Suede Newsroom

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:

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.

Founder-led mode. When the recurring source is a founder interview, voice note, or operator debrief, or when personal and company accounts need different jobs, read references/founder-content-loop.md before Step 1. It adds an intake contract and account lanes to this pipeline; the six role contracts, evidence gates, collision checks, approval boundary, and Keep/Test/Stop thresholds below remain unchanged.

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.

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

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:

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.