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.
Sift did the editorial work — stories picked, reporters assigned, citizens verified, briefs on disk. This skill executes. Launch reporters, review what they write, compile the edition, validate, get Mara's approval, publish.
If sift is right, this is mechanical.
Prerequisites
Verify these exist before starting:
output/dispatch_c{XX}.json — from /sift — the mechanical launch spec. Per-article entries carry {slot, section, briefFile, reporter, desk, headline, outputPath, voiceDirective, spine, threeLayerKeys, initsTouched}; plus letters[], quickTakes[] (QT entries may have reporter: null), and header fields (cycle, edition, generatedAt, slateLockedBy, ...). This file is the source of truth for Step 1 — launch from it directly.
output/reporters/{reporter}/c{XX}_{SLOT}_brief.md — per-slot angle briefs from /sift (path is in each dispatch entry's briefFile; QT briefs live at output/quick-takes/c{XX}_{SLOT}_brief.md). Naming is per-slot, not per-reporter (G-W50 — the old c{XX}_brief.md per-reporter form was skill drift; a reporter with two slots has two briefs).
output/production_log_c{XX}.md — from /sift (story picks, assignments, citizen table). S225 pipeline.23 consolidated edition-flow to this canonical S195 path; pipeline.32 (S248) folded the civic side in too — civic decisions now live in the ## /city-hall section of this same unified log, not a separate production_log_city_hall file.
If dispatch_c{XX}.json is missing, /sift didn't complete its slate lock. Don't proceed.
Step 0: Bootstrap production log if missing (G-W47)
If output/production_log_c{XX}.md is absent but dispatch_c{XX}.json is present, bootstrap the log from the dispatch header rather than halting — /sift produces the dispatch spec even on runs where the consolidated log wasn't written (pipeline.32 unified-path convention). Seed it with cycle/edition/generatedAt + the slate (one line per articles[]/letters[]/quickTakes[] entry: slot | section | reporter | headline), then continue. If BOTH are missing, didn't run — stop.
Before launching reporters, recall what Mags has been developing on her citizen page (POP-00005) — the inner life she brings to compiling this edition:
node scripts/magsPageRecall.js --cycle={XX} --context="<dispatch slate summary: the headlines/sections from dispatch_c{XX}.json>"
Read-only — NO --mark (that's /sift's job; write-edition never rotates staleness and never appends to her page). Output is a memory-fenced block; it conditions editorial judgment during compile, never quoted as fact, never enters an article as canon.
Fail-open: empty/failure → empty output, exit 0, continue. A missing block never blocks the edition.
Every citizen quote in the edition should be the citizen actually speaking — one cheap citizenVoice.js call carrying their dials, bonds, tensions, and page memory — not a desk model inventing their voice. Runs in the main session (desk agents have no Bash), after the dispatch read, before any Step 1 launch.
Build the batch file. For each dispatch.articles[] entry, read its briefFile and collect the verified citizens attached to the story (POP ids from the brief's citizen section). One batch entry per citizen in output/voices/voices_batch_c{XX}.json:
{"pop": "POP-NNNNN", "ask": "The Tribune is writing about <spine / headline angle>. What's your honest take, as someone living it? <citizen's context line from the brief>", "record": true}
Ask wording is a starting value — tune against live quote quality (plan §Open questions).
Letters ride the same pass (pipeline.43 Task 4). For each dispatch.letters[] entry whose writer is a ledger citizen (has a POP id), add a batch entry whose ask is the letter seed: "Write to the Tribune about <topic> — what would you actually say?" Letters tolerate longer output — pass --max-tokens accordingly. Generic invented writers (S227 letters-texture exception) get no batch entry.
Run the batch:node scripts/citizenVoice.js --batch=output/voices/voices_batch_c{XX}.json > output/voices/voices_c{XX}.json. One output entry per citizen: {pop, name, quote, disp, recorded, fallback}. A fallback non-null entry (no DialState / call failure) means no supplied line for that citizen — explicit marker, never a silent gap. --record writes the speaking to the citizen's page (daypart='PRESS') plus one gated Reflection_Intake row (applied='no') at speak time; dials still apply only at the cycle drain. A fetched quote cut in editing still happened — page/intake record the speaking, publication records the printing.
Log cost to the production log: call count + token totals from stderr (expect ~10–25 DeepSeek calls per edition — pennies).
Availability gate: if citizenVoice.js doesn't support --batch/--record yet (engine-sheet pipeline.43 T1–2 pending), log voice packet unavailable in the production log and skip this step — desks fall back to the S227 quote rules for all citizens.
Rules
Brief-led mode is canonical (S215, closes G-W4). When a desk agent is launched by /write-edition, the agent's job is: read brief + IDENTITY.md, write. The desk SKILL.md boot sequences (LENS + RULES + workspace + voice files) are heavier than the brief-led model needs. The agents have been trained to short-circuit boot under brief-led invocation; this skill's rule overrides desk SKILL.md when there's conflict.
Reporters read their brief + IDENTITY.md. Nothing else. No world summary, no city-hall log, no sheet queries.
No calendar dates. Cycles, not months, not years (except past-event canon citations like "2032 MVP" or "2041 roster"). Day-of-week names (Monday/Tuesday/Wednesday) ARE allowed as scene texture (S227, closes G-W36) — the simulation has a 7-day week even if cycles span multiple weeks, and day-of-week reads as the kind of texture S208 anti-cookie-cutter discipline wants. "Tuesday morning light through stained glass" works; "October 17th" does not.
No engine language. No "cycle weight," "civic load," "sentiment score," "domain count." Citizens don't know these terms. Weather is "cool evening, northwest breeze" not "Weather: 67F from engine."
Story-driven layout. No fixed sections to fill. No filler.
