SWITCHYARD

Editorial workflow automation

Content that routes itself.

Switchyard takes an idea from the editorial calendar through structure, SEO, brand voice and compliance review — then stops for a human exactly where you decide it should. Every change versioned. Every decision logged.

standard editorial · live 3 runs in flight
AI agent waiting on a human approved rewound to an earlier step

Runs advance

in the background

No one clicks “run agent”. Queued jobs move every run forward.

Every edit

is a new version

Versions are immutable. Nothing is ever overwritten.

Workflows are

data, not code

Build and reorder pipelines in the UI. No deploy.

Output is

publish-ready

HTML, DOCX, PDF and assets in a single archive.

How it works

An idea goes in. A publish-ready package comes out.

The steps between are yours to define. This is the seeded “Standard Editorial” pipeline — five AI passes, two human gates, one packaging step.

  1. 01

    Capture the idea

    Log an idea by hand, or let Switchyard suggest a batch from current industry trends. Pick the workflow it should travel — that choice is made here, at intake.

  2. 02

    Place it on the calendar

    Schedule against campaign themes and date ranges. Drag to reschedule; the article and its run follow.

  3. 03

    Write, or generate

    A WYSIWYG editor with AI drafting, selection-scoped rewriting, and image generation built in.

  4. 04

    Agents do their passes

    Structure, SEO and brand voice rewrite. Marketing and medical review flag findings and can block. Each pass produces a version.

  5. 05

    Humans decide

    The run halts at your review and approval gates, assigns the right people, and waits — with an SLA clock running.

  6. 06

    Package and publish

    Approved content is built into a downloadable archive: HTML, DOCX, PDF, assets and metadata.

The agents

Five editors that never miss a pass.

Each agent has its own prompt, model, effort level and output schema — all editable by an administrator without a deploy. Two modes:

Revise
Rewrites the article and produces a new version, with a summary and a rationale for every change it made.
Review
Doesn’t touch the copy. Records findings with a severity, and can hard-block the run when configured to.

Test any agent against a real article version before it goes live — see the raw response, the token usage and the cost, without writing a version.

Structure Editor

revise

Reorders sections, splits dense blocks, enforces a scannable shape.

SEO Editor

revise

Headings, meta description, internal link anchors, keyword coverage.

Brand Voice Editor

revise

Applies your brand kit — tone rules, banned phrases, audience register.

Marketing Reviewer

review

Flags unsupported claims and off-strategy positioning.

Medical Guidelines Reviewer

review

Checks copy against current clinical guidance. Can block.

Your own

custom

Add an agent with its own prompt and schema, and drop it anywhere in a pipeline.

Versioning & attribution

Know exactly what the AI changed — and why.

Every agent pass and every human edit writes an immutable version. The diff view puts them side by side and attributes each change to the agent or person who made it.

  • Structured rationale. Agents return a category, a description and the reason for each change — not just new copy.
  • Cost per change. Tokens, latency and spend recorded against the version that resulted.
  • Restore without loss. Rolling back writes a new version. History is never rewritten.
spring-wellness-guide v5

Human in the loop

Automation up to the point where judgement matters.

You choose where a run stops. It assigns the right people, starts an SLA clock, and waits — surfacing everything the agents already flagged so a reviewer isn’t starting cold.

Assign by role or person

Route a gate to every editor, or to named reviewers. Reassign mid-run without losing the round.

Multiple rounds

Request changes and the run rewinds to the step you nominated. Agents re-run; the next review is round two.

Approval quorums

Require one approval or several. The gate shows who has decided and who hasn’t.

Recorded overrides

An authorised approver can override a quorum — but only with a written reason, stored in the audit trail.

Full control while it’s moving

Pause a run, cancel it, re-run a single step, or rewind to any earlier position — which cancels everything after it and starts a fresh attempt. Nothing is destroyed: earlier attempts stay on the record.

Pause Resume Cancel Re-run step Rewind to step Reassign

Publishing

Approved content, packaged for wherever it’s going.

A packaging step builds a single archive from one specific version — so what ships is exactly what was approved, not whatever the article looks like today.

Direct CMS publishing, starting with WordPress and scheduled against the content calendar, is next on the roadmap.

index.html

Self-contained, semantic, with SEO meta and relative asset paths.

article.docx

Headings, lists, tables and images mapped to real Word styles.

article.pdf

Print-ready, assets embedded.

metadata.json

Title, SEO fields, campaign, author, version and a run summary.

Roles & permissions

Everyone sees their own work first.

A personal dashboard shows what you owe. A fleet view shows every run in motion, where each one is stuck, and what it has cost.

Role Can
Viewer Read articles, ideas, the calendar and dashboards.
Editor Everything above, plus create and edit content, use the editor, start runs and submit reviews.
Approver Everything above, plus grant or deny approvals at gates.
Publisher Everything above, plus build packages and push to publish targets.
Administrator Everything, plus manage users, workflows, agents and settings — and override approvals.

Stop pushing content through the pipeline.

Let it route itself, and step in only where your judgement is the thing that matters.