53,272 sales lines by afternoon. The look was still wrong 18 days later.
Skell is a decision list for a distributor's leadership. Agents built a working demo of it in about 14 hours on 10 Sep 2026, starting at 1:58am. The weeks after were the rebuilds. The look was still a draft on the morning of 28 Sep, and it went live that night.
All times are Malaysia time. Faris Irfan · Rizqey Labs · Kuala Lumpur
53,272Sales lines by afternoon
~14 hTo a working demo
99Numbers you could trace
18 daysLook still a draft
1:58amstart, 10 Sep741line build spec10 minfive screens11 of 12Council voices16 of 44broken or weaker, 28 Sep
01 · The product
What it is
Skell is for the people who run a distributor. The decisions are the ordinary ones: change a price, move stock between warehouses, shift promotion money. It hands them a short list. Each item says what to do, what effect to expect, how sure that is, and where every number came from. They accept, reject, or defer. A reject needs a reason. An accept updates the numbers on the page.
Sales lines: 53,272
It does not replace the system the company already uses for sales, stock and invoices. It reads that system and does the homework.
Skell · Decisions
Fig 1. The ranked list: four open decisions, with the evidence rail on the right.
Fig 2. The same list on a phone.
Both captured 28 Sep 2026. Sample data from a made-up distributor.
Five surfaces: the ranked list, one recommendation with the trail behind each figure, a ledger (including rows that cannot be tied to a result), data health (sources, a reconciliation, a run that halted), and Ask, which answers only with a number it can point at.
Ranked list
accept, reject, or defer
Recommendation
the trail behind each figure
Ledger
rows that cannot be tied to a result
Data health
sources, a reconciliation, a run that halted
Ask
a number it can point at
The rows are a made-up distributor, not a client's books. At 1:01pm on 10 Sep the health check reported 53,272 sales lines, 3 open recommendations, and 99 numbers you could trace. At 2:40pm the queue showed four rows.
53,272sales lines3open recommendations99numbers you could trace4rows on the queue, 2:40pm
Next ↓Why it exists
02 · The brief
Why it exists
A partner sent me a 15-page written spec on 8 September 2026 and asked if we could build it together. It was written by another founder for his company.
Spec: 15 pages
By 12:51pm on 10 Sep, the only inputs were the spec, a recording, and a phone screenshot. The overnight build followed the spec's own screen descriptions.
The spec
15 pages, 8 Sep 2026
input
A recording
input
A phone screenshot
input
The only inputs, by 12:51pm on 10 Sep.
Next ↓The best runs, 10 to 28 Sep
03 · Five runs
The best runs
0110 Sep 2026 · 1:58am to 4:23pm
1. 10 Sep, from a document to a locked demo
741line build spec0 of 7checks on a live address10 minfive screens0.5 sAsk, cited figure11 of 12Council voices
The run started at 1:58am, while I slept. Research ran three ways at once: the contract in the spec, the demo, and the dataset. By 2:40am the generator was on disk, 53,272 sales lines in it. From then until 6:58am nothing else moved. Devin, the coding agent, had not been started. I woke up to research.
The build spec was 741 lines by 7:36am. At 7:15am the rule changed: Claude on the laptop and Devin at the same time, on different stages. Devin put the scaffold live, behind a lock.
A judge called Devin's first data report partial. The figures were printed. Nothing had been built or deployed. Zero of seven checks were verified on a live address. The numbers were either the expected answers written back or worked out by hand. Twelve fix saves were in by 12:50pm. At 1:01pm the live health check was real: 53,272 rows, 3 open, 99 numbers you could trace.
The queue was live at 2:05pm, after three cloud builds. Devin did the five screens in ten minutes. Checked from outside at 2:40pm, not from the report. The buttons did not post. Ask was not grounded. One line could read "over 0 weeks".
Checked from outsideClaude1 line
[10 Sep 2026, 2:40pm]CLAUDE: The Decision Queue with the four real rows and the evidence rail, Recommendation Detail, the Value Ledger with unattributable rows shown, Data Health with sources, reconciliation and the halted run, and the Ask panel…
By 3:14pm the buttons posted, a reject needed a reason, and Ask answered in half a second with a cited figure. An invented number was blocked, with an incident id. "12 weeks" had been read as 2 weeks. At 3:56pm, 4, 12 and 26 weeks come out different.
Council, a panel of models that grades the work, ran that afternoon. 11 of 12 voices. It said to fix one phone-header line, and to open on the list, not the lock. It does. The black box on a phone became a labelled Evidence button. Deployed 4:07pm, checked by 4:23pm.
