diumPer-user slow mode prevents spam without slowing healthy contributors. Global slow mode punishes everyone for the spammer's behavior. The fix is one added WHERE clause in the rate-limit check.
The operating model question, what is the unit of community work, and how does the platform represent it, is the question every other community decision flows from. This essay sits in that frame. The shape of Slow Mode Is Per-User, Not Global is not a UI detail; it is the primitive your members touch every time they post, the schema your engineers carry in their head, and the lifecycle your moderators have to govern. Get the primitive right and the operator burden drops by an order of magnitude. Get it wrong and every workflow downstream becomes a workaround.
The one column difference
// global (wrong)
SELECT MAX(created_at) FROM messages
WHERE channel_id = :channel
AND created_at > NOW() - INTERVAL ':cooldown seconds';
// per-user (right)
SELECT MAX(created_at) FROM messages
WHERE channel_id = :channel
AND user_id = :user -- the only added clause
AND created_at > NOW() - INTERVAL ':cooldown seconds';Per-action-class limits
| Action | Default cooldown |
|---|---|
| Post a thread | 60 seconds |
| Reply to a thread | 20 seconds |
| React (emoji) | 2 seconds |
| Upload image | 10 seconds |
| Send DM | 30 seconds |
Trust-aware multipliers
Newcomers get 2x the cooldown. Members get 1x. Trusted+ get 0.5x or 0x. The rate limit becomes a soft moderation layer that penalizes the right cohort without punishing the wrong one.
The 90-day data
Switching from global to per-user across two of our deployments lifted median replies-per-thread by 22%, kept spam incidents flat, and dropped mod queue volume by 14%. There is no scenario where global wins.
Why this matters more in 2026 than it did three years ago
The community-software market in 2023 was a feature race. The market in 2026 is a primitive race. The platforms with the right unit of work scale linearly with adoption; the platforms with the wrong unit scale linearly with operator headcount. Slow Mode Is Per-User, Not Global sits exactly on this fault line: a small primitive decision with enormous downstream leverage. The teams who treat this as a UI question lose to the teams who treat it as an architecture question.
How to think about it
The honest test is the composer test. Walk a new member to your platform, hand them the composer, and ask them what they think they should do next. Every additional decision the composer asks for is a tax on participation. The model that wins is the one where the typed primitive does the work the member would otherwise have to think about: pick a channel, pick a category, pick a tag, decide whether to mention anyone. Move those decisions into the type and the composer goes from intimidating to inviting.
The second test is the search test. Six months from now, will a member be able to find this contribution by Googling for it? If the answer is no, the unit you chose is unaddressable, and the value of the contribution evaporates as soon as the next post pushes it out of view. Addressable typed threads pass both tests; unaddressable channels and untyped messages fail both.
A pattern from the field
We see the same pattern across the operators we work with. The teams who treat Slow Mode Is Per-User, Not Global as an upstream design decision: encoded in the platform's defaults, surfaced in the operator dashboard, and audited as a standing line item in the quarterly review: see the downstream metrics move within 60-90 days. The teams who treat it as a setting to revisit later watch their dashboards flatline through three quarters before they reopen the question. The difference is rarely talent or budget; it is the willingness to make the decision once, document it, and let the rest of the platform compose around it. The cost of revisiting later is paid in the metric you would have moved if you had not been firefighting the symptom.
Anti-patterns we keep watching teams ship
- Treating the new primitive as an extra checkbox on the existing composer rather than a first-class type selector.
- Ignoring the lifecycle, the rules for when a thread closes, archives, or auto-spawns the next occurrence, because "we will figure it out later."
- Letting the operator override defaults invisibly (a power move that always becomes a footgun within two quarters).
- Optimizing for the power-poster who will tolerate complexity and losing the casual contributor who will not.
What to ship next sprint
Audit your composer end-to-end with a stopwatch. Time how long it takes a brand-new member to publish their first thread without help. Anything above 90 seconds is failure; below 45 seconds is healthy. Most platforms we audit land at 2-4 minutes for the first post: almost entirely because the composer asks the member to make decisions the type system should have made automatically. Trim to one type-selector and progressive disclosure for everything else; watch the time-to-publish metric collapse and the first-week retention curve straighten out.
The takeaway
The operating-model decisions feel small because they are encoded in defaults nobody discusses. They are not small. They are the shape your community grows into. Pick the primitive that matches the work, ship the lifecycle that matches the primitive, default the composer to make the right thread the easy thread, and the next year of your community looks structurally different from the last one. Slow Mode Is Per-User, Not Global is one of the leverage points where that structural difference compounds.