Every app looks the same. This is where the way out is.
Same screens, same chat box in the middle, same words. That's not a technology problem, it's an experience one — and it happens because everyone is looking at each other instead of at people. This skill goes and reads how real people are actually talking right now, so what you build next comes from somewhere other than your competitor's homepage.
Here's the part that makes it possible: people are telling you what they care about, constantly. Not in surveys. In a video someone filmed at 1am, in a joke that's funny because it's true, in a comment thread nobody was performing for. Almost never in the words you'd think to search.
Nobody says “I'm lonely.” Nobody says burnout, or that they're stuck between the life they expected and the one they've got. They say something sideways, and they say it constantly, and it's all sitting there for free.
It reads culture, not your customers — if you need to know what your users think, go and ask them. It reads English-language, mostly online, and it reaches people who write under their own name more easily than people venting anonymously, because most of the open forums have shut their doors to machines. It tells you which rooms it couldn't get into, every time.
Why bother
Because the alternative is shipping the thing that looks like everything else, and hoping. Emotional resonance isn't a soft goal you argue for at the end of the quarter — it's the difference between a product people stay with and one that reads as generated. If you don't want yours to feel like AI slop, this is the work.
- Search your category, find your competitors. Type the words your industry uses and you get vendor blogs and search-bait, because those are the only people who write them. Nobody with the actual problem talks that way.
- Your agent already knows — from last year. Ask any AI how people feel about something and it answers confidently from a training cutoff. Nothing about the answer will look wrong.
- Research that changes nothing. A lovely summary of the discourse that leaves you exactly where you started, with no idea what it means for the screen you're about to write.
How it works
Five steps. The second one is the whole trick.
- 1.Admits what it doesn't know
Its own knowledge stops at a date. It says so, and treats that as a starting position rather than an answer. How far it turns out to be off is itself a reading of how fast your world moves.
- 2.Works out where the feeling actually lives
Not your category's words — those belong to whoever's selling into it. It works out who really has this problem, the moment it bites, and what they call it. Then it goes looking for evidence that it's got that wrong, because a hunch you only ever confirmed is still a hunch.
- 3.Goes and looks, live, every time
Never from memory. And if what comes back is marketing and search-bait, it changes where it's looking rather than writing up what it found.
- 4.Reads the good critics — afterwards
People like Jasmine Bina, Eugene Healey, Zoe Scaman, Ana Andjelic and Alex Murrell, but only once the looking is done, so their frames don't decide what it notices. Kept separate from what ordinary people said, and credited by name.
- 5.Hands you two or three things, in plain words
Each with the quotes that produced it, and how many people said it. About two screens, so you can read it between two meetings. Then it stops and lets you pick what to pull on.
- 6.Gives you the rest when you ask
The choice a finding forces, what each direction costs, how long it holds, and the pass it ran to try to prove itself wrong. That's the part worth acting on, and it lands once you've said which finding matters to you.
Nothing gets saved to your machine unless you ask. The read lands in the conversation; if it's worth keeping, it asks first.
How sure it's allowed to sound
The most useful thing here is also the least obvious: how confident the writing is permitted to be is set by how much evidence it actually found, not by how good the sentence would be. So it tells you what it counted, right next to the finding, and you decide how far to lean on it.
| What it found | What it says |
|---|---|
| Several people, in different places | “This held up everywhere I looked.” |
| Two people, or three in one place | “Two people said this, both on the same forum.” |
| One conversation | “One thread, twelve replies, and nothing like it anywhere else.” |
| A real argument both ways | “There's a real argument here, and it turns on X.” Both sides reported, never averaged into a verdict. |
| Almost nothing | “The field was quiet. Here's what would answer it.” And it stops there. |
It won't write your copy. Ask what a finding puts at risk and it names the choice that forces, then leaves the choice with you. Ask it for drafted options and it will give you some.
Install it
› /plugin marketplace add hgambas/think-like-a-pm
Then /plugin install think-like-a-pm, which brings
this skill along with the others. It starts on its own when you're naming,
writing or positioning something, or you can call it by name.
Where you come in
Once, near the start, it shows you where it's about to point the research — who it thinks this is really for, and the words it's going to search instead of yours. That isn't a form to approve. It's a position you can argue with, and it starts searching straight away either way. If you correct it, it folds that in and tells you what changed.
The other moment is at the end, when there's a decision the evidence can't make for you. It puts the choice in front of you with what each direction costs, and leaves it there.