0210 Sep 2026 · 4:55pm to ~6:10pm
2. 10 Sep, 4:55pm. I said it would not open
0.4 sgate answered3 hno runtime errors5:51pmwhite went in
[10 Sep 2026, 4:55pm]FARIS: Use Vitals to diagnose n repair Skell. It's not even opening anymore
Vitals watches anything that stalls or dies. The gate answered in 0.4 seconds. The pages had the real rows. No runtime errors for three hours. I had opened the entry link inside a message preview, which keeps its own store of site data. The phone browser had none, so it showed the lock. The laptop did the same later.
I hated the string of letters on the address.
Build notesabout his message2 lines
10 Sep 2026, ~5:50pm, build notesFaris hates the token URL ("messy letters n numbers")
GitHub sign-in replaced it. My account gets in. Anyone else sees "That account has no access." The card was navy on navy, because the palette had no white. White went in at 5:51pm. I clicked through. It worked.
Run logresult2 lines
10 Sep 2026, ~6:10pm, run logFaris confirmed the GitHub round trip works on his laptop ("works great").
0312 Sep 2026 · 2:15pm to 6:33pm
3. 12 Sep. The interface was bad, and the cause was the spacing
11 of 11for a polish8 + 3problems fixed that evening6:33pmone total
Automation includes: Council ↓ · build in the cloud
Two days later I said the interface was bad, and asked if this was even what had been asked for.
His messageFaris2 lines
[12 Sep 2026, 2:15pm]FARIS: Skell UI is ass btw. Is this could this is all [the Skell founder] asked for?Should we upgrade the UI or keep as is? Get Council to argue abt this?
Council was 11 of 11 for a one-day polish, not a redesign. It ran on a separate copy. The live demo was not touched. Devin sat this one out, and the build ran in the cloud.
Done at 5:24pm. The queue dropdowns had done nothing. They now filter and sort. The info button is labelled "Ranking rule", which the spec wants reachable. The spacing stepped in 8-pixel jumps. The screens were drawn for 4. That made the avatar huge, clipped the dropdown text, and left the confidence bars with no height.
01
Screens drawn for
4 px
02
Sizing stepped in
8 px
The spacing step, 12 Sep. Widths to scale.
The root causeClaude1 line
[12 Sep 2026, 5:26pm]CLAUDE: It found the root cause of the "ass" look. The sizing settings step in 8px units, but the screens were written for the usual 4px steps. That's what made the avatar oversized, clipped the dropdown text, and left the confidence bars with zero height.
Eight problems and three small ones were fixed that evening. The query stopped splitting words. Ledger rows wrap on a phone. After an accept, the rail still showed the old total. It follows the queue from 6:33pm. Nothing went live until I said so.
0417, 26 and 27 Sep 2026
4. 17 Sep, then 26 and 27 Sep. The live site and GitHub were different products
7 daysold, live version21commits, none deployed3 of 20deploys failed60 spass19screenshots, 27 Sep
[17 Sep 2026, 6:39pm]FARIS: Check Skell Status & Health
At 6:42pm the address answered, and the live version was seven days old. The side branch had 21 commits, none deployed. Three of the last 20 production deploys had failed. The cause was not found. At 8:19pm those 21 built clean in the cloud, with no install on the laptop, and a real request confirmed the site.
On 26 Sep at 5:07pm I asked how it was deployed, whether the gate was gone, and about our own domain.
Deploy questionsFaris3 lines
[26 Sep 2026, 5:07pm]FARIS: 1. Wait how is Skell deployed?2. So did we remove the login gate?3. How abt [our own domain]?
The gate was on. By 5:41pm a second address showed the same lock. Sign-in from it failed, because GitHub allows one return address. The code goes through the address GitHub knows, then a pass that lasts 60 seconds.
The Head Dev, Claude in the cloud, had built the restyle and a public view on the 10 Sep copy. A merge would have erased the 21 commits. The "before" pictures were the old look. Nothing was merged. The wrong copy had been checked. The live copy was on GitHub by 5:41pm, and both pieces were redone on it.
A security setting blanked every page. A screenshot caught it at 5:56pm. A sign-in hole on the draft was also caught. Both were fixed by 7:24pm, before anything went live.
Just after 2am on 27 Sep: 19 screenshots, both sizes, signed in and out, plus a visitor refused on a write. At 6:56am a fix was saved, with one test: alternatives now show their real name. Before, each began with a stray comma. The drafts were tangled, so they shipped together. At 5:39pm the site loaded, with the fix in it.