Every citizen name was verified in sift. If a reporter introduces a new name not in the brief, flag it. Letters texture exception (S227, closes G-W40): letter writers may name generic neighbors (Mrs. Chen, the guy who runs the bodega) and small new local businesses without prior sift verification — letters are voice-from-the-stoop and strict pre-clearance kills the texture. BUT any new business named in a letter must be promoted to BIZ-NEW in NAMES INDEX / BUSINESSES NAMED for ingest, and any new citizen named with enough specificity (full name + role + neighborhood) lands in NAMES INDEX em-dash form for POP-pending promotion.
Reporter-range quote invention (S227, closes G-W37) — DEMOTED to fallback by pipeline.43. Supplied voice-packet lines (Step 0.7) are the primary quote source; in-scene invention now applies ONLY to (a) citizens with a fallback marker in the voice packet, (b) scene extras never attached to the story, or (c) cycles where the packet is unavailable (Step 0.7 availability gate). Within that scope the S227 craft rule is unchanged: when a beat reporter is "at the scene" of a covered moment, inventing direct attributed quotes from citizens present is within craft; a citizen referenced but not in-scene gets paraphrase only — no fabricated direct quotes. Capability + Mara graders should distinguish in-scene quote invention (allowed craft, fallback-scope only) from off-scene attribution (fabrication) and from paraphrase-then-attribute of a supplied line (also fabrication).
Output path (S215, closes G-W5). Per-reporter path is canonical: output/reporters/{reporter}/articles/c{XX}_{slug}.md. Disregard any desk SKILL.md instruction to write to output/desk-output/{desk}_c{XX}.md — that's the legacy desk-packet path, retained only as historical reference. Single source: per-reporter, matches sift output structure.
Memory Fence (Phase 40.6 Layer 2). The brief file handed to each reporter carries recalled canon excerpts from sift. Those excerpts must be wrapped via require('/root/GodWorld/lib/memoryFence').wrap(text, 'bay-tribune') before being embedded in the brief, so the reporter model receives them tagged as data, not instructions. Full convention: [[SUPERMEMORY]] §Memory Fence.
Context Scan (Phase 40.6 Layer 4). Before Step 1 launches any reporter agent, scan the brief file with require('/root/GodWorld/lib/contextScan').scanFile(briefPath). If r.safe === false, abort the launch and surface r.matches to Mags. Blocks are logged to output/injection_blocks.log. Never run a reporter against a flagged brief.
Step 1: Launch Reporter Agents
Read the production log for assignments. Launch each reporter with direct editorial direction.
Prompt structure per reporter — fully mechanical from the dispatch entry (G-W51, G-W52). Every field below is a literal dispatch.articles[i] value. Pass the brief PATH, never paste brief content — the agent reads its own brief; pasting it doubles tokens and invites the agent to treat recalled-canon excerpts as instructions (defeats the Memory Fence).
You are {reporter}. Brief-led mode.
ARTICLE — {headline}
Read your brief at {briefFile}. Read your IDENTITY.md.
Editorial direction: {voiceDirective}
Use ONLY citizens named in the brief. Write to {outputPath}.
QUOTE DISCIPLINE: citizens listed under SUPPLIED CITIZEN LINES below speak in their own words — quote those lines verbatim, or trim from the front/back; never paraphrase-then-attribute, never rewrite wording inside a supplied line. For any citizen WITHOUT a supplied line, attribute a direct quote ONLY when the verbatim line is supplied in your brief (or the packet it cites). If the brief says "VERIFY before quoting," that is a HARD STOP — no quote unless the exact filed statement is in hand; otherwise paraphrase the action. A synthesized attributed quote is the highest-severity fabrication class.
SUPPLIED CITIZEN LINES (their own words, fetched from their real state):
{POP-NNNNN | Name | "quote"} ← one row per non-fallback voice-packet entry for THIS article's citizens; omit the block entirely if the packet is unavailable or all entries are fallbacks
Do NOT read other files — write.
Voice-packet injection (pipeline.43). The SUPPLIED CITIZEN LINES rows come from output/voices/voices_c{XX}.json (Step 0.7) — filter to the citizens attached to this article's brief, skip fallback entries. Fallback-marked citizens are covered by the in-scene invention rules (§Rules), same as before the packet existed.
Launch via the {desk} agent (the dispatch entry names it). {reporter} + {briefFile} + {outputPath} come straight off the entry; no lookups.
Why the QUOTE DISCIPLINE line is in the launch prompt, not desk RULES (RB-1, C99 G-W2). Under brief-led mode the agent reads only its brief + IDENTITY.md — desk RULES.md is trimmed out (see §Rules) — so a quote gate placed in RULES would never fire. The launch prompt is the one channel every reporter receives, so the universal gate lives here. C99: Jordan Velez ran a synthesized quote "These should have been on the record two cycles ago" attributed to Keisha Ramos despite the B1 brief's explicit "VERIFY before quoting" gate; the actual filed line was "The process is clean — and now it's documented." Step 2 Pass 2 (below) is the editor-side backstop.
Unnamed-reporter QT sub-template (G-W49). Quick-Take entries in dispatch.quickTakes[] may carry reporter: null (no beat reporter assigned — the QT is a compile-time short piece). Do NOT invent a reporter name. Launch the {desk} agent with:
You are the {desk} desk. Brief-led mode.
ARTICLE — {headline} (Quick Take, short)
Read the brief at {briefFile}. Write a tight short piece to {outputPath}.
Byline: "By Bay Tribune {Section}" (no individual reporter). Do NOT read other files — write.
If a QT has no briefFile or source material, it is a compile-time drop — see Step 3 QUICK TAKES handling, don't launch a placeholder.
The E90 lesson: Agents told what to write produce articles. Agents told to figure it out spend all their tokens reading files and produce nothing.
Launch order (S215, closes G-W1 — rule, not named list):
Sports first — assigned sports reporters (typically Anthony / Hal / P Slayer) launch first; sports articles are the lowest editorial-judgment work and stabilize fastest.
