A few weeks back I posted some thoughts on not using 5-star review mechanics for ATStore reviews, tried to distil what a good micro-review mechanism should do, and offered my pet “is/not” micro-review mechanism.
That (longer) post on Github
I pontificated earlier (
) about the challenge of using a 5-star review system (anywhere, but particularly on the ATStore). I’d love to see us embrace other mechanisms that could be more informative on the breadth of what the Atmosphere offers, and might even provide a lexicon that other Atmospheric apps might lean on for reviews too!
The big challenges I see with written & 5-star reviews:
- As Ariel mentioned, stale reviews mean you often need to pay close attention to when the review was and what changed on the platform since — who has time for that when browsing lots of potential things?
- Having a time component to how reviews are surfaced seems important
- It requires a bunch of effort from a potential reviewer to summarize the entire app, and to reduce their mixed views into a single output; which means folks either don’t bother, or pick a particular aspect to shower praise or criticism upon.
- Making it super easy and fast for anyone to offer a high-context review (with scope for conflicted feelings) would be excellent… if possible
- If folks find something they don’t like about an app, the reviews can become a place to ‘shout into the void’ in a way that can appear overly entitled, stir discontent, and produce pile-ons, when it’s a reasonable & genuine frustration (if perhaps ungracefully framed)
- Finding a way to channel frustration into something owners can channel into improvement seems important
- Ensuring there’s good context as to what a reviewer should expect from leaving a review seems important
- In small communities, and especially in open systems like atproto, your rating is also a means of support — a “critique in private, praise in public” kinda thing — so early ratings can seem rather one dimensional or fake
- Making sure there’s a way to offer balanced feedback without being unsupportive seems important
- There are many interesting critiques of 5 star review systems, often pointing out how folks often vote 5 for ‘good’, 1 for ‘bad’ and 4 for ‘excellent except for that one thing I want’, making it very hard to get useful data out, even when there are lots of respondents.
- Gathering data that maximises the ability for useful analysis seems important
- A single thing often has different value in different contexts. 5 stars for novelty — I want to support this, 1 star for dependability — I wouldn’t use it foundationally, yet.
- Allowing expression for nuance, without inflating complexity, seems important
What (I think) people want from a review system:
- To know the thing is worth investing a little time in (via social proof)
- it does what it says
- it’s not scammy/doesn’t exhibit any dark patterns
- To evaluate A vs. B, if there are two similar things
- To have expectations set. Eg. a thing trying to be artistic and beautiful at the expense of broad feature set vs. a thing trying to be fully featured at the expense of good looks
- … (probably much else not in my head right now )
To that end, here’s a pet idea I’ve had knocking around for a long time, that feels rather suited to atproto, that I call “is/not” reviewing.
- There is an open list of adjectives a review platform has defined for writing reviews.
- A reviewer is offered a handful of them* as three-way toggles; eg. “beautiful” → “is beautiful”, “is not beautiful” or not asserted (the default)
- The reviewer picks the ones they have strong opinions about, and offers that “thumbs-up/thumbs-down” verdict on each of the adjective dimensions.
- You could capture a text review too, if you like
- The review system then aggregates these votes in any number of dimensions: how polished it is (is beautiful/is clean/is not work-in-progress), how fun it is (is exciting, is not boring, is artistic), and so on.
You can then extend this as the dataset grows
- Paired with a “Like” button, you can extract that a particular person always rates “is beautiful” for the things they Like → now you can up-rank things specifically marked “is beautiful” for that person
- You can use different languages for the adjectives, and map them in a non-1:1 fashion; “es rico” could boost an “is expensive” signal in Spain, but an “is delicious” signal in Latin America.
- (Hell; with enough data you can eventually ignore what the tag’s text is entirely, and just use clustering algorithms to see how a person’s ratings fit in with the wider ecosystem)
I feel this system could be particularly good across the Atmosphere particularly (given this kind of methodology could be a lexicon used across review platforms too!) — but this idea hasn’t seen the light of day yet, please critique and adapt it; I’d love to hear your thoughts!
Generally: has the ATStore given any thought to other rating/review mechanisms? Do you have any other goals for what you have or will implement?
After a prompt from @erlend.sh (thanks!) figured I’d break that idea out here, including a possible lexicon, and ask for feedback — particularly from folks who’ve worked with the psychology & statistics of review systems before!
Here’s my proposed lexicon for micro reviews that could replace any 5-star mechanism (the justification is in the expander box above).
I’d love some feedback — I’ll be happy to try out this lexicon on its own (marque.at is calling ) but if it’d be useful for others I’d happily get it ready for the lexicon community.
An example review record, declaring that the ATStore app entry “is trustworthy”:
{
"$type": "community.lexicon.review.isnot",
"subject": {
"uri": "at://did:web:atstore.fyi/community.lexicon.app.entry/at-store-tid",
"cid": "bafywhateverwhateverwhateverwhateverwhatever"
},
"adjective": "trustworthy",
"direction": 1,
"createdAt": "2026-05-17T14:38:51.423Z"
}
With prototypey format lexicon (for readability here) like this:
import { lx } from "prototypey";
export const review = lx.lexicon("community.lexicon.review.isnot", {
main: lx.record({
key: "tid",
description: "A lightweight 'is/not' review of something represented by another atproto record",
record: lx.object({
subject: lx.ref("com.atproto.repo.strongRef", {
required: true,
description: "The subject of this is/not review",
}),
adjective: lx.string({
required: true,
minGraphemes: 1,
maxGraphemes: 16,
maxLength: 160,
description: "An adjective that the record publisher believes describes the subject either in the positive or negative direction.",
}),
direction: lx.integer({
enum: [-1, 0, 1],
required: true,
description: "Whether the record publisher thinks the subject is the adjective (1), is not the adjective (-1), or is explicitly making no comment (0).",
}),
createdAt: lx.string({ required: true, format: "datetime" }),
}),
}),
});
First thoughts on what I’ve suggested here:
- Repo owners can use whatever
adjectives they like, but an AppView ingesting them would only pay attention to the ones they were interested in. - I’ve gone with
enumfor thedirectionbecause it should be a permanently closed set. We could leave it open to allow for +2/-2 for “is really trustworthy” or similar. strongReffor thesubject, so AppViews can make their own choice as to how to handle a review of a record which has changed since it was issued.- I’ve explicitly made this one-adjective-per-record; I could also have made it one-review-per-record (containing an array of adjective/direction pairs) — I think this is the right approach, as my opinion about whether something “is funny” may change independently of whether it “is good”.
can be too small. I mean yes it’s clearly “good or bad” but I know I’ve definitely skipped giving reviews with those because my experience was murkier (for lack of a better word).