An Instagram grid that survives the feed

Cutting nine squares out of a photograph is arithmetic and it always works. Whether the nine squares still read as one picture once a profile has drawn them is a different question, and it is decided by three things that happen after the cut.

The gutter is the design constraint

A profile grid draws a gap between every tile. It is small — a couple of pixels at the size most phones render — but it is present on both axes and it never goes away. Anything that depends on two tiles touching does not survive it: a face split down the middle, a line of text across a boundary, a horizon that has to be continuous to mean anything.

What does survive is composition that treats each tile as a panel. A subject that sits inside one tile, repeated shapes that rhyme across the grid, a colour field that changes gradually enough that a two-pixel interruption reads as a frame rather than a break. The useful test is to look at the gutter preview and ask whether any single tile is unreadable on its own, because in a feed that is exactly how each of them first appears.

Where to leave margin, and how much

The grid crops each tile to a square and then trims it slightly to fit its cell, so the outer edge of every tile is at risk. About two percent of a tile’s width is a fair working figure — on a 1080-pixel tile that is roughly twenty pixels on each side.

Keep eyes, mouths and hands well inside a tile rather than near an edge. A face interrupted by a gutter is the single most common way a grid fails.
Do not place text across a boundary. If text has to appear, put it inside one tile with room around it.
Watch the four inner corners. A point where four tiles meet is the least forgiving place in the whole layout.
Leave the outermost border of the whole picture quieter than the middle: it is the part most likely to be trimmed.

What the platform’s re-encode does to an edge

Every file handed to a platform is decoded and written again to that platform’s own settings. Two tiles that were one continuous photograph are re-encoded independently, and a JPEG encoder makes its decisions in eight-by-eight pixel blocks with no knowledge of what used to be next door. A gradient that crossed the boundary can come back a shade different on either side.

This is not a cutting error and there is nothing at this end that prevents it. What reduces it is giving the encoder less to argue about: hand over the largest tiles you can, avoid placing a smooth gradient exactly on a boundary, and accept that a flat sky spread across three tiles is the hardest case there is.

Cutting one, start to finish

Start from the largest copy you have. Nine tiles from a 1080-pixel-wide photograph are 360 pixels each, and the grid will display them at roughly three times that. Begin at 3240 pixels across or better and the tiles arrive at their displayed size instead of being stretched to it.
Decide the framing before the grid. Square tiles need a square source. Drag the framing box until what matters sits inside it, or switch to pad-to-fit if nothing may be lost and bars are acceptable.
Read the gutter preview, not the lattice. The lattice over the photograph shows the cut. The preview below it shows the destination, with the gap drawn in. Judge the composition there.
Keep the format on JPEG unless you have a reason. The platform re-encodes to JPEG whatever you send, so a PNG upload buys nothing except a slower transfer. Quality 82 is a reasonable place to hand it over.
Post from the bottom right. Upload in the order printed on the tiles, which runs backwards through the grid. The archive carries the same list as posting-order.txt.

What this guide cannot promise

Two things, and both are outside anyone’s control at this end.

The compression a platform applies is its own and it changes without notice, so a hairline mismatch along a tile edge can appear on a cut that is exact to the pixel. And the convention that a profile fills from the top left is a platform behaviour, not arithmetic — the day it changes, every posting order written anywhere, including the one in the archive this tool produces, is wrong.