Civic / business / culture in parallel — assigned reporters across these beats launch concurrently. No serial ordering required.
Conditional beats — accountability (Jax) + health (Mezran) launch when assigned by sift.
Letters LAST — letters react to the edition's topics, so the slate needs to know what shipped. Supplied-writer rule (pipeline.43 Task 4): a letter whose writer has a non-fallback voice-packet entry (Step 0.7 step 2) uses that voiced text as the letter's substance — the letters desk formats (salutation, trim, house style) but may not author substance for supplied writers. Include the writer's voiced text in the letters-desk launch prompt as a SUPPLIED CITIZEN LINES row. Generic invented writers (S227 letters-texture exception) are unaffected.
EIC-written sections — no agent launch (S227, closes G-W34). EDITOR'S DESK + any QUICK TAKE that survived Step 3 routing are written by Mags at Step 3 compile time, not launched as desk agents. The slate may carry them as proposals; treat them as Mags's compile-time work, not Task-tool launches. Future-cold-read of this skill should not try to launch a "mags-corliss" desk agent — there isn't one.
Don't hard-code reporter names in this skill — sift's assignment table is the source of truth. Reporters not assigned this cycle don't launch.
Reporter→desk-agent routing (S215, closes G-W2 + G-W3 + G-W9). Reporter briefs come from /sift as per-reporter files, but the Task-tool agent catalog only exposes desk agents (civic-desk, sports-desk, culture-desk, business-desk, letters-desk, freelance-firebrand, podcast-desk, chicago-desk) — not individual reporters. The routing table that maps reporter → desk agent lives at .claude/agents/REPORTER_DESK_INDEX.md (single source of truth).
Sift may also write output/dispatch_c{XX}.json with the mechanical mapping {reporter, story, desk_agent, brief_path, output_path, voice_directive} per launch — when present, read it and launch from it directly (skips the index lookup). When absent, fall back to REPORTER_DESK_INDEX.md.
Beat-axis routing — some reporters cover multiple beats and need beat-conditional routing:
Beat is editor-judgment per launch unless the dispatch.json names the agent. When in doubt: brief topic + reporter IDENTITY.md decide.
Firebrand boot trim (S215, closes G-W11). Freelance-firebrand IDENTITY/RULES/SKILL chain is heavier than the brief-led model needs. Under brief-led invocation, the agent self-trims to "read brief + IDENTITY, write." Same trim recommendation as G-W4 — desk and conditional agents alike.
Output: Each reporter writes to the outputPath named in its dispatch entry (output/reporters/{reporter}/articles/c{XX}_{SLOT}.md; QTs to output/quick-takes/c{XX}_{SLOT}.md).
Mandatory dispatch-result table (G-W61). After all launches complete, write one row per dispatch entry (articles[] + letters[] + quickTakes[] — every slot, no exceptions) to the production log:
| Slot | Desk Agent | Dispatch ID | Output Path | File Exists | Words | Status |
|------|-----------|-------------|-------------|-------------|-------|--------|
File Exists is checked at completion time (outputPath present on disk, non-empty). A slot that dispatched but produced no file is DROPPED — record it with the reason (e.g., agent returned empty, session-limit kill — the S231 G-S2 signature: <total_tokens>0</total_tokens> + ~500ms duration + session-limit string). Never let a slot vanish silently: C95 lost three culture pieces this way. Step 3 compile reconciles against this table — a slot with File Exists = no must appear in compile's dropped-slot report, not just disappear.
Step 2: Two-Pass Review (G-W63)
Pass 1 — completeness (mechanical). Before reading for quality, reconcile coverage against the dispatch spec. Walk every dispatch.articles[] + dispatch.letters[] + dispatch.quickTakes[] slot and the Step 1 dispatch-result table; confirm each slot either (a) produced a file that will be placed in a section, or (b) is an explicit, reason-logged drop. A slot that is neither is a silent loss — recover the file or re-launch it before compiling. No slot leaves Step 2 unaccounted for. This pass is what makes the G-W61 culture-drop class structurally visible instead of discovered at print.
Pass 2 — quality + framing. Mags reads every surviving article. Not a scan — a read. Check:
Did the agent follow the angle brief?
Are citizen names correct? (verify any you're unsure of via MCP)
Does the voice match the reporter?
Any fabricated facts, stats, game results?
Any direct attributed quote not traceable to the voice packet or brief? (RB-1, G-W2 + pipeline.43) Check every attributed quote against output/voices/voices_c{XX}.json FIRST (verbatim or front/back-trim match for packet-supplied citizens — a paraphrase attributed as their words is fabrication even if the meaning holds), the brief's supplied statement second. If the brief said "VERIFY before quoting" and the reporter quoted anyway, cut or paraphrase. In-scene reporter-range quote invention (§Rules, fallback-scope only) is allowed craft; off-scene synthesized attribution to a named civic figure is not.
Any calendar dates that should be cycle references?
Any engine language?
Any names not in the brief? (hallucination flag)
Two-pass canon verification order (S215, closes G-W13). When verifying any cycle-current stat or fact (player batting line, council member name, initiative dollar figure), read output/world_summary_c{XX}.md FIRST — cycle-current ground truth lives there. Only if the world summary is silent on the fact should you grep .claude/agents/civic-office-*/ files (career-canon roster lines) or docs/media/2041_athletics_roster.md (player career baseline). C93 lesson (G-W13): Aitken's .243 / 3 HR / 11 RBI line was flagged as fabricated because the verification order pulled career-baseline first; the world_summary carried the cycle-current line exactly. Cycle-stats live in world_summary; career-stats live in roster docs; both true at different time horizons.
Flag problems. Fix what's fixable. Cut what's broken. Better to have 8 clean articles than 13 with canon violations.
Update production log with editorial review (articles passed, cut, fixes applied).
