + reviewer lanes, ~2 artifacts at a time) picks it up; publication happens there via /post-publish.
Measurement hooks (the fork's proof criteria + ADR-0012 cost gate)
Record per desk run: total token cost of the chain (charge + ≤3 searches + reconcile + voices + writer) vs the frozen path's per-desk share, and the review grade. The fork proves itself on: quality parity, token burn down, moves the world, per-article wiki taggability. Depth is proven; the bill is not.
