Your team's intelligence stays yours, and it compounds.

Rema helps any model, Claude, GPT, Gemini, Kimi, GLM or Llama, find where to look, cite the source, remember the fix and catch what's out of date.

Answers from what your team already wrote, each with its source.

Every fix becomes something the next one can find.

A fix your team writes down reaches the next question, with its source.

paraphrase45%

asked in different words, 20 questions

typo73%

asked with the word misspelled, 15 questions

two sources75%

the answer needs two sources, 12 questions

wrong premise80%

asked on a false assumption, 10 questions

direct88%

asked close to how the source puts it, 32 questions

negation90%

asked in the negative, 10 questions

Chart data
paraphrase45%asked in different words20 questions
typo73%asked with the word misspelled15 questions
two sources75%the answer needs two sources12 questions
wrong premise80%asked on a false assumption10 questions
direct88%asked close to how the source puts it32 questions
negation90%asked in the negative10 questions

The one person who knows is out.

Rema answers from the thread, the runbook and the ticket they already wrote, with each source.

#ingestChannel31 people
Hana4:12 PM

@Walt (Data platform, staff engineer) why is ingest-worker pinned to one replica? I want four for tonight's backfill.

Only visible to you

Walt Hendricks left in September. This account is deactivated.

1 reply
RemaApp4:12 PM

Walt wrote down why, in three places:

Slack, #ingest, March: “Dedup lives in each worker's memory. With two, a redelivered event can land on the one that never saw it and gets written twice.”

Ingest runbook: “Do not scale ingest-worker past one replica until dedup moves to Redis.”

Jira, the ticket he filed, updated 6 weeks ago: “Move the dedup cache to Redis so ingest can scale out.”

Nothing newer says the cache moved, so one replica is still the limit. A backfill at one replica is slower, not wrong.

Hana4:14 PM

Didn't know about the dedup cache. Keeping it at one for the backfill, thanks @Rema

SlackMarch 12

PriyaIndexes on events go in with CONCURRENTLY. The last plain CREATE INDEX blocked writes for 40 minutes.

ZoomWeekly sync

DanaAgreed: schema changes on the big tables ship in their own PR, never with app code.

ConfluenceMigrations runbook

CONCURRENTLY can't run inside a transaction. Give it its own migration with the transaction off.

GitHubReview on the last PR

TomaszSame as before: its own file, no transaction, and IF NOT EXISTS so a retry is safe.

Data platformGroup DM4 people
RemaAppThu 10:02

Hi @Lee (Infra), your agent is adding an index to events. The team builds those concurrently, in their own migration, so I gave it the note from @Priya (Backend) and the runbook first.

Your docs say 3 retries. Your team said 5, in Slack.

checkout · claude

❯ Try "write a test for retry.ts"

⏵⏵ accept edits on (shift+tab to cycle)

the payment webhook keeps failing, raise the retry budget

Your team never ships the default blue, and neither does your agent.

Mara, on the Figma frame

Welcome / Desktop

Mara Feb 4

Primary buttons are Stackline green. Never the default blue.

Without Rema

Welcome to Stackline

you@company.com
Create account

With Rema

Welcome to Stackline

you@company.com
Create account
web · claude

❯ Try "write a test for retry.ts"

⏵⏵ accept edits on (shift+tab to cycle)

add a sign up button to the welcome page

Your design doc says cache it. A newer comment says don't.

search · claude

❯ Try "write a test for retry.ts"

⏵⏵ accept edits on (shift+tab to cycle)

cache the search results endpoint, the design doc has the details

Your rule says every call goes through the rate limiter. Your team made one exception.

Tomasz in Slack, after the platform guide in Confluence
Confluence · Platform

Platform / Guides

Service guide

Owned by Ruth Calloway · updated last year

4. Rate limits

Every internal call goes through the shared rate limiter. No exceptions without a note in #platform.

#platformChannel40 people
TomaszMarch 21

Health checks go ahead of the limiter from now on. Last night it throttled them and the load balancer pulled every healthy pod.