Step 3: Compile
The edition is story-driven, not section-driven.
Take all articles that passed review
Order by what's most worth reading — front page was picked in sift, but may change if an article exceeded expectations
Label each with its section tag from sift assignments
If a section has no story this cycle, it doesn't appear
Canonical format exemplar (S227 — ADR-0006 Contract A): Use [[../../docs/media/EDITION_FORMAT_TEMPLATE]] as the authoritative format reference. Copy + fill placeholders. The template is the contract. The format snippet below is explanatory (rule definitions); the template carries the canonical literal shape (masthead ===-frame / ----divider sandwich for section labels / ### headline + optional **tagline** + plain By Reporter | Bay Tribune Section byline + --- body block / NAMES INDEX strict pipe-format / CITIZEN USAGE LOG (NEW CANON THIS CYCLE) sub-format the emit script parses for biz/faith).
If template + skill text disagree on detail, template is canonical. Skill drift gets caught next cycle; template stays the literal artifact reviewers parse against.
Article body header convention (S227, closes G-W38). Reporters may emit a headline line + *By {Reporter}* + --- separator at the top of their article body file, or write prose-only. Compile (this step) is responsible for canonical headline + byline placement per template — the headline comes from the dispatch headline field, the byline format is By {Reporter Name} | Bay Tribune {Section/Bureau}. Strip any reporter-emitted headline line / byline-line / leading --- before placing the article in its section block; emit the canonical headline + byline from dispatch metadata + template format. This keeps the compiled edition uniform regardless of which desk agents emit headers vs prose-only. Until desk-agent IDENTITY/RULES standardize emit shape, compile is the canonicalization layer.
Parser contract — the three coupled shapes compile MUST emit (S235/S240 regression, closes G-W62 + G-P-NEW1).lib/editionParser.js (hardened in ES-1) is the fixed point; compile emits what it parses. Three things, all three required together — S240's first PDF shipped unusable (taglines-and-bylines rendered as headlines, empty body divs) because the compiled .txt violated all three, and the parser's canonical-binding path silently fell back:
Headlines are ### Headline (H3), never # Headline (H1). The parser detects a headline only as ^### or ^**…**$ — a # H1 line is invisible to it, so the **tagline** underneath gets grabbed as the headline. Emit ### {dispatch.headline}.
Bylines are plain By {Reporter} | Bay Tribune {Section}, never **By …**. The parser's article filter matches ^By\s+; a bold byline doesn't match, collapsing the byline-article count and throwing the fail-loud bylineArticles.length !== rows.length guard in bindCanonicalHeadlines.
ARTICLE TABLE is the 4-column canonical shape — header | Slot | Section | Reporter | Headline | (extra cols like Words tolerated), every row's Slot a canonical ID matching ^(FP\d+|ED|C\d+|N\d+|S\d+|L\d+|O\d+|B\d+|CH\d+|Q\d+)$. A bare ordinal (1, 2) fails the pattern and drops the whole table to the legacy silent-skip path. Regenerate the ARTICLE TABLE from FINAL placement (the slots that actually shipped, in print order), not from sift metadata — sift order drifts from final placement (G-W62). The Headline cell is the binding source of truth; a parsed section with N bylined articles must have exactly N table rows or the parser throws.
Compile-time parser gate — run it, FAIL LOUD (G-P-NEW1). After writing the .txt, before continuing, run the shipped parser against the file and gate on it:
What this catches and how (precise — don't overclaim). The gate exits non-zero on all three malformations above, but by two distinct paths, because canonicalShape is a table-only boolean (hasAllRequired 4 cols && every Slot matches the canonical pattern) — it does NOT inspect headline level or byline style:
Modes 1 & 2 (H1 headline / bold byline):parseEdition itself throws inside bindCanonicalHeadlines (the bylineArticles.length !== rows.length fail-loud guard — H1 and bold bylines both break which chunks register as byline-bearing articles) → the catch exits 1. Empirically verified C95: bold byline → "found 0 byline-bearing"; H1 → "found 1 of 3".
So the command closes the full fatal S235/S240 class, but ### Headline (mode 1) is enforced by this compile instruction + the parser throw, not by the canonicalShape flag. The ### form also keeps stray H1/tagline cruft out of the retained body text. If the gate exits non-zero, the edition will not render — fix the table/headlines/bylines and re-run before any downstream step.
Edition format (rule reference):
------------------------------------------------------------
THE CYCLE PULSE — EDITION {XX}
Bay Tribune | Cycle {N} | Y{n}C{m} | {Season}, {Week}
Weather: [plain language — from production log] | City Mood: [plain language]
------------------------------------------------------------
Masthead format (per EDITION_PIPELINE.md §Published `.txt` Format Contract):
- Cycle: integer (matches edition number for full editions)
- Y{n}C{m}: cycle math — n = floor((cycle-1)/52) + 1, m = ((cycle-1)%52) + 1.
Cycle 92 = Y2C40. Replaces month-year (real calendar months don't align with cycle clock).
- Season: Spring | Summer | Fall | Winter (from cycle-to-season mapping)
- Week: First Friday | Second Friday | etc. (week within the cycle, optional)
Do NOT emit "October 2041" or any month-year token. The simulation is cycle-paced, not calendar-paced.
Section dividers MUST be `^-{10,}$` (ten or more dashes on a line by themselves).
The capability reviewer at `scripts/capability-reviewer/parseEdition.js` parses on this regex
and rejects `===` or other characters silently (zero sections detected → reviewer-lane skips).
G-W19 (S196) — DOC drift between this template and the parser cost a full reformat round in C93.
Section labels MUST be from this fixed allowlist (parser-enforced):
FRONT PAGE | EDITOR'S DESK | CIVIC | CULTURE | BUSINESS | OPINION | SPORTS | LETTERS
Do NOT use rich variants (CIVIC AFFAIRS, CITY LIFE, ACCOUNTABILITY, QUICK TAKE, FEATURES) —
the parser ignores them and the section's articles silently drop from review.
