I built a memory store for AI agents: a Postgres database behind six MCP tools. It runs today as a local Claude Code plugin. The first post is about what it takes to share one memory across agents.
The paid, hosted version doesn't exist yet. Before writing any of it, I put up a waitlist page as a gate. I wrote down the number that decides whether I build it, 280 people confirming their email inside 28 days, and what each possible verdict makes me do, including stopping.
On day 28 the gate read TOO FEW VISITORS. Too few people had seen the page to reach 280, and the rules I wrote down answer that with one more round through different channels. This post is how I set the numbers, how a script turns them into a verdict, and what each verdict makes me do.
Every number below is computed from inputs I give, so you can redo it, and figures from outside link to their source.
Why a gate before the build
The checks on whether anyone wants the hosted version, a checkpoint at month 6 and a final decision at month 12, only fire after launch. Launch itself is at least 12 weeks of building away, so the earliest either could tell me no is six months after launch, with all of that building already spent.
I moved one of the two questions ahead of the build: does a pitch for the hosted version make people raise a hand? A waitlist page can test that for the cost of a page and a few posts.
It cannot answer the other question, whether people who try the product pay for it. That needs a working product, and the month-6 checkpoint still decides it. A confirmed email is a weaker signal than a payment. The gate can tell me that nobody raises a hand. It can't tell me that nobody pays.
What counts as a raised hand
The page shows the price before the form: $10 a month or $100 a year, with a 14-day free trial and no card. Someone who raises a hand without knowing the price tells me less.
Only confirmed signups count. Entering an address sends one email with a confirm button, and the signup counts once you click it. It is a button and not a plain link because mail scanners and link previews fetch URLs on their own. A plain link would let them confirm signups nobody agreed to, and that would inflate the one number that decides whether I build. A test pins this: opening the confirm link renders the button page and leaves the signup unconfirmed. The count itself only includes signups made after the gate opened, and a signup that its owner deletes stops counting.
The page is framed as build-if: "If 280 people confirm, this gets built," with a closing date. It shows no running count. A live counter would let visitors' reactions move the rate I am trying to measure, so the threshold is public and the total is not.
The gate runs for a window of 28 days, and the page has the date. That date can move, so the page never says "final" or counts days down, and its wording names the date and nothing else about time. Moving it is a one-token edit.
Where 280 and 1.48% come from
The gate rests on two numbers. The first is a floor on the share of visitors who confirm. I took Unbounce's median landing-page conversion rate for SaaS, 3.8%, and multiplied it by Mailchimp's report that 61% of people who start a double opt-in don't finish it, which leaves 39%. The product is 1.48%. It is a rough floor: Unbounce's data mixes click and form conversions, and neither figure is specific to waitlists. A page at or above it hasn't failed; it just hasn't been seen enough.
At 1.48%, 280 confirmed signups needs about 19,000 visitors, or about 675 a day for 28 days. My design spec says 18,893, because it used the unrounded 1.482%.
The second number, 280, is what I would need to see. At 5% to 10% of a list becoming paying customers, 280 signups is 14 to 28 customers. Those rates come from Lenny Rachitsky's 2022 piece on waitlist conversion. He asked founders "over Twitter, email, and IRL" and wrote that conversion averages "around 20% if you get people off the waitlist in under a month and falling below 10% if you wait over three months." That is an anecdote, not a study. The wait on my list would be close to that line or past it, so I used the lower range. 28 is my month-6 number for continuing.
A rate measured on few visitors is noisy, so the script also reads a 95% Wilson interval. A rate of 1.48% seen across 1,000 visitors has an interval of 0.90% to 2.44%, which straddles the floor. Zero signups put the whole interval below the floor only at 256 visitors. At 255 the script still can't read it.
I fixed these numbers before the data. The design spec holding them was committed September 14, and the constants went into code September 17. That is my record, and you can't check it, because both repositories are private. What you can check: post 1 on dev.to, published October 9, says "if 280 people confirm", and so does the Internet Archive's snapshot of the same post from October 10. dev.to lists a later edit date on post 1, but the snapshot doesn't depend on it.
One rule is mine, enforced by hand and not by code: changing a threshold after the window opens needs a written reason in the results document where I record each daily readout. It holds no such entry, and again that is my record.
A script reads the result
A script reads the database every day, starting on day one, so an early GO counts. It computes S, the confirmed signups; V, the visitors; r = S ÷ V; and the 95% Wilson interval for r. Then it checks these conditions in order and returns at the first one that holds.
| Order | Condition | Verdict |
|---|---|---|
| 1 | S ≥ 280 | GO |
| 2 | the follow-up round has closed | STOP |
| 3 | V = 0 | UNREADABLE |
| 4 | r ≥ 1.48% | TOO FEW VISITORS |
| 5 | the interval's upper bound is below 1.48% | PITCH FAILS |
| 6 | anything else | UNREADABLE |
Two choices in that order are deliberate. GO depends on S alone, and a test pins that 280 returns it whatever the visitor count. The two "no" readings are asymmetric: TOO FEW VISITORS reads the point estimate, so a rate at or above the floor, even exactly at it, says the pitch hasn't failed and the page needs more visitors. PITCH FAILS needs the whole interval below the floor, so a bad number from a small sample can't condemn the pitch. When the evidence is thin, the script says UNREADABLE.
