The situation
Movisvami posts management and leadership quotes from films and television as branded images, captions and articles across Instagram, Facebook, X, LinkedIn and a website. Done properly, one week of that is a research job, a writing job, a design job and a scheduling job.
Done with a generic AI content tool, it's worse than doing nothing — because the single thing a movie-quote brand cannot survive is a made-up quote. Language models are extremely good at producing a line that sounds exactly like something a famous actor said in a film he was never in. Publish three of those and the brand's entire premise is gone; the audience is precisely the sort of people who will notice.
So the interesting problem was never generation. It was verification.
What was built
An agent living in a Telegram group, operating as several specialists depending on what's asked:
- Calendar Agent — a theme produces a week's calendar. It searches a film database for ten to twelve candidates across film and television, filtering for rating, cultural recognition, leadership relevance and variety of era. Then it verifies, builds the leadership angle, scene context and character action for each survivor, and drafts headlines.
- Production Agent — turns the approved calendar into finished assets: branded images, captions per platform, and the long-form article.
- Publishing Agent — publishes now or schedules against the post dates, across all five platforms.
- Special Posts Agent — for events and milestones, routed separately so a timely post never gets mistaken for a calendar request.
The whole thing is driven from chat. A greeting returns the command menu; the operator types approve, approve a range of days, skip a day, swap a day with an instruction, produce, publish. A dashboard with calendar history handles anything that doesn't fit in a message.
Brand rules are encoded, not remembered: thoughtful, cinematic, executive — never motivational-poster. The quote anchors; the caption builds the leadership insight around it. Television counts as much as film. No exclamation marks. LinkedIn analytical, Instagram narrative. Hashtag banks by theme.
A script decides what's real
Where this system can be wrong
A quote does not get published unless a script proves it exists.
For every candidate, the agent must run a verification utility that calls a search API directly and does deterministic string matching against real results. The instruction to the model is unusually blunt, and deliberately so: this is not a judgment call, and the model must not substitute its own opinion for the script's output.
- Verified — the source URL is recorded against the quote and travels with it into production.
- Not verified — drop the quote immediately, no exceptions, and try another candidate, or replace the film entirely. Regardless of how plausible it looks.
This is the strongest form of the pattern running through all our work. The model is excellent at finding candidate quotes and genuinely good at drawing the leadership insight — so it keeps both those jobs. But on the one question where its confidence is worth nothing, its opinion is removed from the loop entirely and replaced with an exit code. Hallucination isn't discouraged here. It's structurally ineligible.
Two further gates, both about not destroying work. Overwrite protection: before creating a calendar or generating production files, a script checks for existing files at those dates and the agent must wait for an explicit confirmation before proceeding. Human approval between every stage: nothing produces until it's approved, nothing publishes until it's produced.
What happened
Theme in, calendar out, approved in chat, produced, published or scheduled across five platforms — and it feeds the cinema section of the Director's Cut newsletter.
The thing worth noticing commercially: this is a full content operation — research, verification, editorial judgment, multi-format production, multi-platform distribution — run by one person typing short commands into a phone. That's the actual shape of AI leverage in a small business, and it's a great deal less glamorous than the demos suggest. It works because the boring parts are scripts and the judgment parts stayed human.
What made it repeatable
The pattern generalises well past cinema: research → deterministic verification gate → human approval → production → scheduled distribution.
Any content operation with a factual substrate has the same shape. A pharma brand citing studies. A law firm citing judgments. A finance newsletter citing numbers. In each case the model should be doing discovery and drafting, and in each case there should be a script — not a prompt, not a careful instruction — standing between the claim and the audience.
The reusable question to ask any content team: what is the one thing you cannot afford to get wrong, and what deterministic check currently stands in front of it? Usually the answer is that a person is supposed to remember to check.
The engagement shape
- Find the falsifiable claim — the quote, the statistic, the citation, the price.
- Build the gate — a script that verifies it against a real source and returns an exit code, not an opinion.
- Keep the human at the approval points, and only there.
- Automate the rest end to end, including distribution.
The result is a content operation one person can run, that cannot publish a claim nothing verified.
What's still open
Named, not hidden. A case study that admits open items is worth more than one that doesn't.
- Verification depends on the search index. A genuine quote from an obscure title may fail the check and get dropped. We prefer that failure direction — losing a real quote costs one slot; publishing a fake one costs the brand — but it is a real limitation, not a solved problem.
- Image production is aesthetic, and aesthetics don't have an exit code. No script can gate whether a branded image looks right. That stays human.
- Platform APIs drift. Five publishing targets means five things that can change under you.
- Single-operator design. The approval flow assumes one person deciding. A team would need roles.
Next case study
The Collections Agent →