FRONT PAGE — [best story]
------------------------------------------------------------
EDITOR'S DESK — Mags, 150-250 words
------------------------------------------------------------
CIVIC — [stories tagged civic by sift]
------------------------------------------------------------
CULTURE — [stories tagged culture by sift]
------------------------------------------------------------
BUSINESS — [stories tagged business by sift]
------------------------------------------------------------
SPORTS — [stories tagged sports by sift]
------------------------------------------------------------
OPINION — [editorial / op-ed pieces if any]
------------------------------------------------------------
LETTERS — [letters to the editor slate]
------------------------------------------------------------
ARTICLE TABLE
NAMES INDEX
BUSINESSES NAMED
CITIZEN USAGE LOG
STORYLINES UPDATED
COMING NEXT EDITION
END EDITION
Format-contract footer sections (REQUIRED — see [[../../docs/media/EDITION_FORMAT_TEMPLATE]] §Footer sections for canonical literal shape). S227 correction (closes G-W42 / partial G-W43): canonical order shipped in E94 is NAMES INDEX → BUSINESSES NAMED → ARTICLE TABLE → CITIZEN USAGE LOG. The template carries this order. Earlier skill text said "after ARTICLE TABLE" for the strict three; that was inaccurate.
NAMES INDEX — one row per named entity. Strict pipe-format:
POP-NNNNN | Full Name | Role/Title for Sim_Ledger citizens
CUL-NNNNNNN | Name | Role for cultural-only entities (musicians, public figures from wd-cultural)
FAITH-NEW | Org Name | Faith Org | Neighborhood for new faith orgs
Name — Role (em-dash) for citizens not yet in canon (ingester promotes to POP-pending row)
BUSINESSES NAMED — one row per named business. Strict pipe-format:
BIZ-NNNNN | Name | Sector | Neighborhood for existing businesses
NEW | Name | Sector | Neighborhood for new businesses (sector/neighborhood may be blank)
CITIZEN USAGE LOG — editorial categorized prose (human-readable; partial parse for ingest).
Subsections like CIVIC / GOVERNMENT, CITIZENS QUOTED OR PROFILED, LETTERS WRITERS,
plus (NEW CANON THIS CYCLE) sub-header. The strict sections above are derived from this
section by scripts/emitFormatContractSections.js — the (NEW CANON THIS CYCLE) subsection
is the canonical input for biz/faith extraction. Sub-format rules:
Each line starts with - (bullet)
Entity name precedes em-dash delimiter (—)
After em-dash: comma-separated metadata
For businesses: include BIZ-pending OR confirmed canon marker, sector, neighborhood
For faith orgs: include confirmed canon, neighborhood, tradition, leader name, congregation size, founding year
For citizens: include citizen + occupation + neighborhood + POP-pending
Wrong-shape entries are silently dropped by emitFormatContractSections.js today (G-W43 — engine-sheet repair pending). Until repair lands, verify emit script output matches author intent before continuing past Step 3a; if NAMES INDEX comes back as bullet-prose or BUSINESSES NAMED writes zero with biz mentions present, hand-restore from template.
Step 3a: Derive strict format-contract sections. After writing CITIZEN USAGE LOG (rich prose),
run the helper to derive the strict NAMES INDEX + BUSINESSES NAMED sections and inject them into
the file before ARTICLE TABLE:
Idempotent — re-running replaces existing strict sections. Fails loud if CITIZEN USAGE LOG is
missing or empty. Without this step, ingestPublishedEntities.js silently no-ops in /post-publish
Step 5 (G-W19/G-P6/G-P8/G-P9 — three new citizens + Atlas Bay Architects + Greater Hope Pentecostal
silently dropped from C93 intake). Verification gate verifyNamesIndexParse.js --strict in
/post-publish Step 5 enforces NAMES INDEX presence — publish blocked if absent.
Section omission rule: If a section has no story this cycle, its label and divider don't appear.
The parser handles missing sections cleanly; never emit an empty section header. Format-contract
footer sections are NOT subject to omission — NAMES INDEX must always appear (even if empty body).
Save to editions/cycle_pulse_edition_{XX}.txt.
Compile complete; file ready at canonical path. Continue to Step 3.25. (S227 correction, closes G-W45: Mike's canon-review point is Step 5, not Step 3. Step 3 produces a complete file; Step 5 is the USER APPROVAL GATE for canon-verify before Mara/ingest exposure.)
Seal the raw generation (G-W4, S256) — mechanical, final action of Step 3. The instant compile completes, before ANY review lane or operator read, hash the raw edition + reporter articles:
This is the measurement-integrity baseline. The review lanes exist to measure where the SKILL fails on RAW output — that signal is destroyed if the edition/articles are hand-edited before the lanes run (C96 G-W4: 12 manual edits + renames pre-lane → lanes measured a laundered hybrid, taught research-build nothing). After this seal, no edits to the edition .txt or any reporter .md are permitted except those routed from a lane REVISE verdict (Step 4 / 4.1). Every sanctioned REVISE round re-seals ONLY the files it changed, by name (--files); an un-sanctioned operator pre-edit is never blessed, so it shows as a hash mismatch at the next verify. Verify runs at first-lane entry (Step 3.25, primary) and again at the Arbiter (Step 5.5, backstop); contamination at either checkpoint is sticky — a later clean verify can't erase it. Flag-not-block (S256): a detected pre-edit doesn't halt the pipeline, but the Final Arbiter marks the run measurementIntegrity: contaminated.
Update production log with compile details (front page, total articles, edition path).