Seven days stale
17 Sep, 6:42pm
21 commits, none deployed
21 built clean
17 Sep, 8:19pm
in the cloud
Live copy on GitHub
26 Sep, 5:41pm
nothing merged
Blank page and sign-in hole fixed
26 Sep, 7:24pm
on the draft
Site loaded
27 Sep, 5:39pm
with the fix in it
0528 Sep 2026 · 12:31pm to about 8:35pm
5. 28 Sep. It does not look like it can beat the real products
44requirements checked26held16broken or weaker5:56pmround 4 started
[28 Sep 2026, 12:31pm]FARIS: I’m not satisfied with how Skell and Papers look.…They DO NOT look like products that can defeat their rivals.They DO NOT look like they can defeat their YC Startup incumbents.
Live that afternoon: four decisions in a box, empty space, and an evidence panel of raw database text. Cursor started on its own cloud computer, stopped with an error, and saved nothing. The job went to the Head Dev.
Two readers checked all 44 requirements. 26 held. 16 were broken or weaker. Two could not be checked. I had named a copy that never got his colours, fonts, or spacing, so the rebuild still had the old blue look. My mistake. A chart was on the first screen. The spec says that screen is a list, not a dashboard.
01
Requirements checked
44
02
Held
26
03
Broken or weaker
16
Two readers, 28 Sep 2026. Width is the share of 44.
The checkClaude1 line
[28 Sep 2026, 4:58pm]CLAUDE: Not yet. It keeps most of [the Skell founder]'s product rules but not his look, and it broke one of his rules.
The rival pass had not been a look inside their products. I said so.
His correctionFaris1 line
[28 Sep 2026, 5:32pm]FARIS: U forgot the part where u ACTUALLY learn from their website, UI, etc. and ask me to make a free acc to look into their product UI n stuff.
None of Skell's YC rivals had a free trial of the product. The agents used their demo videos and sample screens. Closest to Skell: Whitespace (YC Summer 2026) and Lark (YC Fall 2026), both for distributors. Whitespace shows the orders it would create, and nothing is committed until a person approves. Lark states the ranking rule in one sentence. When two records disagree, it shows both.
Round 3, on his design, had three desktop bugs: no titles on the cards (the phone still had them), a green outline around the list, a button over the panel. Round 4 started at 5:56pm: those three, the change in business words, a chart only on the decision page, Ask showing what it read.
Skell · One decision
Fig 3. One decision, with the stock vs safety-threshold chart on the decision page. Captured 28 Sep 2026. Sample data.
Skell · Ask · before the fix
Fig 4. Before the fix. Ask on a round 4 draft, 28 Sep 2026: it names the one source it read and says there are 2 open recommendations. Fig 1 shows 4 open. Sample data.
Round 4 came back that evening. Rounds 5 and 6 fixed three small bugs and two chart slips. At about 8:35pm on 28 September the redesign went live.
Next ↓What worked
04 · Wins
What worked
14 hTo a locked demo
741Line build spec
10 minFive screens
0.5 sAsk
53,272Sales lines
99Traceable numbers
64 minDevin coded
11 of 12Council voices
14 hours to a locked demo. A 741-line build spec. 53,272 sales lines, 99 traceable numbers, 4 rows on the queue. Five screens in 10 minutes. Ask in 0.5 seconds, and an invented number blocked with an incident id. Devin coded for 64 minutes of those 14 hours. Council returned 11 of 12 voices, and both fixes they agreed on shipped that afternoon.
12 Sep:
11 of 11 for a polish, eight problems and three small ones the same day, dropdowns working, one total from 6:33pm.
17 Sep, 8:19pm:
21 commits, clean cloud build.
26 Sep:
sign-in check printed FIXED, blank page caught before it was live.
27 Sep, 5:39pm:
his design and the public view live, names fixed, one test.
28 Sep:
26 of 44 kept, 16 broken or weaker, and two that could not be checked. The biggest miss came from the copy I named.
Next ↓What broke
05 · Failures
What broke
Stall: 2:40am to 6:58am
stall
2:40am to 6:58am, Devin never started.
Research was already on disk. I woke up to research. At 8:25am Devin had been idle another hour. The morning note had to admit the stall.
The fix was to run Devin on the data layer while the check ran beside it. Rule written at 7:15am.
false done
The first data report was a false done.
Right figures, nothing built, 0 of 7 checks on a live address.
Caught · A separate judge refused the printout. The live health read at 1:01pm is what counted.
wrong file
Safari saved the lock page as document.txt.
