Suede Newsroom

Six chatbots pointed at the same brand are not a content team. A newsroom is the loop that connects them, and every arrow in it is a handoff with a contract:

idea -> research -> angle -> flagship -> distribution -> review -> performance -> playbooks

Break one arrow and the failure is predictable, not mysterious:

Broken arrow	What ships instead
research never reaches the strategist	a generic angle any competitor could have written
the writer never sees the evidence packet	claims that drift one step past what the source says
distribution starts from the finished draft	five surfaces carrying the same shortened article
performance never reaches the playbooks	the same mistake repeated at a higher cadence

This skill owns the pipeline: who decides what, what travels between them, and how one idea becomes several assets that argue different things. Read .agents/product-marketing.md first if it exists — it holds product, audience, positioning, and proof, and nothing here restates it.

Step 1 — Put the campaign in a file, not a conversation

One record per campaign at .agents/newsroom/<slug>.md. It is the only thing that has to survive between roles, sessions, and restarts.

A conversation between roles has two failure modes a file does not: context dies on restart, and the next role has to infer which sentences were load-bearing. The record gives each role a predictable input and lets a human audit a campaign without reopening six sessions.

Keep it narrow — decisions and the evidence behind them. Drafts live beside it, not in it. The template and per-stage completion rules are in references/campaign-record.md.

Step 2 — Give each role one decision to own

The fastest way to ruin a multi-agent content team is to hand every role the same instruction: make great content. Each role gets one decision, one artifact, and a named point where it stops.

Role	Owns	Returns	Hands to
Signal scout	whether an idea has a reason to exist now	scored candidates	researcher
Researcher	what is verified versus inferred	evidence packet	strategist
Strategist	the single editorial angle	angle brief	writer
Writer	the deepest version of the idea	flagship piece	distribution
Distribution	what each surface argues	asset set with entryways	editor
Editor	whether the package holds together	approve, revise, or reject	human

Every contract fills five fields — owns, reads, returns, must not, done when. Full text for all six, with the must not and done when clauses, is in references/role-contracts.md. Copy them in verbatim; a paraphrased must not is how a writer starts picking angles.

Thresholds that make the contracts checkable:

• Scout returns at most three candidates, each with a decay window, and names a rejection reason for every candidate it discarded. A run that approves everything it found did not filter anything.
• Researcher returns three to seven verified claims, a URL for every claim with consequence, and an explicit what the sources do not prove list.
• Strategist returns exactly one recommended angle plus at least two named rejected directions and why each is weaker. A menu pushes the decision back to the human, which is the decision this role exists to make.
• Distribution returns one entryway per asset and no entryway twice.
• Editor reviews the assets together in one pass, never one at a time.

Step 3 — Gate every handoff

A stage is complete when every field it owns is filled and readable by the next role — not drafted, not discussed. Written into the record.

Return-to-sender. When a field the current stage needs is empty, the role sends the record back and names the field. It does not fill the gap with a plausible assumption. A pipeline whose roles patch each other's gaps produces confident output with no traceable source, which is worse than a stalled campaign because nothing signals the problem.

Halt format. Stop. Name the empty field and the stage that owns it in one line. Offer the human these options and wait: (1) supply the field now, (2) send the record back one stage for another attempt, (3) narrow the campaign until the field is unnecessary, or (4) kill it. Do not advance the record, and do not present a partial package as complete.

Two returns on the same field means the pipeline cannot supply it. Escalate rather than attempting a third pass — a third attempt on the same gap is where fabrication starts.

Step 4 — Distribute by argument, not by format

This is the stage that fails most often, and it fails quietly.

An article does not become a short post by losing 1,500 words, and a newsletter does not become a carousel because its paragraphs were laid onto slides. Both moves produce assets carrying the same argument at different lengths — which is why a week of them reads like one post repeated.

Work from the angle brief, not the finished draft. A draft carries one argument, so anything derived from it inherits that argument. Reading the draft first is the mechanism that produces shortened articles.

Assign one entryway per asset

Entryway	The argument it makes	Needs
Proof	this is real, and here is the artifact	something a stranger can check
Mechanism	here is why it works	a causal chain, not a description
Workflow	you can run this today	ordered steps someone could follow
Risk	here is what breaks, and when	a named failure mode with its trigger
Result	here is what it produced	an outcome plus its conditions
Critique	the common approach is worse than it looks	the real default, and how it fails
Compression	here is the whole shape	the idea reduced to one line or image

One entryway per asset. An asset carrying two is a shortened article wearing a label. An entryway the source cannot supply is unavailable — report it and say what evidence would unlock it, rather than inventing an artifact to fill it.

Worked examples, surface fit, and the failure mode each entryway falls into are in references/entryways.md.

Score each asset for standing alone

Three checks, one point each, run against the whole set:

Check	Passes when
New information	it carries a claim, step, number, or object no other asset carries
Second reader	someone who already read the flagship still gets something; "more at the link" is not something
Removal	deleting it costs the campaign a specific, nameable thing

3 ships. 2 goes back for revision, naming the failed check. 0–1 gets cut, and say what the set loses — usually nothing.

Run the collision gate

Reduce each asset to one sentence first; comparing full drafts hides collisions.

