What this page is for: a niche page for experienced builders and maintainers: how the morning show is made, and how to build your own.
In short
Every morning I listen to a short radio comedy before I start work. Two AI crews, the Night Shift and the Day Shift, talk through what happened in the box overnight and what bubbled up in the river of news. The humor is the kind I like first thing: deadpan, nerdy, engineering-office jokes about job queues, flaky tests and never enough electricity. It gives me the overview I need, sets my mood and intention over coffee, and then I go remove bottlenecks and direct the agent team for the day.
It's fiction about real work. It's also one example of a bigger idea: a conversational interface doesn't have to be a support line with no humor and no feeling. You can design one for almost any job.
What I actually get from it
The headlines, lightly
The crews mention a few top items from the latest daily briefing, the stories that rose to the top of the river. The briefing reads about the last day and a half of new material, with a 10-day net so nothing I missed slips past. The full reports are on News.
The state of the shop
The Day Shift gets real numbers from the operations logs, such as what the release queue published yesterday and how many jobs are waiting. It only uses numbers it's given.
A mood
I start the day laughing at problems I'm about to go fix. For me, that works better than a dashboard.
How it's made, in plain words
- Sources. Our nightly engine reads the channels and people I follow and writes the daily briefing.
- Ranking. The show takes the briefing's top headings, drops anything about investing, and puts my own interests first. That's my bias, and I'm saying so.
- Script. A local model on our own Macs writes each half from a persona and tone guide: who each character is, how they talk, the running gags, what they must never do. Each crew leaves the other one hand-off line.
- Checks. Plain code, not a model, checks every draft: only our cast speaks, every line passes our clean-language gate, the length is right, the hand-off is there. A failed draft is rewritten a few times. If none passes, the night is skipped.
- Voices. Each character has its own synthetic voice, made on our own machines. Sound effects are generated in code. The cast credits live in each episode's notes and audio metadata.
- Publish. The two halves are joined into one episode with chapters and even loudness, and it goes out to podcast apps by 5:00 AM Eastern. When I hear a mispronounced word or a wrong pronoun, it gets fixed and re-voiced.
Build your own
You don't need our cast. The recipe works for a build pipeline, a content-tone review, a teaching team, or a status meeting nobody wants to attend.
- Sources: logs, tickets, CI results, feedback, a news feed. Real inputs only.
- Ranking: decide what the listener needs first. Write the rule down, biases included.
- Script with a persona and tone guide: a few characters with clear jobs and voices, a humor style, hard limits.
- Voices: one voice per character, credited in the episode metadata.
- Checks: code that rejects invented numbers, off-limits words and broken format before anything is voiced.
- Publish: a feed, a page, a chat, wherever the person already is in the morning.
Cheap: script writing on a local model, local text-to-speech, sound effects made in code, and the checks. Our show uses no paid AI service.
Costly: time. Dozens of model calls and a long voicing run per episode, a machine left running, and the hardest part, a human who listens and keeps tuning the tone.
Builder course planned, not built yet
This will be a course on building your own conversational interface. These are the planned topics:
- several agent personalities, and testing how they play off each other;
- adjusting tone and intent;
- story arcs of different lengths, from one episode to a whole season;
- inside jokes, humor styles and choosing humor topics;
- randomizing priorities across real elapsed days, so the show doesn't repeat itself;
- how much memory the team keeps, so each role stays itself in the style you chose. Today our writer sees only the last three nights' log.
There's nothing to sign up for yet. When the course exists, it will be on Build.
Where to listen or read
On the site
Becoming Stellar has every episode, with show notes and transcripts.
In your podcast app
Video and the virtual campus
Both are coming later.
plannedWhy I bother
The Manifesto explains the why. The short answer is that I want people to build with these tools, not just watch them, and to build in a way that helps us not to merely survive, but to truly thrive! This is how I start my day, the CognitoRiverDelta way!
Fiction about real work, performed by an AI-agent cast. Naming a tool is not an endorsement.