QUICK TAKES handling (S227, closes G-W44). /sift slates sometimes carry "Quick Takes" (QT) entries; the section allowlist has no QUICK TAKES section. Two routes:
(a) Fold into existing section — if a QT has a topical home, place it as a standalone short piece inside CIVIC / CULTURE / etc. It reads as a short article and therefore takes the same canonical block as any article — all four lines required (RB-3, C99 G-W11):
### {headline}
By {Reporter} | Bay Tribune {Section}
---
{body}
The --- separator after the byline is REQUIRED, not optional. lib/editionParser.js counts a byline-bearing article by the ^By\s+ line and binds it to the headline; the --- is what closes the byline block so the body registers. Omitting it (C99: the first QT compile) makes the parser undercount bylines vs ARTICLE TABLE rows and throws the fail-loud bylineArticles.length !== rows.length guard. Headline is ### (H3), never # (H1) — same parser contract as Step 3 ¶Parser contract.
(b) Drop from the cycle — if no reporter source material exists or the QT lacks anchor (no civic-office statement for an Okoro mini-take, no DJ photo for a walk feature), cut it from compile. Don't carry through to a half-written placeholder.
Default: (a) when source material exists, (b) when it doesn't. Editor's Desk may also absorb spine framing the QT was carrying — that's the third path. The intent of QT was lightweight texture, not architectural separation; the section allowlist deliberately omits a QT section so texture lands inside topical sections.
FIRST — measurement-integrity verify (G-W4, S256). This is the non-negotiable checkpoint. Before any lane consumes the files, confirm the raw generation sealed at compile is intact:
No legit REVISE has happened yet at this point, so a CLEAN result is expected. A contaminated result means the edition or a reporter article was hand-edited between compile and now with no lane REVISE verdict behind it — the exact G-W4 failure. Flag-not-block: the pipeline continues, but record it in the production log and know that this cycle's lane findings now describe a hand-edited hybrid, not raw generation. The right recovery is to re-run the lanes against the raw output; if that's not possible, the run is a contaminated measurement (the Final Arbiter will stamp it).
Three deterministic pre-review steps, all run in parallel after compile:
# 1. Adversarial review — devil's advocate probe (5 lenses)# Run /adversarial-review or do manually per the skill# Outputs: inline findings (contradictions, unsourced claims, narrative gaps)# 2. Tier classification — assigns A/B/C to each article
node scripts/tierClassifier.js {XX}
# Outputs: output/tier_assignments_c{XX}.json# 3. Reward hacking scan — checks for evaluator gaming
node scripts/rewardHackingScanner.js {XX}
# Outputs: output/reward_hacking_scan_c{XX}.json
Read the tier assignments — they control how much review each article gets downstream:
Tier A (full review): front page, Tier-1 citizens, engine-flagged ailments, contested civic stories → all three reviewer lanes + capability + two-pass hallucination
Tier B (editor pass): neighborhood features, routine council, sports recaps → Rhea + cycle-review only
Tier C (automated only): letters, baseline briefs, box-score equivalents → Rhea regex + anomaly flag only
If the reward hacking scanner flags HIGH severity (rubric gaming or rubric execution detected), investigate before proceeding — a reporter may be optimizing for the rubric instead of writing journalism.
If adversarial review recommends HALT, fix the findings before proceeding to capability review.
Update production log with tier counts (A/B/C), reward hacking scan results, adversarial review recommendation.
Step 3.5: Capability Review (Phase 39.1, S146)
Run the editorial capability gate before validation. Catches structural editorial gaps that Rhea + Mara don't check — the front page missing the highest-severity engine ailment, citizen names that don't resolve to canon, engine metrics leaking into journalism. The Varek anti-example (E91 front-paged NBA expansion while Temescal ran four cycles uncovered) is exactly what this gate makes structurally impossible.
node scripts/capabilityReviewer.js {XX}
Or invoke /capability-review for the wrapped flow with the markdown summary.
Read output/capability_review_c{XX}.json. Show Mike the summary.
Blocking failures halt this step. For each, choose with Mike: (a) fix and re-run (route back to relevant reporter or /sift), (b) override and proceed (logs the failure for next sift), or (c) defer publish entirely. Advisory failures ship with a flag in the production log and don't gate.
Editor override propagation to Final Arbiter (S215, closes G-W28). When option (b) is taken — Mike approves an override of a blocking failure — append the override to output/capability_review_c{XX}.json as a structured field:
"editorOverride":[{"ruleId":"<failing rule>","approver":"Mike","approvedAt":"<ISO>","reason":"<one-line — e.g., 'phase-advanced this cycle so highest-severity-ailment rule doesn't apply'>"}]
Final Arbiter (Step 5.5) MUST read editorOverride[] and demote any overridden rule from blocking → advisory before computing its verdict. Without this propagation, Mike has to override the same rule twice (once at Step 3.5 capability, once again at Step 5.5 Arbiter). C93 hit this on front-page-leads-with-highest-severity-ailment for the INIT-005 phase-advance case.
Capability rule "phase-advanced this cycle" exception (S215, closes G-W21). The front-page-leads-with-highest-severity-ailment rule should NOT fire when the highest-severity-ailment initiative advanced phase this cycle (per Initiative_Tracker C{XX} writeback). A stuck-and-just-unstuck story is editorially different from a stuck-and-still-stuck story; the Varek anti-example that motivated this rule was about ignoring continuing crises, not about ignoring resolved ones. Until the rule is updated in scripts/capability-reviewer/ (engine-sheet pipeline.19 work), document override use cases in production log + editorOverride field. Default treatment when override applied: advisory-only.
Anonymous-source convention (S215, closes G-W10). When a reporter beat is accountability (freelance-firebrand) or investigative, anonymous sources ("an East Oakland resident waiting on signature 109-191") are valid craft — Jax Caldera's signature move and load-bearing for accountability journalism. The capability + Mara graders need to recognize this class: anonymous-source is allowed for firebrand and explicit-accountability beats; flagged elsewhere. Until the capability rubric is updated (engine-sheet pipeline.19 work), document anonymous-source use in production log under capability findings; do NOT treat as fabrication.
