HeyNews

Another update from @heynews (the AI tool we're building to help newsletter operators draft issues in their voice).

This release was mostly about trust in generated output, plus a less cramped Compose screen. The part I'd flag for other builders is the link validation loop, because "the model added a link that doesn't exist" is the generative-app failure mode nobody notices until a user ships it to 10,000 inboxes.

What changed:

1. Generated drafts get a link audit before saving. Every link in a draft has to come from a story the user selected or a reusable link they approved in their AI Writer settings, and the check also catches unfinished placeholders. If validation fails, we rewrite once. If the rewrite still fails, we save nothing and let the user retry. We decided a failed generation is cheaper than a plausible-looking draft with an invented URL in it.

2. Drafts keep the user's closing lines. Sign-offs, "hit reply" invitations, plugs for their other newsletters. These kept getting dropped or paraphrased because they look like boilerplate to a model, even though they're the most personal part of an issue. Now they survive, and their calls to action only get linked under the same rule as above.

3. Paste a link list straight into Compose. A lot of issues start life as a pile of links in a spreadsheet or Slack thread. Users can now paste the whole list; we extract the links, dedupe, preview what will be added, then fetch them as one-off stories in paste order. Large lists are processed in the background and persist after the user leaves the page, with per-link retries on failure.

4. Long-running AI refinements are resumable. Users can refresh or navigate away while a saved draft is being refined. The draft locks during the run, so a manual edit can't race the result, and on return, we restore the in-progress state or apply the finished copy as unsaved changes for review. On failure, the previous copy is untouched.

5. Scheduled drafts use the reader's issue date. If an automation prepares a Friday issue on Thursday, the AI Writer now writes with Friday in mind, respecting the saved lead time and timezone, including seasonal clock changes. This killed a class of "happy Thursday" bugs in weekday-specific copy.

Plus a Compose layout rework (more space for story selection, the AI Writer picker moved ahead of the story workspace, the auto-select action demoted to a shortcut with an Undo) and navigation regrouped by publication.

If you're building anything that generates user-facing content with links, curious how you handle validation. Rewrite-once-then-fail felt right for us, but I can see the argument for more retries.

I'm a founder who spent most of my career doing non-technical work, and I now build our product. I got there by talking to AI coding agents night after night. The problem with that: the agents are good company right up until 4am, when you're still hitting approve every thirty seconds, and the thing you actually want is sleep.

So I wrote up how I stopped babysitting them and built a loop I can actually walk away from: https://blog.heynews.co/afk-coding-agent-loop/

The short version. I fill a queue of GitHub issues during the day, each with enough context for an agent to complete it without coming back to me for a product decision. The ready ones get a label. From there, the agents carry each issue through implementation, review, fixes, merge, and deploy checks, then unblock the next one. I stay in charge of what gets built. They do the rest.

A few things that made it work as a founder, not just a demo:

• The decision of what to build stays human; only the labor moves to the agent

• Labels as the state machine, so I can read the status of any work at a glance

• Review runs as a separate loop with fresh context, not the same agent that wrote the code and is now sure it's right

• A gate that checks GitHub directly instead of trusting the agent's "I handled it"

• Nothing reaches staging without a PR, ever

The real lesson wasn't about code. Getting an agent to write code was the easy part a year ago. The hard part is the structure around it: enough that the work stays legible and safe to leave running while you're not watching. That structure is the moat, not the model.

If you're building solo or running a small team, where do you keep yourself in the loop versus letting the system run? And what's the first repetitive part of your business you'd hand to a queue like this?

Another quick update from @heynews (the AI tool we're building to help newsletter operators draft issues in their voice).

This week split across two areas of the product: subscription controls and content quality. Sharing here because both are parts that most builders postpone longer than they should.

What changed:

1. Pause your subscription for a month. New control inside Settings > Subscription. Billing stops, your existing newsletters and settings stay readable, and HeyNews resumes automatically at the end of the month. The use case: people who want to skip a billing cycle without losing their setup, or who'd rather pause than cancel outright. Pause is also a retention surface most products underinvest in, so we wanted ours to behave the way users expect rather than the way most billing dashboards do.

2. Cleaner cancellation. If you want out, you share a short reason, see any save options (including switching to a cheaper plan if the current one is too much), then schedule cancellation for the end of the billing period. "Keep my subscription" reverses it if you change your mind before then. The goal was to make the cancel path honest, not friction-heavy.

3. Story titles closer to the source, search across more fields. Content Desk now keeps story titles closer to the original article or video instead of paraphrasing them. YouTube stories show the native video title and published date. Renaming a source no longer rewrites the titles of stories that came from it. Search also broadened to check titles, summaries, source names, and URLs together, so you can find a story by topic or domain when you don't remember the exact headline.

Plus the usual round of fixes. Stripe Portal cancellations now sync back to HeyNews correctly, including the option to keep your subscription before the billing period ends. Scheduled downgrades use the correct billing period that Stripe stores on the subscription item. Trial cancellation and missing-payment-method handling are more consistent. Retrying a stuck story extraction clears old error messages once the retry succeeds.

If you're building anything with subscription controls or content metadata, happy to compare notes in the thread, including the boring engineering ones about how we wired up the schedule-cancellation and "keep my subscription" state.

Cagri, co-founder

Quick update from @heynews (the AI tool we're building to help newsletter operators draft issues in their own voice).

This week's release is mostly the unglamorous kind of work: making the product behave the way people expect when they run more than one newsletter in the same account. Sharing it here in case any of you are dealing with similar multi-tenant edge cases in your own products.

What changed:

1. Story picks stay in the right publication. If you manage two or three newsletters under one account, story suggestions, filters, and the "selected stories" list no longer leak across publications. Each one now sees only its own sources and audience.

2. Clearer access to older stories. If a story falls outside your plan's access window, we now keep the locked content out of the draft flow and show a real upgrade path, instead of letting the draft attempt run and then fail. We also stopped overwriting the original date on stories you intentionally picked, so older selections don't get silently rejected.

3. Smarter source checks. RSS and social source checks now pause when no active publication needs them and resume the moment one does. The practical effect: source libraries stop quietly drifting in accounts with many paused or archived publications.

Plus the usual round of fixes. Shared RSS feeds resume correctly when a publication resumes using them. Story detail pages and usage tracking now stay tied to the current publication. A class of AI drafts that had been failing due to model settings completes successfully.

If you run a newsletter (your own or a client's) and want to see how HeyNews handles your sources and voice, the product is at heynews.co. Happy to answer questions in the thread, including the boring engineering ones about how we keep multi-publication context isolated.

Cagri, co-founder