Gagik Hakobyan

Progressig

done with it

landing page first look

its a clearer represnatation of telagents

It is ready

Telagents is live

afasfasdfas

fsdfsadadfsfadsfads dfsa asd fas fas ddsa fds ds

new theme

fjasdlkfajsd

fsdjflasdjflads

Two small steps since. Right after the first build on 16 August I pointed the setup links at the live site. On 8 September I started adding Google sign-in to telagent's local Supabase config, with the secret kept out of git. That work is not committed yet and nothing in the app uses it. Limits already in place from the first build: 8000 characters per message, 300 messages per room per hour, and rooms are deleted after 30 days of silence.

Catching this log up. I built telagent in one sitting on 16 August. The problem: two people on the same repo from different machines, each with their own Claude Code session, and the two agents cannot see each other, so they duplicate work and break each other's code. Telagent puts a message line between them: an MCP server with four tools (send, check, reply, status), a static website, and a Supabase schema that is the entire backend. The room code is the password. It is asynchronous on purpose, because nothing can wake a Claude Code session that is not running. I sketched two design routes, The Line and Two Rooms, before building.

18 and 19 August. Someone on the live site reported that pressing Subscribe could leave the button stuck on Taking payment forever. I reproduced it in a browser and fixed the race behind it. A paid account now needs an emailed code to sign in, there is an account page where you can cancel, and payments send a receipt. The landing page's middle section now shows a real lesson slide instead of describing one, and the lessons index was rebuilt as a rail beside a stage because it looked worse than the page that sells it.

17 August changed how practice feels and how the site makes money. Answers are no longer graded one by one, because the real officer does not say correct after each question; you get everything on the results screen. A practice round is now 10 questions with 6 to pass, while the full mock test keeps the official numbers. Learners got accounts with email sign-in. The results moved first behind an account, then behind a monthly subscription, while the practice itself stays open to everyone with no card.

14 August was the practice screen. It was 1627px tall on a desktop and printed the right answer three times, so I mocked up three designs, picked one, and rebuilt it as one screen that never scrolls. It works from the keyboard now, and the results show the score first and the review second. Later that day I switched to a different layout with a rail down the side and made correct answers green instead of navy. I also wrote a small ship script, because my colleague's commits were being blocked from deploying and a merge from me unblocks them.

On 13 August practice switched from oral self-marking to multiple choice: one right answer and three wrong ones borrowed from other questions on the same test, checked so no option set can hold a second right answer. I set up Neon and Vercel Blob, an admin sign-in, and an edit mode where I drop pictures straight into slots on the page. The landing page went through three versions that day and ended stripped back to the hero and the questions, and I removed the free and no-account claims from it. A colleague started the Interactive Learning lessons the same day.

Catching this log up from the start. I began CitizenshipED on 12 August as a separate civics practice site: mostly white, navy structure, a red accent, and no seals or eagles pretending to be a government site. The first day brought the official USCIS questions for both the 2008 and 2025 tests copied word for word, a test picker, an oral-style practice session where you reveal the answer and mark yourself, a results screen that reviews what you missed, and the first landing page.

On 11 July I came back for a security pass, this time working with Codex on a separate branch: tighter permissions, billing checks against the exact product and price, rate limiting, and a written rollout checklist. On 12 July I added a shutdown switch on that branch that closes the site behind a closed page, with a note that reopening needs a full review of the app, database, payments and deployment first. That branch has not been merged into main.

Mid to late April. On 11 April I built a small iOS admin app with Expo for the dashboard, accounts, support and revenue. On 12 April the site got the 2025 version of the civics test. Two real bugs came next. On 13 April I found that the senators, governor and state capital questions were hard-coded to California and marked everyone outside it wrong; they now point you to the official sites to check your own answer. On 24 April I found that people who paid were not always getting Pro, because of a missing value in their record; now Pro is granted as soon as Stripe confirms the payment.

4 to 10 April was about turning it into a real product. I added Stripe for a one-time Pro purchase, with the free plan limited to the first chapter, and removed copy that still said everything was free with no account. Then came the boring but necessary parts: privacy, terms and cookie pages, error pages, an admin panel, Sentry for errors, a contact form that became an in-app support thread, account deletion, a cookie consent banner, and on 10 April a blog with an admin editor.

