- A wall board for launch day, built to be read from across the room. There is a new page in the admin console for watching a rollout live: sign-ups, active clients, messages, searches, offers, picks and new projects, each leading with the last hour and each carrying its own minute-by-minute chart, plus three dials for the error rate, offers per search and picks per run, and a conversion dial showing how many of the people who ever wrote to Mira reached a search and how many went on to pick someone. It goes fullscreen for a TV. The TV in the room has no Google sign-in, so an operator can mint a one-time six-character code, type it on the TV, and that unlocks the board and nothing else for twelve hours, reading only totals and never a single person's details. Four rounds of follow-ups landed the same day: it was counting only the searches that come through the logged-out funnel, so the main flow, a signed-in client's search, read as zero; picks are now counted at the moment someone is actually picked instead of when the run started; the "search myself" click was only ever written to the logs and is now countable; a danger tile counts the clients whose frustration flag fired; and the whole thing gained a dark mode reachable from the page itself, since the TV has no sidebar to reach for.
- "Why did the agent do that" can finally be asked in production. The dashboard that answers it, every call an agent made on a turn with the exact text going in and coming out, lived only in the old internal admin app, and that app has never been deployed to production. So in the one place where the question really gets asked, there was no way to read a trace. It is now a Developers tab in the new console, on its own access grant that no product role hands out, and the whole selection lives in the address bar, so a specific trace is now a link you can send to someone.
- The nudge to a silent talent was never actually leaving the building. A talent who did not reply was supposed to get a reminder twenty minutes later. In production they got nothing. The message was written and saved, and the step that hands it to Fiverr was quietly skipped, so it existed with us and nowhere else. It looked perfectly healthy in our test environment, because the test talent screen reads our saved copy rather than a real inbox. All five of these closing messages rode the same broken call, the reminder, the last call, the timeout note, the declined note and the hired note, so all five were silent. All five are now addressed and sent properly, keyed so a retry cannot send the same message twice.
- Mira stops signing her name at the bottom of every message to a talent. Each message to a talent ended on a line like "Mira, on behalf of Dana". She already introduces herself in the opening line, so the sign-off said nothing twice, and four of the five carried a long dash, which the house writing style bans on anything a client or a talent reads. It is gone, helper and all, and the client's first name still reaches the talent where it belongs, in the prose and in the project's own name. This was invisible until the fix above made those closing messages reach a real inbox for the first time, so it ships alongside it rather than after it.
- A rejected offer must not be answered with a receipt. On production a talent quoted $720 against a $500 to $600 range, was asked to re-quote inside it, and answered "719". They were told "I've recorded your estimate as $719, I'll be back in touch when there's news", on an offer that had just been rejected, and the thread closed two turns later. The trap was structural rather than a slip: on a budget-only mismatch, the second round had nothing it was permitted to say and fell through to a script that reads as a yes. Now a talent who actually refuses settles the point for good, while one who moved a dollar gets the requirement stated once more; the reply states the outcome plainly instead of thanking them for it; the three phrasings that read as acceptance are named and banned; the final round announces itself rather than promising news that never comes; and there are more rounds now that the later ones have something honest to say.
- A finished project stops taking questions from talent. A talent asked something after the work was already handed over. The assistant correctly told them the question could not go to the client, and then passed it to the client anyway, because writing the refusal and forwarding the question were two separate actions and nothing made them agree. The forward is now blocked at the seam once the run is over, counted so we can see how often it happens, and the thread stays with the talent. A finalist asking a real question between the reveal and the client's pick is deliberately still passed on, because at that moment the client is actively choosing.
- Mira's Choice now leads with fit rather than fame. The choice among the finalists was decided first by a seller's standing on Fiverr, so a well-decorated seller won it before anyone read how well they fit the job. Reproduced on live code: a German-language project crowned an English-only top-rated seller over the perfect-fit person the scout had ranked first. The judge's verdict now leads, in bands, so a two-point wobble is treated as noise rather than a verdict, then the scout's own ranking, then standing only to break a tie.
- The scout carries what you asked for onto every single search. Three fixes from an audit of two real reruns. A stated preference, "preferring Italy or Spain", was treated as decoration and never reached the search at all, so a pool came back from the wrong countries with nothing to show it. Preferences now ride the first waves, are relaxed only when the pool proves genuinely thin, and an unmet one is named in the reasoning and costs the candidate something rather than being free. Country and language were being forwarded to only two of the search engines, and Fiverr's own screen can only filter on what it is handed, so the others came back unfiltered and we threw out 155 people ourselves on a single German rerun. And a candidate whose quality grade simply is not recorded now scores as average rather than bottom, because a missing grade is a gap in our records, not a fact about the person.
- A card should say what makes this person different, not what everyone here has in common. The "why they fit" line on the talent cards was filling up with the requirements every finalist already met, "US based", "speaks English and Spanish", because the pool was filtered on exactly those things. Across two reruns, all 24 cards led with the same shared tags and told the client nothing. Those lines now name what distinguishes this pick from the others in front of you. Two more corrections in the same pass: "full-time hire starting on the 20th" was being read as a two-day delivery deadline, which quietly starved the search, and a start date is now understood as timeline detail rather than a deadline; and "$100 an hour" had nowhere to live in the record, so it was being turned into an invented $8,000 total that then lit up the big-client alarm on a number nobody said. An hourly budget is now stored and searched as an hourly rate, end to end.