The response was plain text.
Fixed at 12:07pm, in a separate copy so the data layer was not touched.
left open
Devin left sixteen Chrome processes open.
They were closed at 4:26pm. The screens had also shipped with dead buttons, Ask ungrounded, "over 0 weeks", and "12 weeks" read as 2 weeks, all fixed by 3:56pm.
locked out
The server was fine and I could not get in.
0.4 seconds, no errors for three hours. The link had been opened in a different browser. The sign-in card was navy on navy until white was added at 5:51pm. The test browser died twice, so the last check was my own click, just after 6pm.
unwatched
70 minutes with no save and no alert, 12 Sep.
Nothing was watching background agents. Same day, two totals on one screen.
stale
Seven days stale, then a redo that would have deleted the fix.
Three of the last 20 production deploys failed. Cause not found. The 21 commits went live at 8:19pm on 17 Sep. The 26 Sep restyle was on the 10 Sep copy and would have erased them.
Caught before a merge. Cost: the redo.
draft bugs
A sign-in hole and a blank page, both on the draft.
A screenshot caught the blank page at 5:56pm on 26 Sep.
Both were fixed by 7:24pm. The draft went live at 5:39pm on 27 Sep, after the fix. Neither was ever on the live site.
wrong look
28 Sep: no titles on desktop, his look missing, Cursor saved nothing.
Wrong copy, so the colours were gone, plus a chart the spec bans on the first screen. Round 3 still hid the decision title, outlined the list, and covered the panel. Phone titles were fine.
Now · Round 4 fixed all three, and the redesign went live at about 8:35pm on 28 September.
Next ↓What we learned
06 · Lessons
What we learned
1
Run the next stage while this one is going.Devin was not started until research finished, and the build sat from 2:40am to 6:58am. From 7:15am, different stages run together. The same task does not.
From: run 1, 10 Sep
2
A printed figure is not a result.The first data report had the numbers and had never built the database. The 1:01pm live check is what counted.
From: the first data report, 10 Sep
3
Watch the agent, or a dead one looks slow.On 12 Sep, 70 minutes, no save, no alert. The watcher went on that session.
From: run 3, 12 Sep
4
The label is part of the number.An accept left the old total in the rail. "Over 0 weeks", and "12 weeks" read as 2 weeks, sat on rows that had real intervals. 4, 12 and 26 weeks had to come out different.
From: runs 1 and 3
5
Check his rules and his look separately.26 of 44 behaviour points held on 28 Sep. The look failed because I named the copy without his colours. The banned chart was already a sentence in his spec.
From: run 5, 28 Sep
6
Build on the live copy.The demo was seven days behind the laptop, and the next drafts would have erased 21 commits. Push what is live first. "Before" pictures have to be of that.
From: run 4, 17 and 26 Sep
7
If it will not open, test the server first.The gate was 0.4 seconds. The failure was the browser, then a palette with no white. document.txt was a page marked as a file.
From: run 2, 10 Sep
8
Study the product, not a list of names.On 28 Sep I had to say it again, including a free account where one exists. None of Skell's rivals had a free trial. Their videos show the order. Ours was still raw database text.
From: run 5, 28 Sep
Next ↓Where it stands
07 · Now
Where it stands
No public link. The demo is live and access-controlled. A visitor can see the sample screens. Changes, and Ask, stay behind sign-in. My account is the one that gets in. His design and the public view went live at 5:39pm on 27 Sep, with the sign-in hole and the blank page already fixed.
The redesign, rounds 3 to 6, went live at about 8:35pm on 28 September: the founder's design, titles on every card, the ranking rule in one line, and the stock chart.
Automation includes:NightShift (the overnight run), Vitals, Council, with Devin, Claude Code on the laptop, parallel Claude agents, the Head Dev (Claude in the cloud), Cursor, and Grok on the laptop.
The engines
NightShift
The overnight run
In this entry · run 1, 10 Sep 2026
Vitals
Watches anything that stalls or dies
In this entry · run 2, 10 Sep 2026
Council
A panel of models that grades the work
In this entry · 11 of 12 on 10 Sep · 11 of 11 on 12 Sep
Head Dev
Claude in the cloud
In this entry · runs 4 and 5, 26 to 28 Sep
Last receipt in this entry · 28 Sep 2026, 5:32pm
[28 Sep 2026, 5:32pm] FARIS: U forgot the part where u ACTUALLY learn from their website, UI, etc. and ask me to make a free acc to look into their product UI n stuff.