These rows are hypothetical inputs run through the same verdict() function. None is a real reading.
| S | V | r | 95% interval | Round closed | Verdict |
|---|---|---|---|---|---|
| 300 | 4,000 | 7.50% | 6.72–8.36% | no | GO |
| 40 | 1,000 | 4.00% | 2.95–5.40% | no | TOO FEW VISITORS |
| 2 | 60 | 3.33% | 0.92–11.36% | no | TOO FEW VISITORS |
| 2 | 600 | 0.33% | 0.09–1.21% | no | PITCH FAILS |
| 1 | 200 | 0.50% | 0.09–2.78% | no | UNREADABLE |
| 150 | 6,000 | 2.50% | 2.13–2.93% | yes | STOP |
The third and fourth rows show the asymmetry. Two signups from 60 visitors is above the floor, so the pitch gets the benefit of the doubt even though the interval runs from 0.92% to 11.36%. Two from 600 is 0.33% with the whole interval under 1.48%, so the pitch fails.
V is the sum, over days, of distinct visitors to the front page. A visit to the post page is recorded but never counted, and tests pin both rules. A visitor who returns on another day counts twice, which lowers r. That can mislabel a no as PITCH FAILS when it should be TOO FEW VISITORS, but it can never turn a GO into a no, because GO is decided on S alone. I documented this and left it unfixed: counting people across days would need a cookie and consent, and would contradict the page's privacy claims.
STOP is the one reading I set by hand. It depends on a constant in the readout script that I flip when the follow-up round ends. The verdict logic has tests. The readout script that calls it, and holds that constant, has none.
What each outcome triggers
Every reading has an action, and the table below was in the same design spec as the numbers, before the window opened.
| Verdict | What I do |
|---|---|
GO |
I start building the following Monday. Every confirmed signup gets an email, and the page changes from "build-if" to "being built", with the form still open. |
TOO FEW VISITORS |
One follow-up round of four weeks, through different channels. |
PITCH FAILS |
One follow-up round of four weeks with a different pitch, aimed at independents who work for several clients. |
UNREADABLE |
One extension of two weeks. The page's date moves, which is why its wording names the date and nothing else about time. |
STOP |
Reached when S is still below 280 at the end of the follow-up round. Every confirmed signup is emailed that it won't be built, the list is deleted 30 days later, and lontar stays a personal tool. |
The STOP email says that and nothing more. It names no alternative and recommends no other product.
S carries over between the first window and a round, so a round adds to the signups already counted and doesn't restart the count at zero.
Three limits keep the gate from running on. There is at most one follow-up round, and a two-week extension counts as that round. The whole gate ends no later than eight weeks after the first post. Re-opening a STOP needs new evidence. These are process rules that I follow by hand, and no code enforces them.
Day 28, and the round it triggered
On day 28 the script returned TOO FEW VISITORS. That word says two things: the share of visitors who confirmed was at or above the 1.48% floor, and not enough people saw the page for the count to reach 280. It does not say the pitch is proven. The word reads the point estimate, and a rate above the floor with few visitors behind it still has a wide interval around it. What it does say is that nothing so far shows the pitch failing, so the page needs more visitors, not a different pitch.
I'm leaving out the counts behind that reading while the round runs. A running total would let reactions move the rate being measured, which is also why the page has no counter.
The action for that word is the one in the table above: one follow-up round of four weeks, through different channels. The page's date has moved to the end of the round, and the page has the date. This post is part of the round. It is about a different subject than my earlier posts, and it goes to the communities whose rules allow it, and to dev.to and LinkedIn as before.
Inside the round only the count matters. After the round the script can return only GO or STOP, and which one depends on whether S, carried over from the first window, has reached 280. Visitors who don't sign up change the reported rate, but they can't change the outcome.
Where this stands
The local build runs today as a Claude Code plugin, for one user: me. The repository is private for now, and it will be open source (AGPL-3.0). The hosted version doesn't exist yet. I haven't written any of it.
Whether it gets built depends on the gate above. If 280 people confirm a signup at lontar.dev, I build it. If the round ends short, lontar stays a personal tool and the list is deleted. The page has the date and the details. If you build with several coding agents and want one memory across them, you are who the page is asking.
If you take one thing from this post, it isn't lontar. Write the number down, and what each reading makes you do, before there is any data, and keep a copy somewhere dated that others can check. A threshold chosen after you have seen the traffic can be fitted to it. If you have run a gate of your own, I would like to compare notes: what you set, how you read it, and what you would change.
Top comments (3)
needing ~18.9k visitors to hit 280 at a 1.48% floor makes the gate a traffic test first. did you decide upfront what happens if round two comes back too few visitors again
Yes, that's decided in the outcomes table. There is exactly one follow-up round, and a two-week extension would count as that round. After it the script can only return GO or STOP, and S carries over from the first window. If S is still under 280 on the closing date (December 4), it's STOP: every confirmed signup gets an email saying it won't be built, the list is deleted 30 days later, and lontar stays a personal tool. "Too few visitors" can't buy a third round; only a reopened STOP with new evidence can.
You're right that it's a traffic test first. At the 1.48% floor, 280 needs about 18.9k visitors. A higher rate needs fewer, so the round's job is to get the page in front of more of the right people, and I fixed the cutoff before looking at any traffic.
fixing the cutoff before looking at traffic and emailing every signup on a STOP is the part most people skip. where are you planning to send people for round two