Initiative budget line-item recognition (S215, closes G-W23).validateEdition.js flags any dollar figure against the Initiative_Tracker total. A $4.5M Atlas Bay Architects contract is a line-item against a $45M Health Center total budget — not a contradiction (~10% architect contract is normal). The validator's warning text should qualify as "possible budget mismatch — verify line-item vs total"; until that lands (engine-sheet pipeline.19 work), accept architect/contractor sub-budgets without revision when they round to a reasonable fraction of the total project budget. Document in production log if a line-item flag is overridden.
Update production log with capability review counts (passed/total, blocking, advisory) and any overrides taken.
Step 4: Validation + Rhea (Sourcing Lane)
Reviewer-lane discipline — run the documented lanes, nothing else (RB-7, C99 G-W12). From here through Step 5.5, the review is the fixed lane sequence: validateEdition → Rhea → cycle-review → Mara → capability → Final Arbiter. Run each lane and capture every issue it surfaces into the cycle gap log (G-W) as you go.* Do NOT substitute an out-of-band reviewer — in particular, do not call the advisor tool here or narrate its reasoning back into chat (that burns Mike's tokens for a check the lanes already do), and do not freelance slate/publish decisions outside the lanes. The advisor tool is for pre-substantive-work checkpoints in an open session, never mid-skill. Mike, C99, explicit: "if you aren't running the skill and reporting the gaps this is a complete waste." If a lane is genuinely blocked, file the gap and surface the infra gap — don't write the lane's verdict from the editor seat.
Fix CRITICALs. Then launch Rhea as the Sourcing Lane (Phase 39.2, weight 0.3).
Rhea has scoped Bash access — dashboard API (localhost:3001), Supermemory (bay-tribune + world-data), world summary. She's a real verifier with live data access. After Phase 39.2 she produces output/rhea_report_c{XX}.json in the reviewer-lane schema.
# After Rhea writes her JSON, validate + emit .txt companion:
node scripts/rheaJsonReport.js {XX}
verdict PASS → proceed
verdict REVISE → fix and rerun, max 2 rounds
verdict FAIL → halt, route back to desks
Re-seal after a sanctioned REVISE fix (G-W4, S256) — mechanical. A Rhea REVISE that edits the edition or a reporter article is a sanctioned change. The final action of each REVISE round, after applying the fix, is to re-seal — naming ONLY the files the fix touched with --files:
--files is required for a revise:* re-seal (the command errors without it). This is the load-bearing rule: a blanket re-seal would re-bless every currently-divergent file — including an un-sanctioned operator pre-edit that happens to precede this REVISE — laundering the exact contamination the gate catches. By naming only the lane-fixed files, any un-named pre-edit stays off-seal and is caught at the next verify. Skipping the re-seal entirely false-flags the legit fix; a blanket re-seal launders. Name the files.
Update production log with validation results and Rhea's lane score.
Step 4.1: Cycle-Review (Reasoning Lane)
Run /cycle-review as the Reasoning Lane (Phase 39.4, weight 0.5). Produces output/cycle_review_c{XX}.json.
Does NOT re-check names, votes, stats, engine language — those belong to Rhea and capability.
verdict PASS/REVISE/FAIL same semantics as Rhea.
Re-seal after a sanctioned REVISE fix (G-W4, S256): if a cycle-review REVISE edits the edition/articles, re-seal naming only the touched files — node scripts/editionSeal.js --seal --cycle {XX} --reason revise:cycle-review-r{N} --files <touched-paths> — as the final action of the round (same --files-required discipline as Step 4).
Update production log with cycle-review lane score.
Step 5: Mara Audit (Result Validity Lane, External)
USER APPROVAL GATE — Mike reviews the compiled edition for canon before uploading to Drive for Mara (S227, closes G-W45). Show Mike the compiled edition + the lane JSONs from Steps 4/4.1/3.5 (validation + Rhea + cycle-review + capability). Canon-verify happens here: this is where the edition crosses from internal-pipeline state to external-ingest exposure (Drive + Mara on claude.ai). Mike's check at Step 3 was deferred from prior skill text — that placement was wrong; review-for-canon lands at Step 5, publish-approval lands at Step 5.5 / 6, the two are distinct gates.
After Mike says go:
Mara is on claude.ai — Result Validity Lane (Phase 39.5, weight 0.2).
Upload edition + sift brief + engine review to Drive: node scripts/saveToDrive.js editions/cycle_pulse_edition_{XX}.txt mara
Tell Mike the edition is ready for Mara
Mike takes it to Mara on claude.ai
Mara produces a markdown audit with the structured top per PHASE_39_PLAN §16.3
Mike saves it to output/mara_audit_c{XX}.md
# Parse Mara's markdown into lane JSON:
node scripts/maraJsonReport.js {XX}
STOP. Wait for Mara.
Mara prose vs structured-top expectation (S215, closes G-W26). Mara's value is editorial judgment expressed in her own voice; the pipeline needs the structured top for lane-JSON parsing. Mara's claude.ai system prompt (per PHASE_39_PLAN §16.3) instructs her to emit the structured-top block as the first ~10 lines, followed by free-form prose. If a given audit lands as prose-only without the structured top, run scripts/maraJsonReport.js {XX} --extract-from-prose (when available — engine-sheet pipeline.19 follow-up) OR manually construct the lane JSON from her required-fixes list while preserving the prose in mara_audit_c{XX}.md. Don't force-template Mara's voice — adapt the parser, not the writer.
