Dossair finds programmes, drafts the answers, fills the forms and keeps each stage. From 29 August to 28 September 2026 it held 91 programmes and 25 applications, and 10 forms went out. Most of what it taught us came from the ways it went wrong.
All times are Malaysia time.
Dossair is the applications agent. It is for a founder who applies to accelerators, grants, hackathons and competitions and does not want to retype the same facts into every form.
It keeps one set of company facts, opens the real form, drafts the answers in that form's own order, and stores each stage so a later form does not wipe the first one. I still approve what goes out, and I press submit.
Next ↓Why it exists
I was applying to a lot of programmes and the answers lived in chat. On 29 August I asked for one place where the copy could be tapped, with the established answers already in it.
The first build was one long page. I said Claude was ignoring the ask, and I wanted one page per programme.
On 1 September I had to say it again: one application had been given its own refill process. Put it in the module.
On 14 September I called Dossair the product version of Blitz and Sift, two of our finders.
Claude corrected me the same hour. Sift finds accelerators, Blitz finds hackathons, and Dossair was the missing file, plus a record of what was sent. I then wrote the definition in the run below.
Next ↓The best runs
Automation includes: Claude Code
I asked for the module on 29 August, and in the same note I said the console was down. Claude checked. The site was up. The network I was on was blocking it. Then it built shared company facts, plus a block per programme already in progress. Each block could be tapped to copy, with a character count.
I rejected the layout the same day. One scrolling page was not usable.
By 1 September a form was only marked submitted once I had actually submitted it. On 2 September Claude put one programme's form up with all 26 fields in the form's own order. 24 were filled. Two were only mine: how I heard about it, and whether I wanted their interview step.
[29 Aug 2026, 8:56pm] FARIS: …Make an Accelerators Module in Rizqey Console. Form copy should be easily copied with a tap. Put all established form copy in there
[2 Sep 2026, 11:35am] CLAUDE: All 26 fields are there in the form's own screen order…
Automation includes: Claude Code
AltaLab accepted the application into its Fall 2026 cohort on 5 September. On 8 September I asked for a split that only makes sense once you are in. One side keeps the application exactly as it was sent. The other tracks the work the programme gives you. Claude's first pass did not put the work there. I said so.
[8 Sep 2026, 9:58am] FARIS: Nah man did u bloody build what I asked u to build in Rizqey Console Accelerator module? …
The second pass, the same day, listed every task released so far, with a section left for me. Chapters not yet released stayed as empty stubs, not guesses.
Automation includes: Codex, Claude Code, NightShift, Council
On 13 September I said we likely did not need a new domain. The product could sit under rizqey.com. I had spelled its name two ways, then told Claude I meant its pick, Dossair. A play on "dossier".
Codex built the first version that night. Claude ran the tests itself. Both sets passed, 6 of 6. The build could make a company file, sign it, hand a receiver only the fields it asked for, and make a receipt the receiver could check. It mapped two real programmes and ran a pretend submission of our own file. Nothing was sent.
A real organisation accepting that file was still open: one receiver, ten submissions that receiver had checked, then a second receiver. That never happened.
On the morning of 14 September, while I slept, NightShift, our unattended job runner, shortlisted 12 programmes, set up the top two, and drafted two emails. It did not send them. I asked what its update meant. Claude's advice was not to send. A cold email was not going to make a programme adopt a new file format. The emails stayed unsent. Both drafts were also missing facts the forms ask for.
That afternoon I locked the definition, and Codex wrote a 14-page product document. Council, a panel of models that grades work, reviewed it before it was treated as final.
[14 Sep 2026, 2:07pm] FARIS: Ok so Dossair is the ultimate competition research n application Agent right? Any competition type, Hackathon, Accelerator, as long as it has a prize and you need to apply, it’ll find good ones for you and apply, or just apply for the ones you want it to.
[13 Sep 2026, 11:01pm] CLAUDE: Dossair's first version is built, and I ran its tests myself. Both sets pass, 6 of 6 each.
Automation includes: Claude Code
One programme sent us a second form. The first application, from 31 August, had 35 answers. I caught the new copy on a job card and asked why it was not a new stage on the same entry, and how an update would avoid deleting the first form.
Claude saved a full copy first, added the new form as its own stage, then checked the original was byte for byte the same. I approved the copy and went to sleep. At 3:23am it was submitted, and the form showed "Your response has been recorded." The first application is still there, untouched, next to the second.
"Your response has been recorded."
Next ↓What worked
The file grew on dates I can check. On 16 September the workspace held 22 applications and a shared catalogue of 87 programmes. On 28 September the checked count was 91 programmes, 25 applications and 10 forms submitted.
Counts · 16 and 28 Sep
Two outcomes are final. AltaLab accepted the application into its Fall 2026 cohort. MYStartup took us into Cohort 6 on 23 September, and the onboarding form went out the next day.
Accepted · 5 Sep
23 Sep · onboarding form out 24 Sep
The stage rule held on its first real test. A second form for the same programme went in at 3:23am on 28 September, and the first application was byte for byte the same afterwards.
First application · byte for byte the sameNext ↓What broke
The first page was the wrong shape. On 1 September an application was missing, because Claude had started a refill workflow instead of a page. On 8 September the programme work was still not in the module.
A sync script was clicking the wrong button. It treated an unfinished check as a finished one, so it always picked the first result.
One application showed as a draft for 23 days after it was sent. The confirmation emails were in an inbox the whole time. The list took its statuses from a fixed file in the code, not from the live records. A full check of every entry against the inboxes followed on 28 September, and the count of forms sent went from 9 to 10.
The signed file never met a receiver. The 14 September emails were stopped, and both drafts were missing facts the forms ask for.
On 21 September a coding agent invented four bot names and left Dossair out of a list it was already on.
Claude rewrote that section from the real list.
On 23 September Claude told the Dossair bot to submit a form. Dossair's rule is that it prepares, does not submit, and does not type a password.
Claude rewrote the step: Dossair fills, Claude checks, I press submit.
The same day the message pipe between us and the bots reported sends that were not in its record. Five were reported sent. Only one was actually in the record.
After that, every send was read back by its id before anyone called it sent.
The bot still did not wake up for the form. The next day Claude found I had already started it, kept the draft, and stopped when a safety check called the next click a real-world transaction.
On 22 September the applications screens did not update until a refresh. I called that five seconds of uncertainty I cannot afford.
The shared facts page was easy to miss, so on 23 September Cursor added a link at the top. That job had been told to write findings and change nothing. It received the title and not the plan, and shipped the fixes.
Next ↓What we learned
Next ↓Where it stands
No public page yet. dossair.rizqey.com was never set up. The applications sit in the team workspace. Checked counts on 28 September: 91 programmes, 25 applications, 10 forms submitted.
Automation includes: Claude Code on the laptop, Codex, NightShift, Council, Grok Build, Cursor, Sift, Blitz, and a Dossair Grok Bot that was given jobs and, in these records, did not finish one.
In this entry · 29 Aug to 28 Sep
In this entry · 13–14 Sep
In this entry · 14 Sep
In this entry · 14 Sep
In this entry · 23 Sep
In this entry · 14 Sep
In this entry · 14 Sep
In this entry · 23–24 Sep
Latest line in this entry · 28 Sep 2026, 3:23am
"Your response has been recorded."
Last updated 28 Sep 2026, 22:05 MYT