I'm building @mato_hq, AI agents that interview humans, starting with podcasts. $0K to $18.5K MRR in 3 months; nobody has churned; about 30 shows running. one of them sits at #10 in News Commentary on Apple Podcasts. before this, I spent years turning other companies' conversion rates into revenue, $150 million in added revenue to be exact, wrote a book about it, and got a wireframe method design-registered in Norway twice. now I do it for my own company. bootstrapped the journey so far.

a rate-limited health check that you dial back is just a check you've decided to run less often. the fix is pacing, not silence. Mato's Apple catalog lookups now use shared caching, request limits, and provider cooldowns. the distribution checks stay intact. which external dependency check have you quietly disabled because killing the noise was easier than fixing the rate limit?

when an integration's setup step is 'paste your API key here,' the secret is now inside the assistant session. Mato MCP connects over hosted OAuth instead: point your assistant at heymato.com/api/mcp, sign in, approve access. no key to copy into a chat, no server to run locally. the credential surface is a design decision, not a default. how many of your assistant integrations still use the paste-a-key path?

your LLM reviewer returns a paragraph. someone reads it, re-decides what it means, and either acts or doesn't. you haven't automated the review. you've added a step. a typed finding names what is wrong and where. that's a different thing: it can be routed. some findings a repair pass can act on, some mean the plan shouldn't be drafted at all. what does your critique step actually return?

if your retry re-reads config instead of replaying a frozen snapshot, it's not a retry. it's a new run that inherited a job id. we ran into this with voice direction in Mato: the selection had to stay identical across every audio part and every retry, including when job data got re-ordered between workflow steps. where in your multi-step pipeline does the retry path actually lock the config it started with?

haha claude telling me times are rough since I'm running so many lanes

one script, three renders: Gemini, ElevenLabs, and Mato Voice. same fifteen-line two-host exchange about a thermostat, identical speaker turns, all loudness-matched to the same target, no timeline edits between them. a voice demo without disclosed normalization is trivially riggable. push your own clip a couple dB hotter and most listeners will call it warmer and more present without knowing why they preferred it. next time you're on a voice vendor's comparison page: what did they tell you they normalized? if the answer is nothing, you're judging the mix, not the model.

you can hear an AI-written podcast inside twenty seconds. one host delivers a whole argument in a single turn, the other agrees, facts keep arriving wrapped in "according to," and every new topic opens with a line announcing a new topic. we had spent a long time on voice quality, pacing and audio assembly. none of that was the problem. the system went from a pile of source material to a script in one leap, and everything synthetic came out of that leap. better speech on top of a monologue gets you a better sounding monologue.

accent color is a budget. spend it on every button and call-to-action row and you've told the user nothing about which moment actually matters. on the new Mato site it's withheld until something human happens: 'interview,' 'finished episode,' 'out loud.' warm white, graphite, Neue Haas Grotesk, and space do the rest. the gradient doesn't decorate the software. it marks the moment a thought crosses into the world. if your accent appears 40 times on a page, it appears zero times with meaning.

deleting or unpublishing an episode in Mato now removes it from the generated feed before its local lifecycle changes. feed first, internal record second. the reason is boring and load-bearing. a listener's podcast app only ever reads the feed. so if you've got a public feed or a CDN sitting in front of your data, the order of those two writes is the user-visible behavior. which one does your unpublish path touch first?

Opus 5.5 performs at the level of Fable 5.1. It's ~30% faster and ~40% cheaper than Opus 5 per task. In Claude Code: - 5-hour session limits increase 20% today - Opus 5.5 is priced lower, so it goes 25% further within limits - Pro, Max, and Team users get a reset to use anytime

we went into transcript measurement expecting to indict other shows. the finding came back against our own back catalogue. that's what happened when we turned editorial taste into properties you can actually check in a transcript. each line either holds or it doesn't. style notes give you an argument. a structural property gives you a test. which of your quality notes are actually checkable, and which ones still rely on whoever wins the review call?

our generation screen had a job sitting at draft-ready, waiting on a person to review it. the screen kept showing progress. it had no state for done-and-pending-you. we fixed it so a review-waiting job treats itself as a completed handoff and routes you to the episode instead of leaving you watching a bar. the interesting design question: how many terminal states does your long-running UI model, and is 'done but requires human action' one of them?

is it better to have fable review code after driving your main session with opus? or is it best to use fable as the driver of opus? the latter eats usage like crazy. any experience?

I don't understand what openai did with Astra. I've been running it on medium for about 6 hours now, and I have 45% usage left on the 20x plan. when astra was released, I could run it on Ultra for 12+ hours, and I would still have like 75% left?

voice cloning in Mato requires the speaker's permission, and both cloning and original voice design run as guided sessions with our team rather than a self-serve upload. that framing is a product decision, not a missing feature. the speaker whose voice gets used and the producer clicking a button are two different people. consent has to live somewhere in that gap, and a form field is not that place.

I wonder what zuk is upto now

getting my new iPhone pro max tomorrow. decided to skip the duo until they ship one with all the fixes