Reviewer-lane canonical order (S215, closes G-W27). Canonical pipeline order is Step 4 (validation + Rhea) → Step 4.1 (cycle-review) → Step 5 (Mara) → Step 5.5 (Final Arbiter). Mara delivering out-of-order (between Step 3 and Step 3.25, e.g.) doesn't break the pipeline but masks lane signal — the reviewer-lane scripts run AFTER editor has already applied Mara's fixes, so the lanes don't independently catch what Mara already caught. If Mike forwards Mara's audit early in a future cycle, ask for canonical-order delivery to preserve lane-attribution. Out-of-order is allowed editorially; just acknowledge the trade-off in production log.
Update production log with Mara's lane score and any editorial notes from her prose.
Step 5.5: Final Arbiter (Phase 39.7)
Backstop measurement-integrity verify FIRST (G-W4, S256). The Step 3.25 verify is the primary checkpoint; this backstop catches any un-sanctioned edit made after the lanes began (e.g., between Rhea and Mara). Run it before the Arbiter so the Arbiter reads a current result:
finalArbiter.js reads every per-checkpoint verify file (output/edition_seal_verify_c{XX}_*.json) and ORs them — contamination at any checkpoint sticks — to stamp the verdict with measurementIntegrity (clean / contaminated / unsealed). Flag-not-block: a contaminated result does NOT change the A/B verdict or the publish recommendation — it adds a measurement-integrity blame entry naming the pre-edited files and the checkpoint that caught each, so research-build knows this cycle's lane findings describe a hand-edited hybrid (G-W4). If unsealed (the seal/verify steps were skipped), the Arbiter records that the gate wasn't exercised rather than asserting clean.
Deterministic computation — reads the four lane JSONs (reasoning, sourcing, result-validity, capability), applies the 0.5/0.3/0.2 weights, enforces the capability gate as a hard block, emits output/final_arbiter_c{XX}.json with a single verdict (A/B), blame attribution, and a publish recommendation:
PROCEED-WITH-NOTES — verdict A, weighted score 0.60–0.75, capability gate passed. Log items for next cycle's briefing.
HALT — verdict B. Exit code 1. Do NOT proceed to Step 6.
The Arbiter is the publication gate — Step 6 runs only if the recommendation is PROCEED or PROCEED-WITH-NOTES.
USER APPROVAL GATE — Mike reviews the Arbiter JSON and says publish or doesn't.
Update production log with Arbiter verdict, weighted score, and blame attribution.
Step 6: Publish
# Save edition to Drive
node scripts/saveToDrive.js editions/cycle_pulse_edition_{XX}.txt edition
Text ingest to bay-tribune is /post-publish's job (S215, closes G-P1). This step previously ran node scripts/ingestEdition.js editions/cycle_pulse_edition_{XX}.txt at close. That call has been removed: /post-publish Step 1b is the single canonical home for edition text ingest. Running ingest here AND in /post-publish created duplicate text records in bay-tribune (different doc IDs, same content), polluting future sift queries. Wiki ingest (per-entity records, different shape, no duplication risk) also lives in /post-publish (Step 1a), not here.
If you need an early text ingest for some rare reason (e.g., emergency canon backfill), explicitly note it in the production log and skip /post-publish Step 1b.
Update production log with publish status — Drive ID + path. Inline doc IDs from /post-publish ingest land in the production log when that skill runs (not here):
All ingest now happens in /post-publish — text ingest (Step 1b), wiki ingest (Step 1a), citizen cards (Step 2a), world summary (Step 2c), coverage ratings (Step 4), grading (Steps 6-8). Doc IDs land in the production log when /post-publish runs.
Handoff
After publish, two separate skills pick up:
Skill
What it does
When
/edition-print
Photos (DJ Hartley), PDF, Drive upload of print assets
Continued from sift — reporter results, review, compile, validation, Mara, publish added
Steps 1-6
Legacy Reference
These elements were part of the old write-edition (pre-S144) and are now handled by other skills:
World summary build → /build-world-summary
Story picks and sifting → /sift
Citizen verification → /sift Step 4
Angle brief writing → /sift Step 5
Production log creation → /sift Step 2
Desk packet building → legacy scripts preserved, not in pipeline
Voice workspace building → /city-hall-prep
Gap log (S212 — see [[plans/GAP_LOG_TEMPLATE]])
At skill close, capture friction observed during edition write as a gap log. /write-edition is the heaviest skill at the media generator terminal; sidecar gap logs catch inefficiency the skill couldn't catch while running.
Destination (RB-1/RB-2 — one-true gap log): append a leg to the cycle's single gap log output/production_log_run_cycle_c{XX}_gaps.md (the file the engine cycle audit opens each cycle). Do not write a separate _write_gaps.md sidecar — that split convention is retired. Open the leg with the fixed header the gate greps for:
## LEG: /write-edition (G-W)
Then the G-W entries below it — or No gaps this run. on a clean run. The header must be present either way.
Gap prefix:G-W* (e.g., G-W1, G-W19, G-W22).
Common categories for /write-edition gaps:
doc-drift (skill compile template vs parser/validator constants — high-recurring class, ~30% of accumulated gaps)
reviewer-handoff (Mara format expectations, Final Arbiter override propagation)
Discipline: write the gap log even on clean runs. File a ROLLOUT row in pipeline.<n> pointing at the gap log per ADR-0005 §How to add work. Promote individual HIGH gaps as bandwidth allows. The S195 G-W16 meta-pattern (HIGHs sit on shelf and compound across cycles) makes promotion-cadence load-bearing.
Close gate (mechanical — RB-1, G-S1). The final action of /write-edition is:
It exits non-zero until the ## LEG: /write-edition (G-W) leg exists in the cycle gap log; skill close is defined as this exit 0. A Stop-hook backstop (gapLogGate.js --stop-gate) blocks session close for the same reason if this step is skipped — the G-S1 failure was the operator skipping a written instruction, so the enforcement is mechanical, not prose. Deliberate bypass: GAPLOG_GATE_OFF=1.
Where This Sits
After /sift. Before /edition-print and /post-publish.