1. Claim collision — two assets whose central claim is the same sentence.
2. Opening collision — the same story, statistic, or line opens two assets. The most common, and the most visible to anyone following on two surfaces.
3. Entryway collision — the same entryway used twice.

Halt format. Stop. Name the two assets, the collision type, and the shared sentence. Offer: (1) reassign one to an unused entryway, (2) cut the weaker and say what the set loses, or (3) record it as a deliberate repeat with a reason. Wait. Do not resolve a collision by rewording — the claim underneath is what collides.

Fewer surfaces than entryways is normal. Three assets with three distinct entryways beats seven assets with two.

Step 5 — Keep the human at the editorial boundary

The first version prepares everything and publishes nothing. The human approves the angle, the flagship draft, every consequential factual claim, every public post, any change to voice or offer, and any performance lesson before it becomes a standing rule.

Each approval and rejection is a labelled example of the human's judgment. When the same decision comes back the same way three campaigns running, move it into the editor's review rules so it is applied before the human sees the package.

Move publishing authority only when both hold: the queue has produced no revisions for three consecutive campaigns, and a wrong post is recoverable by deleting it. Autonomy should remove repeated decisions, not remove taste.

Step 6 — Make performance rewrite the playbooks

Reporting numbers is not learning. After each campaign log the signal, angle, entryway, surface, reach, meaningful engagement, and the goal-linked action. Then have the editor return three lists:

List	Bar to qualify
Keep	held across two or more campaigns and still matches strategy
Test	one promising result — a hypothesis, run it again
Stop	failed twice, duplicated another asset, or cost more than it returned

Every proposal names the campaigns supporting it. One strong result is a Test, never a Keep. The human approves before any rule changes. Without this arrow, every campaign starts from the same generic prompt forever.

Log by entryway, not only by surface — over a few campaigns that reveals which door the audience actually walks through, which is more useful than which platform performed.

Running it on the host you have

Six roles is the model, not a hosting requirement.

Host	Roles map to	What breaks
One session	one pass per role, record reread between	role bleed — it edits its own draft
Subagents	one per role, dispatched in stage order	cost, if each drags full context
Separate profiles	one per role, isolated memory	drift, if they coordinate by chat
One person	one sitting per role, in order	nothing; this is the honest first week

Two rules survive every host. Conversation coordinates; files carry state. And isolate role memory where the host allows it — one shared memory turns six specialists back into one generalist that has read everything and distinguishes nothing.

Start smaller than six roles

Stand up scout, researcher, and editor first, and run three real ideas through them. Add the strategist once two of three evidence packets need no rework, the writer once the briefs constrain a draft without further questions, and distribution once one flagship has shipped and been measured.

The first useful version returns one researched opportunity, one angle brief, and one approval-ready post with a source attached. That is already more than most content operations produce.

Anti-patterns

• One shared memory for all six roles — specialists collapse into one generalist with opinions about everything.
• Prose handoffs — the next role guesses which sentences were load-bearing.
• Roles that fill each other's gaps — confident output, no traceable source.
• Splitting the draft instead of the argument — produces length variants.
• One hook rewritten seven ways — the claim underneath is identical, so the reader sees the repeat even when the wording differs.
• Adding the writer first — the most visible role, and the one that produces least value without an evidence packet in front of it.
• A performance log nobody converts to rules — analytics theatre.

Banned vocabulary

Avoid "fully autonomous media company", "content on autopilot", "replaces a marketing team", "repurpose", "atomize", and "one piece into fifty posts". The first three hide the approval steps; the last three describe reformatting, and naming the work that way produces the work. Say which argument each asset makes.

Boundaries

• Do not publish, schedule, or send anything. The pipeline prepares packages; the human releases them.
• Do not record a claim without the source that supports it, and do not upgrade an inference to a verified claim at any stage.
• Do not introduce a claim, number, or quote during distribution that the evidence packet does not carry.
• Do not invent a proof object, result, or failure mode to fill an entryway the source does not support. Report the entryway as unavailable.
• Do not present an unscored asset set as ready, and do not resolve a collision by rewording rather than reassigning.
• Do not report a stage complete without the record fields that stage owns being filled, and do not fabricate reach or engagement. An unmeasured campaign is logged as unmeasured.

Routing

• Which topics, pillars, or clusters the pipeline should work on -> use suede-content-strategy. That skill decides what it operates on; this one decides how it moves.
• Platform mix, cadence, calendars, listening, and pulling atoms out of long-form -> use suede-social. Instagram-specific programs -> use suede-instagram-growth.
• Cadence, state, idempotency, and stop conditions for a scheduled stage -> use suede-marketing-loops.
• Writing the flagship for a high-stakes public surface -> use suede-ship-copy. A single shorter surface -> use suede-copy.
• A video moment bridged to one long-form guide -> use suede-clip-to-guide.
• Stripping AI writing patterns from the drafted assets -> use suede-deslop.
• Measurement plumbing behind the performance log -> use suede-analytics.
• Paid variants of these angles -> use suede-ad-creative.
• Agents editing code in parallel with file ownership and rollback -> use suede-agent-teams. This skill covers editorial lanes, not repository lanes.
• From suede-content-strategy and suede-social: route role contracts, the handoff record, and entryway assignment back here.