Chiblu is a curated marketplace for Indian craft — weavers, ceramicists, woodworkers, the kind of work that takes weeks to make and seconds to scroll past. Instagram is where buyers find us and where makers see whether we treat the work with care. The feed mattered.
The feed also kept dying. Week one we'd post daily. Week two, every other day. Week three, silence. Hiring a social-media person felt premature for the size of the storefront. So we built a team of small agents instead — each one handling a single event — and shifted our job from making posts to approving them.
This is how it works at Chiblu.
A few things had to be true at the same time.
- The feed needed multiple kinds of posts — featured products, new arrivals, maker spotlights, festival wishes, recruitment for the marketplace itself.
- Each kind had a different rhythm. Featured products weekly. Festivals when the calendar said so. Maker recruitment whenever there was an angle worth using.
- We wanted to spend 15 minutes a week on this, not 5 hours.
- We wanted to know on Sunday morning what would post on Wednesday.
The default answer — "draft, polish, schedule manually" — collapsed under those four constraints. So we unbundled it.
Six agents, each watching for one kind of event.
Every agent has a narrow job and a clear trigger. The schedule agent doesn't know about festivals. The festival watcher doesn't know about new arrivals. The auto-publisher doesn't care what created the draft — it just looks for approvals.
Picks a featured product · writes a caption · creates four design variations.
Checks today against the Indian festival calendar. Pongal, Holi, Eid, Diwali, Christmas.
Rotates through twelve maker-recruitment angles for chiblu.com/seller.
Scans Chiblu's catalog · queues a draft for each new product across the next week's slots.
An AI writer trained to match our voice · picks one italic-gold accent word per post.
Publishes approved drafts to Instagram exactly when they're scheduled to go out.
The surface we look at when we want to know what's coming.
Each chip is one fire. Festivals show up on their date. Recurring posts on their weekly cadence. Pre-generated drafts on the date they'll publish. One button at the top — Generate all drafts for May — runs the pipeline for every empty cell in the month. It skips cells that already have a draft, so we can re-run it anytime without duplicates.
Clicking a chip drops us into a per-fire page: the caption that'll go out, the four design variations, a template override picker if we want this Wednesday to be a carousel instead of a single post. Generation happens server-side and lands as a draft we can approve later.
The web view is for planning. The approving happens on Telegram.
Sunday morning at 8 AM IST, the team gets a digest message listing every draft scheduled for the coming week. Each draft arrives as a swipeable album of four design variations, with a message underneath holding six buttons: 1, 2, 3, 4, ↻ (regenerate), ✕ (skip). The video at the top of this piece is one of these reviews, end to end.
Tapping a number doesn't publish immediately. It marks the draft as scheduled. When its scheduled time arrives, the auto-publisher posts it to Instagram. Sunday's tap → Wednesday's post.
The split between approval and publish is the load-bearing part of this. We clear the week in 10 minutes without thinking about timing.
The agents reach for different layouts depending on the event.
Editorial when the photograph carries the story, typography when the message itself is the visual, festive blocks when the calendar says it's the day. A few of them, side by side:
The agents don't pick the design; we do, once. They just keep using it. Each schedule has a single layout by default, or rotates between two — a single-post one week, a carousel the next. The festival watcher uses one designated festival template across the year.
The job changed shape.
When we started we imagined the agents writing the posts and us approving them. That's accurate but it undersells the shift. The actual change: the editorial decisions moved earlier. We're not picking words on Wednesday. We're picking a rotation in January. We're not laying out a Diwali post in October. We're flagging enabled on the festival in advance. The work compresses into a few moments of taste — pick the template, pick the rotation, pick which festivals matter — and then six agents make the consequences happen.
Curator, not creator. Approval, not authoring.
The other thing that emerged: predictability stopped being a constraint and became the point. We can tell you what'll post next Sunday at 10 AM IST. That's not a limitation. That's the whole reason the calendar works.
Four principles that kept showing up across rewrites.
The skeleton is small.
Three things start a post — a timer, a date on the festival calendar, a new product showing up in the catalog. Agents in the middle turn that signal into a draft. We see drafts on Telegram and on the calendar. The auto-publisher closes the loop into Instagram.
None of this is novel on its own — a timer, a writer, a renderer, a phone-friendly review surface. The interesting part isn't the pieces. It's that each one does a single thing and stays out of the others' way. Changing how captions get written doesn't touch how the calendar shows what's coming up. Adding a new festival doesn't ripple into the Telegram review flow.