
Weekend Creative Agent Challenge: Codice
Códice is an autonomous agent that draws one page a day. Every morning it wakes on a schedule, reads the real sky over Mexico City — temperature, wind, cloud cover, rain, the length of the day, the phase of the moon — turns that reading into a visual language, draws an original SVG page, checks its own work, and publishes it to a gallery. Nobody triggers it. Nobody prompts it. When you come back, the page is already there, and it is a page that did not exist yesterday and cannot exist again tomorrow.
The vision
Last weekend I built something creative that I had to open. This weekend the
brief was to set it free, and that reframing turned out to be the whole
exercise. "Always-on" is not a deployment detail. It changes what the thing is.
An app waits to be asked. An agent decides on its own that today is worth
making something for.
Códice is an autonomous agent that draws one page a day. Every morning it
wakes on a schedule, reads the real sky over Mexico City — temperature, wind,
cloud cover, rain, the length of the day, the phase of the moon — turns that
reading into a visual language, draws an original SVG page, checks its own work,
and publishes it to a gallery. Nobody triggers it. Nobody prompts it. When you
come back, the page is already there, and it is a page that did not exist
yesterday and cannot exist again tomorrow.
The problem it solves is small and honest: generative art tools make you do the
work. You open a tab, you think of a prompt, you tune it, you pick a favourite.
Códice inverts that. The tool never has to be opened. Its output accumulates
into something with a shape — a codex, one page per day, where the interesting
thing is not any single page but the run of them.
The decision the whole thing rests on
My first version passed the weather straight into the prompt.
16 °C, 89 % humidity, 100 % cloud cover, wind 4.4 km/h . It drew a weather report. Bands
labelled like a chart, a wind arrow, something that wanted to be a dashboard.
Of course it did — I had handed it data, so it visualised data.So I moved the interpretation out of the prompt and into the code. Now every
measurement is spent in a module called
grammar.py , and what reaches the
model is only words and hex values:archetype basin — nested open curves cupping one another, opening upwardcomposition sits near the middle, evenly dividedgesture still — lines hold their direction; nothing leanssurface submerged — almost everything semi-transparent and overlappingtexture heavy grain — dense stippling across the larger shapesfalling a few short falling strokes, widely spacedpalette #25302e #d6dbda #465f67 #729782 #c171a6That is the same day. The model cannot draw a weather report from it, because
there is no weather in it. Temperature becomes hue through a piecewise
interpolation — violet when it freezes, teal when it is mild, amber and then red
as it climbs. Cloud cover drains chroma and thickens the veiling. Wind decides
whether lines lean. Day length puts the horizon high or low. A test in the suite
asserts that no measurement ever appears as a string in the brief, because this
is the kind of rule that quietly breaks.
The moon takes it one step further: the code draws the moon, not the model.
The agent is told only to leave a circular void in a given region and is never
told the phase. Afterwards the Lambda composites in an exact terminator — an
ellipse of semi-minor axis
r·|1 − 2f| for illuminated fraction f . Two hands
on the same page.How it was built
Python 3.13 on Lambda, no dependencies beyond the boto3 already in the runtime.
One CloudFormation stack. Nine small modules, each with one job.
Services: Amazon Bedrock (Claude Sonnet 4.5) draws the page. AWS Lambda runs
the routine. Amazon EventBridge Scheduler wakes it at 06:15 local — Scheduler
rather than an EventBridge rule specifically for
ScheduleExpressionTimezone ,
so the agent keeps local mornings across daylight saving. Amazon DynamoDB holds
the memory. Amazon S3 hosts the gallery as a static site. CloudWatch Logs keeps
the record. Weather comes from Open-Meteo, which needs no API key.EventBridge Scheduler → Lambda → ┌ Open-Meteo (today's sky) ├ DynamoDB (memory + style genome) ├ Bedrock (draws the SVG) ├ validator (rejects, argues, retries) └ S3 (publishes, rebuilds gallery)The validator is the part I'd defend hardest. The page is written by a
language model and then served from a public bucket, so it is checked before
anyone sees it: well-formed XML, no
<script> , no <foreignObject> , no on*
handlers, no external references, only same-document url(#id) , only the five
colours it was given, inside an element budget, under 320 KB. Violations are
written as English sentences and handed straight back to the model, which
redraws. Three rounds. If it still fails, the day is skipped and the refusal is
recorded — publishing a page that failed every check would quietly corrupt the
codex, and an agent nobody is watching needs to be able to decline.Memory and drift. DynamoDB carries a style genome — stroke weight, symmetry,
negative space, complexity, hue drift — that mutates a little each night. Four
traits wander; complexity trends upward with a ceiling, so the agent grows more
ambitious as its body of work grows. The last six archetypes are remembered and
forbidden, so it cannot repeat itself.
What I learned
Constrain the model by removing information, not by adding instructions. I
spent a while writing longer and longer prompts telling it not to make a chart.
What actually worked was making the chart unmakeable.
Negative constraints beat positive ones for style. "Draw beautifully" did
nothing. The line that changed the output was: if it could be mistaken for a
dashboard, a blueprint or a wireframe, it has failed — with the specific
banned vocabulary listed. Axis lines, full-width rules, evenly spaced repeats,
ladders of identical steps.
I ran a bake-off and the honest result surprised me. On the same brief,
Nova Pro produced genuinely attractive pages but under-populated the canvas —
17 drawn elements against a floor of 44 — and passed validation inconsistently,
sometimes burning all three attempts. Claude Sonnet 4.5 passed first try on
every brief I threw at it. For a scheduled agent that nobody is watching,
reliability beat charm, so Sonnet draws and the model stays configurable.
Test the geometry you think is obvious. My first moon rendered every new
moon as full and every full moon as dark: I had the SVG arc sweep flags
backwards. I only caught it because I rendered eight phases and looked at them.
There is now a test.
24 offline tests, no AWS or Bedrock needed to run them.
The retry loop earns its keep in production. The very first page the
deployed agent drew was rejected once before it passed — exactly the failure
mode the local rehearsal had not shown me. An agent nobody is watching needs to
be able to catch itself.
Links
Built for the AWS Weekend Challenge. One page a day, and I never have to open it.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article