AWS Builder Center
Weekend Creative Agent Challenge: Chalk

Weekend Creative Agent Challenge: Chalk

An agent that writes one scene of an action comic every morning at six and deletes it at midnight. Its entire memory is a single DynamoDB row it writes to itself, so the story keeps going even though none of its chapters survive.

AI Developer
One scene a day. Gone by midnight. The story keeps going.
Right now, scene 14 of Chalk ends with Nutmeg's grip slipping and a tendril closing in.
At midnight tonight that scene stops existing. Tomorrow at six in the morning, without me, the story will pick up from it anyway — and it will do that without ever reading the scene it's continuing.
That gap is the whole project.


Vision and what it does.

Chalk is an agent that writes one scene of an action comic every morning at six, and deletes it at midnight.
Not a prompt. Not a summary. An actual scene — four to six panels, each with a line of art direction, a line of dialogue, a sound effect, ending on a cliffhanger, with a six-second establishing shot generated to sit above it.
Then midnight comes and the scene is gone. There is no archive, no back button, no "see previous scenes." If you missed yesterday, you missed it.
The serial is on scene fourteen. Not one of the first thirteen still exists.
To be exact about where those fourteen came from, because it matters: I seeded the chain in one sitting before launch, each scene generated from the last one's memory so the drift is real, and every run stamped with the time it actually happened rather than backdated to look like a fortnight of mornings. The page says so plainly — "Chain seeded 21 Aug. First scheduled run: 22 Aug, 06:00." The chain is genuine. The unattended part starts tomorrow, and I'd rather you watched it happen than take my word for it.
I built it this way because I've noticed something about myself and everyone I know who makes things: we hoard. Bookmarked prompts, screenshotted ideas, folders of references we will absolutely come back to. The hoard becomes the hobby. Chalk removes the option. There is one scene, it's today's, and tomorrow it won't be there.

How I built it

So how does a story survive when none of its chapters do?
The agent's entire memory is one row in DynamoDB. Not a log. Not a vector store. One row, overwritten every single morning, containing a block the agent wrote for itself the day before.
This is the real thing, as it stands after fourteen scenes:
heroes = Timmy, Lila, Nutmeg
allies = none
villain = Mist Tendrils, creatures of the dark forest that want to trap the kids
goal = Retrieve the Crystal of Light to save the forest
where = Deep within the dark forest
just_happened = Timmy and Lila desperately try to hold onto Nutmeg,
a tendril comes close to snatching her
unresolved = Nutmeg's grip is slipping
running = Nutmeg's constant chattering and Timmy's fear of squirrels
established = Tendrils are fast and strong; heroes must retrieve
the Crystal of Light to save the forest
Nine fields, and every word of that was written by the agent, about itself, for itself.
Look at  running . Nobody asked for Timmy's fear of squirrels. That's a gag the agent invented somewhere in the first fourteen scenes and has decided, at every hop since, is worth one of only nine slots it has. Nothing forces it to keep that. It just does.
And look at  just_happened . Two sentences about a scene that no longer exists anywhere.
So the agent never reads yesterday's comic. It reads the note it wrote to itself about yesterday's comic. Then, before it finishes, it writes tomorrow's note.
Scene 14 knows scene 13. It has no way to reach scene 12. Every morning it decides what about yesterday is worth carrying forward and what to let go — and the thing it's summarising has already been deleted, so that decision is final. Get the note wrong and the story doesn't get worse, it breaks. There's nothing to check it against.
That's the part I'd defend if someone told me this is a cron job with a language model bolted on. It isn't looping over a fixed prompt. It's maintaining state it authored itself, and choosing what to forget.
Two smaller decisions hold it up. The scene expires at local midnight; the memory row has no TTL at all — story persists, output doesn't, and those are separate promises that need separate mechanisms. And every day gets one anchor word, picked by  hash(date)  from two hundred concrete nouns. It never appears in the scene, it just nudges. Without it, a model answering its own output collapses into vapour within a few weeks. I've watched that happen, which is why the list is there.

AWS services and architecture


Amazon Bedrock does two jobs. Nova Lite writes the scene — in delimited  ART:  /  SAY:  /  SFX:  lines rather than JSON, because Nova Lite fumbles nested structure and I learned that the expensive way on an earlier build. Nova Reel renders the clip at 1280×720, seeded from the scene's own seed.
Amazon DynamoDB holds three things: the carry row, the scene row with its midnight TTL, and an append-only runlog that stores timestamps and the scene title only. Never the panels. Never the seed. If the seed survived, nothing was really destroyed and the whole premise would be a lie.
Amazon S3 and CloudFront serve a flat  prompt.json  written on each run, which means the site keeps working when Lambda doesn't — it just shows an honest, stale timestamp instead of pretending. An S3 lifecycle rule expires the clips after a day, but lifecycle runs on its own schedule, so the real guarantee lives in the read path: if  expiresAt  has passed, the page serves nothing regardless of what's still physically sitting in the bucket.
Amazon CloudWatch alarms on errors, on  Invocations = 0  over 25 hours with  TreatMissingData: breaching , and on a render stuck in  rendering  — a failure mode that only exists once you go async. There's an AWS Budget on Bedrock spend too, because a scheduled job that generates video is precisely the shape of thing that bills you quietly for a month.
One validation gate is there for a reason worth naming: the comic is for a nine-year-old, so there's a ban list on violence and name-calling sitting alongside the structural checks. The strictest check, though, is that all nine carry fields must be present. A short scene is a bad day. An incomplete carry block loses the story.

What I learned

The latency that disqualifies a model everywhere else is free in an agent.
Nova Reel takes minutes. That rules it out of anything a person is sitting in front of. At six in the morning it costs nothing, because nobody is waiting — and once I saw that, I stopped treating async as a problem to engineer around and started seeing it as a reason agents suit slow models. So the two Lambdas aren't joined by a polling loop or a state machine. They're joined by an S3 event. The job's own completion is the trigger, which is both the cheapest design and the only one I couldn't get wrong.
A classifier that guesses will guess in your favour.
My run accounting asked whether  trigger != "seed"  to decide if a run had been autonomous. Every runlog row written before that attribute existed had no  trigger  at all — so it defaulted to autonomous, and the page cheerfully displayed five past mornings, each stamped exactly  06:00:00 .
I hadn't faked anything. The code resolved an unknown in the flattering direction, which is worse, because I wasn't looking for it. And it got worse still when I checked the template: EventBridge Scheduler was invoking the function with  {}  — byte-identical to me running  aws lambda invoke  by hand. Every test I'd run had minted a fake unattended morning.
The fix is to fail closed. The schedule now sends  {"scheduled": true} , nothing else is allowed to claim it's the scheduler, and anything untagged counts as backfill. The page's entire claim is that this happens without me. That's the one number the code must never be allowed to round up.

On the brief

The challenge asks for something ready when you get back, and Chalk only waits a day. That's deliberate. An agent that accumulates forever isn't choosing anything — it's filing.
So what's waiting for you in the morning isn't a pile of output. It's one scene, made while you weren't there, ending on a cliffhanger you have until midnight to enjoy. And underneath it, a note the agent wrote to itself about a story you can no longer read, which it will pick up tomorrow at six whether you show up or not.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article