- The brief stops repeating your talent preferences back at you as if they were the work. A brief for a dog grooming business carried "the developer should be able to communicate in Urdu" in its overview and "support Urdu communication during the project" among the things the client needs, which then travelled onward into the reasoning and the messages talent see. Those are preferences about who to hire, not a description of the job. The cleaner only caught them when they were clearly labelled; it now catches them woven into ordinary sentences too, and the writer is told the trap by name. A genuine product requirement, "the store itself must support Urdu for shoppers", is scope and stays. Separately, the checklist row that asks for your business now insists on the address of your website rather than ticking on the business name alone, since the whole research chain runs on that page, and a client who says they have no site closes the row for good and is never asked again.
- The document you take away, and the one you never asked for. Three fixes to the exported job description: the page fitter was a page optimistic about where the printer breaks and could strand a heading alone at the foot of a page; there were two Mira marks in the document and now there is exactly one, in the page footer, measured off the app's own top bar rather than eyeballed; and no invented logo. A client without a logo on file was getting a grey square with their initials in it, a made-up mark standing in for their brand on their own document, with a generic globe behind it from a favicon service that cannot say it has nothing. Only a real capture of their own mark is used now, and the tile removes itself if there is none. The gate that reprints until the last page carries real content also stopped shipping a document that grew a whole extra sheet in the attempt. And the client is no longer handed a brief they did not ask for: the automatic download that fired on continue, on declining the sign-in prompt and again after signing in is gone, along with all of its machinery. The Export button inside the brief is untouched, and About You's own export button, which nobody was asking for either, is deleted down to the last route.
- Text now sits on the optical centre of every pill, chip, badge and label in the product. A single line of centred text was riding high everywhere at once, from two causes that stacked. The brand face ships an off-centre em box, so what the browser centres is not what the eye centres, and that was corrected once at the font rather than with some seventy hand nudges. And a fractional line height is not free: the browser rounds it down to the screen's pixel grid, which painted every chip half a pixel high. Of the 225 pill, chip, badge and label styles in the client app, the number sitting off the pixel grid went from 90 to zero. The audit also swept up an avatar nudge that would now over-correct, a style variable that was defined nowhere, and fourteen hand-rolled typewriter font stacks that collapse into one.
- The home page knows who you are before it paints, and a few other rough edges. On a refresh the greeting painted "welcome" first and swapped to "Hi, Dana" a second later, in the largest type on the screen, because every source for the name is a network call. The name of whoever just refreshed is now remembered from the previous load and read instantly, and a first-ever load holds on the loading screen for up to two seconds rather than showing a greeting it would have to correct. The second line of that greeting was also five sizes smaller than the first and read as fine print, so it is now sized off the greeting and finishes the sentence. Elsewhere: the "Job description" tab stopped sitting on top of the last talent card, since the label had grown and nobody re-measured the lane it hangs in; the hand-over dialog now leads with "Got it" as its main button instead of the Fiverr inbox link; the profile initial sits on the centre of its circle; and the Help centre points at the new article.
- And an easter egg. Ten clicks on the header wordmark inside a second and a bit, and Mira climbs up behind the chat box with her hands on the top edge, looks at you, blinks twice, and drops back down. Four hand-generated frames of the same character used in the brief artwork, clipped to the box rather than layered over it so she really is behind it, re-measured every frame so she stays gripping the edge while the box moves around her. Nothing is remembered between visits, keyboard and drag clicks do not count, and she stays away entirely for anyone who has asked their machine to reduce motion.
- Reporting, links and plumbing. The link in a lifecycle email pointed at a rendered brief behind a token that expired after seven days, so an email read late hit a dead page; it now opens the client's project, which is the place they can actually do something. The reporting warehouse was publishing a talent's Fiverr username under a field named "seller id", so an analyst's join against Fiverr's own tables was silently matching nothing; the real numeric id was being handed to us all along and thrown away, and is now kept on the broadest table we have. The talent dashboard was empty in production and pre-production because only the development environment had been told where the numbers live, so the hourly refresh ran, found nothing, and wrote nothing, every hour, forever. A failed warehouse promotion no longer turns a perfectly successful deploy red. And a storm of repeated requests when a workspace opens is closed, the curated VIP list is refreshed, and the test-talent roster gains a sixth account.
A very full Sunday, 41 commits across the crew, spent mostly on being able to watch the thing run and on messages that actually arrive. A launch-day wall board went up in the admin console, realtime to the minute, openable on a TV with a one-time code that unlocks nothing else, and was corrected four times the same day, including the discovery that it had been counting only the logged-out funnel's searches. The dashboard that explains why an agent did what it did finally reaches production. Underneath, a genuinely serious find: the reminder and the four closing messages to talent were being written and saved but never handed to Fiverr, silent in production for everyone while looking perfectly healthy in test. Alongside it, Mira stopped signing off, a rejected offer stopped being answered with a receipt, a finished project stopped accepting questions, and Mira's Choice stopped crowning the best-decorated seller before anyone read the fit. The scout now carries stated preferences and the country and language screen onto every engine, cards say what makes each person different, an hourly budget stays hourly, and a start date is no longer read as a deadline. The exported document lost its invented client logo and its second mark, the client stopped being handed a brief nobody asked for, and every pill, chip and badge in the app finally sits on its optical centre.