There's one place it will stop and ask: when the field is silent on the thing you actually care about. You know things search can't return — something a customer said, something you saw on a call. That's worth more than five confident paragraphs, and it will ask for it rather than fill the gap.
Keep your own list
General search is polluted enough now that a hand-kept list of people worth
reading beats searching for them. The skill keeps one for you at
~/.claude/read-the-room/sources-i-trust.md — drop in anyone whose
writing you rate, and it reads them first on every run. That file is yours;
updating the skill never touches it, and it never edits it unless you ask.
There's nothing to learn
No commands, no syntax. You say “read the room” and then describe what you're doing, the way you'd say it to a colleague. Your agent works out the rest.
| You say | What happens |
|---|---|
| “…on our pricing page” | Reads your page first, then goes out |
| “…on how people feel about X” | Straight out to the topic |
| “Does effortless still land?” | Same work, one phrase |
| “Has anything moved since?” | Diffs your saved read |
A skill about how the industry's words fail real people has no business handing you a vocabulary to memorise.
On something you've made
A page, a screen, a name, a whole position. You point at it and say what you're doing. There's nothing to learn: “read the room on our pricing page” is the entire syntax.
What it does with it
First it reads your thing properly. Every word: the headline, the buttons,
the browser tab title, the line that shows up when someone pastes the link.
Then it looks at what sits around it, because a page with ?v=b on
the end is one of several and the differences are usually the interesting bit.
Then it closes your page and goes outside. It isn't reviewing your work. It needs your exact words so that when the field says something, there's something real to hold it against.
What comes back
The most surprising thing it found about the world, in the first line or two. If you asked something direct, that gets answered first. Then two or three things the world is doing, each with the quote underneath and how many people said it. Last, one question only you can answer, and the rooms it couldn't get into so you know what's missing.
Roughly two screens. Ask about any of the findings and it gives you the choice that finding forces, what each direction costs, and how long it's likely to hold.
Try it
› read the room on our pricing page
One word is a fair question too
“Does effortless still land?” is a real thing to ask, and it does the same work, scoped small. It comes back with one of five answers: it lands, it's dated, it's saturated so it now says nothing, it collides with a meaning you didn't want, or the field was too quiet to tell.
If it isn't a keeper, it offers alternatives it actually found in someone's mouth, with the quote attached. It won't invent a replacement, because a word nobody said is worth nothing.
Pitfalls
- Asking it about your own users. It reads the wider world. What your customers think is a different job, and a better one to do by talking to them.
- Expecting a verdict. It hands you the room and the choice. Ask it to pick and it will, and it will tell you how sure it is.
- Treating a quiet field as a quiet market. Sometimes nobody is talking because the doors are shut, not because there's nothing to say. It tells you which.
On a topic
No page, no product. Just: how are people actually feeling about this right now? This is the one to use when you're about to start something, or when you've been in the same category so long you've stopped being able to hear it.
Where it goes looking
Not your category's keywords. Those belong to whoever is selling into it, and searching them gets you competitors and search-bait. It works out who really has this problem, when it bites them, and what they call it between themselves. Then it goes there.
Before it trusts any of that, it goes looking for evidence it's got the wrong people. That's the difference between reading a room and inventing one.
Try it
› read the room on how people talk about meal planning now
Pitfalls
- Asking too narrow a question. “How do people feel about our category” gets the category's own voice back. “What's it like to be the person who has this problem” gets you somewhere.
- Taking one loud thread as the room. A single scornful post is a single scornful post, and it will say so rather than turn it into a rule.
Months later
Culture moves faster than product ships. A call you made against a read from four months ago might be standing on ground that isn't there any more.
How it happens
Mostly you don't have to ask. If you saved a read on this topic and it's still recent, your agent notices and picks it up rather than starting over, and says so in a line. You can also just ask: “has anything moved since we looked at this in March?”
It needs a saved read to compare against, and reads are only saved when you ask for one. If there's no file it will say so and offer to start fresh, rather than reconstructing a baseline from memory. That's the exact thing this skill exists to prevent.
What comes back
Four buckets and nothing else in the headline: what held, what moved and which way, what died, and what's new. Then the part that matters. The decisions you made on the old reading, and whether they still stand.
Try it
› has anything moved since we read this in March?
Pitfalls
- Going back too soon. Every read says how long it's likely to hold. Six weeks and a year are very different answers.
- Reading the movement and stopping there. What moved is interesting. What it broke is the point.