The first two days of this week went to one idea: I don't want to edit my videos, I want to review them. video-tool got rebuilt around that. The rest of the week was quieter and a bit more humbling. Mind Journeys backed out of a paywall decision, language-learn got a page that shows exactly how far from launch it is, and MyFoodJournal had bugs hiding in plain sight.
video-tool Goes Autopilot-First
Last week Claude was a helper inside my editor. This week it became the editor. The planning passes moved to Claude, with OpenAI as a fallback, and an agent got the project tools, with every change undoable. Autopilot is one agent turn with a standing brief: transcribe locally, make the cut, set the style, direct the effects, add captions, make shorts, export.
The home screen is now basically a drop zone. You land on a review screen with a player and a chat, and the full timeline sits behind a button called Edit manually. That button placement is the whole product decision.
I also added an MCP server, so Claude Code or Claude Desktop can drive a project from outside the app. Every write becomes a revision and shows up in the same chat, so undo works no matter who made the change.
Then came 18 new effects in three packs. The fun ones need the speaker cut out, so Apple Vision's person segmentation now runs on every frame of every camera recording, at about 36 ms a frame on my M4 Mac mini. With that matte I get text behind me, a window I get pulled out of, and a paper sheet I burst through. A day later I retired sketch boil, which is why Settings now has a list of effects Claude may use.
The best changes came from watching real output. My exports measured −27.6 LUFS against the −14 that YouTube, Instagram and TikTok normalise to. Now every export gets voice cleanup and two-pass loudness with local ffmpeg, and the main cut landed at −14.3. Cards stay off my face, captions move out of the way of titles, and Claude's instructions now include timing rules: a graphic starts on the clip where its idea is said, and no more than two of the same kind run in a row. I haven't tried those rules on a real edit yet.
Mind Journeys: I Took the Lock Back Off
Last week I was marking ten journeys as premium. The data talked me out of it. Free users finished the original free journeys and abandoned most of the others in the first 10%. My read is that a long commitment was the problem, and a lock wouldn't fix that.
So every bundled journey now plays free in a version of at most 12 minutes: same steps, same tones, squeezed to fit. Full length is what Pro adds.
Onboarding had its own number. About 30% of people left before picking a goal, which was screen seven of eleven. Now the goal is screen three, and the whole flow is seven screens, or eight if you pick Meditate. There's also an optional trial reminder, two days before the trial ends.
1.9 is in App Store Connect with two Pro-only 15-minute journeys, a tone slider and a rating card on every finish.
On YouTube, one 8-hour dark-screen Delta Sleep video earned about half the channel's watch minutes in a week: 1,536 of 3,144. So a script now makes more in that format. It renders each video, measures the tone in each ear, confirms the frame is near black and uploads it as private. Three are up. It can't publish, on purpose. That part stays with me.
language-learn: Measuring the Real Blocker
Content is what's holding this app back, so most of the week went into the admin tools for making it. A new Coverage page lays out the course: 100 units, two started, one signed off, 98 still outlines. Not a fun page to open. It's the right one, though.
The Prepare & approve flow takes each unit from partly done to signed off, with AI drafting and me reviewing. Nothing the AI writes is saved until I accept it, and approvals store a hash, so a step reopens if its content changes later.
I also imported the 2,500 most common Spanish words from open frequency data. AI suggests a CEFR level, topic and meaning for each one, and tagged the first 100 in about 24 seconds. It put usted at A2 and kept forms I didn't want, so the prompt got tuned before going further.
The learner side moved too: Learn and Practice tabs on iOS, recorded audio for words, flashcards that play it, and scenario checklists with a little confetti. Claude Haiku checks tasks off after each turn for about $0.0001.
The Smaller Fixes That Mattered
MyFoodJournal had a bad one. Saving food also fetched the user's profile, so if
that fetch failed, a perfectly good save showed an error. Those are now
separate, and new anonymous users get their profile created right after sign-in
instead of hitting a 404 on first launch. A backend safety net is on a branch.
Manual, voice and photo logging also never fired log_created, so my analytics
had been missing three whole ways of logging. PDF nutrition reports shipped too,
and the marketing site finally has the sitemap and robots file it was returning
404s for.
Calculador Neto got a 2026 engine comparing autónomo against sociedad limitada, plus three Spanish guides built on it. The honest answer in those guides is that there's no official threshold. An SL mostly pays off when the profit stays in the company.
Closing
Autopilot only counts if I'd publish what it makes, so next week I'll feed it real recordings and see how often I reach for Edit manually.