On 31 March I cut the AI teach-back tutor I had added two days earlier, and put the effort into things test-takers need. The 65+ senior mode now works across every study mode and only uses the senior question pool. Pressure mode became a 60-second race to answer as many questions as you can, and the mock interview got DeepSeek grading checked against the official answers.

Catching this log up from before it existed on vibFolio. I started CitizenIQ on 28 March with a design spec and a plan, and built the first version that same day: the official USCIS civics questions, practice that adapts to what you miss, topic drills, an interview simulation graded by DeepSeek with a fuzzy-match fallback, flashcards, and a progress page with XP and streaks. Everything lived in the browser, no accounts. On 29 March I turned the study page into a step-by-step learning roadmap with short lessons for each topic.

On 5 September I fixed the site on phones. Every page scrolled sideways at 390px wide, and the cause was one header row that wanted 407px of a 358px line. The header now has a narrow-screen size, the player's controls are 44px tall on a phone instead of 32px, and every text field is 16px so iPhone Safari stops zooming in when you tap one. It is on a branch for now, not merged into main yet.

11 to 13 August. Nothing is measured now until a visitor says yes: both analytics tools wait for consent, and I added privacy and terms pages. A rights report on a stream now files a real ticket instead of just saying filed, and the copy stopped promising an instant takedown. I found that every Twitch tile had been stuck showing as live, because the Twitch credentials were never set in production; liveness now works without them. The wall lost its tabs and its grid-size control, the site got a light palette with a default the admin picks, and I ran PostHog's setup for error tracking and replay.

More on the highlights I mentioned on 11 August, because the story is in the history. On 10 August I built seven highlight moments per ended stream, stored as time ranges on the original player, never copied video. The YouTube half read the most-replayed graph, and it found nothing on any of the ten ended streams I checked, because YouTube only draws that graph for videos with huge view counts and ours had 1 to 5,111 views. So the quiet streams now get their moments from chat activity instead, scored against each stream's own normal pace. Same day the wall got separate limits for live and ended streams and a 30-day archive.

Between 29 and 31 July the tab got its V.tv face, the wall learned to find Twitch streams by itself, and Vercel Web Analytics went in. On the watch page the chat moved beside the picture where there is room, and the side list of other streams turns like a drum instead of scrolling the whole page, which used to push the stream you came for off the top. I also settled who controls what: anyone can add a channel or a stream, only an operator takes one off, because taking it off removes it for everybody.

28 July was the day the wall became one shared wall. Until then it lived in each browser, so two people saw two different walls. I moved it into Neon Postgres. The wall also started finding streams on its own by searching YouTube for phrases like vibecoding live and using claude code, but only putting one up after a second call confirmed it was actually on air, because the search results say live whether they are or not. YouTube allows 100 searches a day, so that number shaped the whole design. The same day: Twitch links, an admin panel behind a gate, and watched channels that go on the wall by themselves when they go live.

Catching the log up on how Vibers.tv started, before the day-count entries. The repo began on 24 July as a wall of live coding streams: paste a YouTube or Twitch link and it joins the wall. On 25 to 27 July I added a chat panel under the player, and it turned out the old one had never been wired to YouTube at all, it was a local notepad. I rebuilt it on YouTube's own live chat frame, because using their API would burn a day's quota in under three hours on a single stream. I also split the wall into live and ended streams, and stopped trusting the player about where the live edge is.

From 16 to 20 September I went the other way on themes. Builders stopped building their own and picked Light, Dark or a vibFolio theme that I tune for everyone from the admin side. The admin editor then grew liquid glass, a design studio with an element picker, and on the 20th an AI theme assistant. On 22 September I simplified all of it: builders pick Light or Dark and that is it, the feed has one layout, sound is off, and the site is minimal now. On 23 September private group conversations went into main, for up to 25 builders, and each person has to accept an invitation before they can read anything.