What got built

src/server.ts+2
const app = express()
+// ahead of the limiter, see #platform
+app.get("/healthz", healthz)
app.use(rateLimit(limits.internal))
app.use("/api", api)
notify · claude

❯ Try "write a test for retry.ts"

⏵⏵ accept edits on (shift+tab to cycle)

add a /healthz endpoint to the notifications service, behind the rate limiter like everything else

Your queue delivers every event at least once. Your team dedupes on event_id, and so does your agent.

Lee, in Slack
#engChannel64 people
LeeSep 3

After Tuesday's double charge: the bus is at-least-once. Every consumer records event_id in processed_events, in the same transaction as its side effect, and skips one it has seen.

What got built

src/consumers/receipt.ts+6
+bus.on("order.paid", (e) => db.tx(async (tx) => {
+ const first = await tx.processedEvents.insert(e.id)
+ .onConflictDoNothing()
+ if (!first) return // seen it: no second receipt
+ await mail.receipt(tx, e.orderId)
+}))
notify · claude

❯ Try "write a test for retry.ts"

⏵⏵ accept edits on (shift+tab to cycle)

add a consumer that emails the receipt when order.paid fires

Your team removed the old auth check on purpose. Your agent was about to bring it back.

Priya's pull request, on GitHub

Remove requireSession, use withAuth everywhere

Mergedpriya merged 3 commits into main last Tuesday
requireSession reads the old session cookie, which stays valid after someone signs out of SSO. Every route uses withAuth now; it checks the SSO session on each request.

What got built

src/pages/admin.tsx+4
+import { withAuth } from "@/auth/withAuth"
+import { AdminPage } from "@/admin/AdminPage"
+
+export default withAuth(AdminPage, { role: "admin" })
web · codex

Ask Codex to do anything

GPT-6.1-Sol xhigh · ~/stackline/web

← for agents · ? for shortcuts

hey rema, add an admin page and protect it with requireSession like we always do

Your designer's video said three steps. Your agent built three.

Mara's walkthrough video, shared in the design doc
123
2:04

“Sign-up is three steps. Email, name, done. Everything else can wait until they're in.”

What got built

Sign up

  1. Email
  2. Name
  3. Done
web · claude

❯ Try "write a test for retry.ts"

⏵⏵ accept edits on (shift+tab to cycle)

build the sign up flow

From the engineers using it

“Some of your observations are far more complex than I would generally expect... not something I was aware anyone other than myself even knew about”
Principal engineer IIEA

Quote 1 of 4: Some of your observations are far more complex than I would generally expect... not something I was aware anyone other than myself even knew about Principal engineer II, EA.

Powering engineering teams at

Join them

300+ engineers in our first pilot

Why I'm building Rema

Diar Shakimov@diarshakimov1 ·

i asked Amjad Masad, the founder of Replit, why he thinks someone who started vibe coding 3 months ago might be better than someone who started 3 years ago.

my takeaway? some engineers have spent 3 years getting better at asking AI to do less.

  1. 1/

    The problem is scar tissue.

    And not knowing how to code might actually be an advantage.

    People who started with early AI learned all the things it couldn't do.

    So they built elaborate prompts, skill files, rules, and workarounds.

    The models improved. Their habits didn't.

    Amjad pointed out that even his own prompts have gotten lazier. Stronger models don't need the same micromanagement.

    I've been thinking about the ambition cost of this. Not just the time wasted on unnecessary workarounds, but all the things we stop attempting because an older model convinced us they weren't possible.

  2. 2/

    Here's where I disagree slightly.

    Knowing how to code is an ENORMOUS advantage.

    But only if you can resist the urge to micromanage implementation.

    The best engineers should be able to delegate an entire outcome, recognize when the AI is wrong, and verify the result.

    Their advantage isn't writing more code. It's knowing what good looks like for the use case.

  3. 3/

    Now apply this to an entire engineering team.

    Your models get smarter every few months.

    But your company's knowledge? It gets buried in Slack, GitHub, old tickets, outdated docs, and decisions nobody remembers making.

    So your agents inherit yesterday's assumptions, while your engineers keep repeating yesterday's explanations.

    I wonder how much of the next leap in productivity comes from better models, and how much comes from finally questioning the limits we've gotten used to.

Read on X

Still not sure? Ask your AI.