Publishing the same technical article everywhere sounds like a copy-and-paste job. It stops feeling simple the moment you actually do it.
One editor wants Markdown. Another quietly rewrites your headings. A third expects a cover image with a very specific aspect ratio. Links acquire tracking parameters, code blocks lose their language labels, and the platform that said “submitted” may still leave the post sitting in a draft queue. By the time the article is live, the author has spent more energy babysitting browser tabs than improving the argument.
That is the real pain in content automation. Writing is only one part of the system. Delivery is a distributed workflow, and treating it like a clipboard macro is why so many publishing automations become fragile.
A publish button is not a workflow
The tempting design is a straight line:
brief -> draft -> image -> publish everywhere
It looks tidy in a diagram, but it hides the decisions that matter. Which version went to which site? Did the destination create a draft or make the post public? Was an image uploaded successfully but the article rejected? If a request timed out, is it safe to retry, or will that create a duplicate?
A more honest model keeps an artifact and a delivery record for every destination:
{
"article_id": "release-notes-2026-10",
"revision": "7f3c9a1",
"destination": "devto",
"requested_state": "published",
"remote_id": "4979999",
"remote_url": "https://dev.to/example/...",
"actual_state": "published",
"attempts": 1,
"last_error": null
}
That little bit of bookkeeping changes the nature of the automation. A vague “done” becomes a result you can inspect. The source revision tells you exactly what was sent. The remote identifier gives the next run something to update instead of blindly creating another post. Most importantly, requested_state and actual_state are separate, because asking a platform to publish something is not proof that it is public.
Keep one source, but do not force one rendering
Developers naturally want a single canonical Markdown file, and that is a good starting point. The mistake is assuming every destination should receive the same bytes.
Think of the source article as structured content rather than a finished page. It contains a title, description, body, code blocks, image references, tags, and perhaps a canonical URL. A destination adapter can then turn that structure into the shape a platform expects.
type Article = {
title: string;
description: string;
body: string;
tags: string[];
canonicalUrl?: string;
};
type DeliveryResult = {
remoteId?: string;
url?: string;
state: "draft" | "published" | "pending" | "failed";
error?: string;
};
interface Publisher {
render(article: Article): Promise<string>;
publish(article: Article): Promise<DeliveryResult>;
verify(result: DeliveryResult): Promise<DeliveryResult>;
}
The adapter is where platform-specific compromises belong. It can cap the tag count, remove unsupported HTML, preserve fenced code blocks, upload local images, or add front matter. The core article stays clean, while the rendering layer absorbs the quirks.
This also makes editorial changes less painful. A conversational opening may work well on a community site, while a newsletter needs a shorter introduction and a stronger subject line. Those are not separate articles; they are deliberate renderings of the same idea.
Research and generation need boundaries
A workflow that can search the web, inspect a repository, generate illustrations, and publish to several accounts is useful, but only if each stage has a clear contract.
Research should produce notes with source URLs, not prose that is immediately pushed downstream. Drafting should consume those notes and produce a versioned article. Media generation should return stable asset URLs plus useful metadata such as alt text. Publishing should accept an immutable revision. Verification should check what the platform actually created.
In practice, the pipeline is easier to debug when its intermediate artifacts are boring files:
work/
research.json
article.md
metadata.json
assets.json
deliveries.json
You can open them, diff them, and rerun only the failed stage. If an API changes its tag rules, there is no reason to repeat research or regenerate a cover image. If an editor changes one paragraph, there is no reason to upload the same media again.
This is also where human review fits naturally when a team needs it. Review is a state transition on a specific revision, not an informal message that may refer to an older draft.
Retries should be dull, not dangerous
Network failures are normal. Duplicate posts are not.
Before creating a post, give the delivery a stable idempotency key derived from the article, revision, and destination. If the destination API supports idempotency directly, use it. If it does not, store the remote ID after creation and search your delivery log before trying again.
A retry policy should distinguish between errors. A rate limit deserves backoff. Invalid tags require a transformed request. An authentication failure needs intervention. A timeout is ambiguous and should trigger a lookup or verification step before another create call.
const key = `${article.id}:${revision}:${destination}`;
const previous = await deliveries.find(key);
if (previous?.remoteId) {
return publisher.verify(previous);
}
const result = await publisher.publish(article);
await deliveries.save(key, result);
return publisher.verify(result);
The final verification is not decorative. Some platforms return a URL for a draft. Others accept a request and process it asynchronously. A reliable workflow records the public URL only after the observed state matches the requested state.
The useful dashboard is a delivery ledger
Teams often build a polished “generate content” screen and leave operations in log files. The priorities should be reversed.
The most useful view answers practical questions: what revision was delivered, where did it go, what state is it in, when was it last checked, and what failed? It should let an operator retry one destination without restarting the entire run. It should also preserve the response body from failed API calls, with secrets removed, because “publication failed” is not actionable.
A compact ledger might look like this:
| Destination | Revision | Requested | Observed | URL | Last check |
|---|---|---|---|---|---|
| DEV | 7f3c9a1 |
Published | Published | Available | 10:42 |
| Newsletter | 7f3c9a1 |
Draft | Draft | Available | 10:41 |
| CMS | 7f3c9a1 |
Published | Failed | — | 10:40 |
This is less glamorous than a one-click publishing demo, but it is what makes automation trustworthy after the demo.
Automation should remove repetition, not editorial judgment
The best content workflow does not pretend every platform is identical, and it does not hide uncertainty behind a green checkmark. It keeps the original idea in one place, adapts the presentation intentionally, and treats each publication as a delivery that can be traced and verified.
Once those foundations exist, adding search, image generation, translation, or another publishing connector becomes straightforward. Without them, every new integration multiplies the number of ways the workflow can fail silently.
The surprising lesson is that multi-platform publishing has more in common with a deployment pipeline than a writing app. Version the artifact. Separate build from release. Make retries safe. Observe the destination. Report the URL and the actual state.
Do that, and the author gets what automation was supposed to provide in the first place: more time for the idea, and less time repairing the route it takes to readers.
Reference
The original workflow case study that inspired this engineering perspective:
Top comments (0)