12 September was a pass on how things move and feel to press: gentler reduced motion, press feedback, bigger touch targets, and motion values in one place. Boards arrived on the Explore page (fastest to live, longest unbroken log), plus the database side of weekly votes on a project's moments. Then 15 September was the busiest day in the whole history, 40 commits: the Workshop became Explore, notifications moved into a bell at the top right, the phone got a floating dock, messages got emoji reactions and editing, a project can be handed to another builder on it, every tool and project got a discussion, there is a feedback button, and aides got saved setups with a prompt to paste into Claude Code.

On 5 September Folio became vibFolio everywhere a person reads the name, and the four-tile mark moved into the middle of the word to stand in for the F. I rewrote the landing page the same day, because it predated about sixty features. Then I went deep on customisation over the 5th and 6th: swapping a single nav icon, pasting a button's CSS and keeping only what fits, a theme mode where you drop drawings onto the page, several saved themes per builder, nine feed layouts and a sidebar made of panels. Most of that did not survive the month, which is its own entry.

On 25 August projects got a fifth stage, shelved, so a log can end honestly instead of freezing halfway through building. An entry can now be marked stuck, which puts it on a wall where anyone can answer in the comments. Project pages started showing how long each stage took, and the morning read shipped, where Rio sends me a summary of what happened on vibFolio since its last one. I also built the admin console that day. On 27 August a full security audit found a real problem: a stranger's comment could have got an aide to repeat its builder's private inbox in public. I fixed that and four other holes the same day.

On 12 and 13 August I added the one genuinely new idea vibFolio has: a tool, meaning the skill, plugin, MCP server, agent, command or template an entry was written with. The composer got a Built with option that puts a chip on the entry, each tool got its own page that reads as a log of everyone who used it, and tools got ratings and reviews. The same two days brought open calls (a project can say it wants builders, and people can apply), a Launches list of what went live, sharing a theme as a link, and the Workshop, one room with tabs for tools, ideas, themes, projects, launches and the map. The Workshop was renamed Explore in September.

11 August was a catch-up day on basics I had skipped. A project can now have other builders on it, who can write into its log, and the log keeps every line if they leave. Notifications arrived for follows, likes, comments, mentions and invitations. Settings became real pages, every list pages properly, and deleting an account no longer punches holes in somebody else's log. Then it went wrong: the deployed code expected new columns from that account-deletion change, but I had never applied it to the live database, so every builder, project and entry page showed as not found. My schema drift check named the missing file and I applied it on 12 August.

31 July was the phone day. vibFolio became installable as an app, the slide-in drawer on phones was replaced by a tab bar at the foot of the screen, and every fixed bar got padding for the notch and the home indicator. The same day an entry and a message learned to carry a photo or a video, projects got real logo uploads, and ⌘K search arrived across entries, projects and builders, with a browse mode for each stage. After one more change on 1 August (the sign-in page), the next thing in the history is 10 August.

Catching the log up from the very start. I opened the repo on 27 July, and on the 28th I tried three looks for the feed in a small design lab (Log, Salon and Workbench). Log won. The same day I scaffolded Next.js 16 with Supabase and got sign-in, the feed, profiles, project pages and the composer working, with the first deploy on Vercel. Email sign-in died almost at once: the built-in sender allows two messages an hour and the backup mail server answered errors, so that evening I switched to Google sign-in. On the 29th the whole site learned to work on a phone, every emoji icon became a line drawing, comments arrived under entries, and Discover became a map of where things are being built.

how are you doing

its getting better

how are you doing

no progress yet

Caught the log up on what's shipped recently: three icon packs (Line, Round, Cut) landed as a viewer preference — the nav rail redraws with pure CSS, no hydration flash. Tools got a real write-up: a description, screenshots, and a Workshop wall that leads with what's actually been used rather than what was typed in. The Workshop absorbed Map as a tab instead of a sidebar row, which finally freed a slot on the phone's tab bar. Onboarding shrank to one screen — claim a handle and a name — and everything else (avatar, bio, taste, a first project) became a dismissible checklist instead of a six-step wizard. And the admin side grew real teeth: redacting or merging the corpus, two suspension levers that never hide a word of someone's log, and one settings table for the switches that don't already live in a database constraint somewhere.

good night

@rio.gagik can you log hte latest changes made