Weekend Creative Agent Challenge: Mot Tissé
Mot Tissé is a daily agent that weaves new A1-level French vocabulary into a short AI-generated story or poem — built with Amazon Bedrock (Nova), Lambda, and EventBridge Scheduler, so it runs entirely on its own schedule with zero clicks.
Weekend Creative Agent Challenge: Mot Tissé
#agents
Vision what it does and why
I'm learning French, and like most people learning a language, my biggest problem isn't
finding vocabulary lists — it's that I keep reviewing the same words I already picked,
in the same order, until they stop meaning anything. What I actually needed was something
that would hand me new words in context, on its own, without me having to decide what to
study each day.
Mot Tissé ("woven word") is a daily agent that solves that: every day, on its own schedule,
it picks a theme, asks Amazon Bedrock (Nova) for six new A1-level French words tied to that
theme, and weaves them into a short French story or poem — plus an English glossary so I can
check myself. No app to open, no prompt to write. I just visit the page and there's a new
piece of French waiting, built from words I didn't choose myself.
The idea was partly inspired by a fellow builder's "Word Loom," which does the same thing in
the other direction — English vocabulary for non-native speakers. I flipped it: a native/fluent
English speaker acquiring French, with the output shaped like an actual study card (word list +
difficulty level) rather than plain prose.
How I built it
I designed this as three independent pieces wired together with Terraform, mirroring the module
structure from an earlier AWS project of mine (a DynamoDB-backed ordering system), so the repo
would stay familiar to maintain:
- a
sitemodule (S3 bucket, static website hosting, public-read policy),
- a
lambdamodule (the generation function + its IAM role, scoped tightly to one bucket and one Bedrock model ARN),
- a
schedulermodule (an EventBridge Scheduler cron job with its own invoke role).
The key design decision was making the word selection itself part of the agent's job, not
something I curate. The Lambda keeps a rolling history of the last ~150 words it has already
used, stored as JSON in S3, and tells Bedrock to avoid them — so the vocabulary keeps moving
forward instead of circling the same set. I also had Bedrock alternate between story and poem
format and pick from a pool of everyday A1 themes (le café, la météo, le marché, and so on), so
the output doesn't feel repetitive day to day.
The one real bug I hit was a subtle S3 permissions gotcha: on the very first run, the Lambda
looks for a history file that doesn't exist yet, and my code caught
NoSuchKey to handle that.
Instead it got AccessDenied . It turns out S3 masks a missing object as 403 instead of 404 when
the caller lacks s3:ListBucket on the bucket — a deliberate anti-enumeration behavior I'd never
run into before. Adding a narrowly-scoped ListBucket permission (on the bucket itself, not the
objects) fixed it. Small thing, but a good reminder that IAM scoping decisions can silently change
what error your own code sees.AWS services used
- Amazon Bedrock (Nova Lite) — generates the themed vocabulary, glossary, and story/poem text as structured JSON, following a system prompt tuned for A1-level output.
- AWS Lambda — the orchestrator: loads word history, calls Bedrock, writes results back to S3.
- Amazon EventBridge Scheduler — fires the Lambda once a day with no human involvement.
- Amazon S3 — doubles as static website hosting for the front end and as the agent's storage for daily output and word-history state, so there's no database and no backend API at all.
- IAM — every role is scoped to exactly what it needs: the Lambda's role can only touch this one bucket and invoke this one Bedrock model; the scheduler's role can only invoke this one function.
What I learned
The biggest thing I took away is how little infrastructure an "always-on" agent actually needs
when the LLM does the interesting work. There's no queue, no database, no API layer — just a
timer, a function, and object storage. The hard parts were in the prompt design (getting Bedrock
to reliably avoid repeated words and return clean JSON) and in IAM edge cases like the one above,
not in wiring services together.
It also reshaped how I think about "personalizing" an agent. Building this for my own French
level meant the prompt had to encode a real constraint (A1 vocabulary, present tense, ~100 words)
rather than just "write something creative" — which made the output far more useful to me
day-to-day than a generic version would have been.
Links
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article