Apps

Frameword is a macOS video editor that Claude drives. You describe the edit in a sentence: “cut the filler words and long silences, add karaoke captions, export.” Claude turns that into tool calls and applies them to the open project.

There is an MCP server inside the app. Claude Code, Claude Desktop, Codex or any other MCP client connects to it and edits the project through 78 validated tools. Every call is one undo step, so nothing Claude does is irreversible.

The transcript is the foundation. When a video is imported it is transcribed on device with Whisper Large v3, with word timings. Effects anchor to words, not to timestamps: “the moment 2026 is said” means the exact frame that word starts. Cut the clip and the effect stays with the word.

The images are not the app’s interface but its output: each one was made with Frameword from a single sentence given to Claude.

No account, no Frameword server, no telemetry. Video, audio and renders stay on the Mac; Claude only sees the transcript text and the project structure. Status: early access, downloaded from the site as a .pkg; not notarized yet.

Tacet is an AI assistant for your documents that runs entirely inside your iPhone. Hand it a spreadsheet and you get one back with the formulas still live.

I built it because of one sentence: I want a document assistant, but I don’t want to upload my company files to a cloud. Every option on the market shipped the file to a server. Tacet has no networking code: not a disabled tracker, not an analytics SDK behind a switch, but no code that can open a connection at all.

The hard part wasn’t the model, it was the approval flow. If an on-device model is going to touch your files, it has to show you what it will do first. Every action surfaces as a proposal, you approve it, then it runs. This is where I learned that adoption of AI features is driven by reversibility, not accuracy.

Rituell reminds you to stay in touch without turning it into a chore. You sort people into circles (family weekly, distant cousins monthly) and each day the app shows only that day’s short list.

The design problem was obvious from the start: this kind of app turns into a guilt machine very easily. Show someone a backlog of 40 overdue people and they delete it. The fix was to never show accumulated debt: a missed day simply disappears, today’s list is rebuilt each morning, and it’s never longer than a few people.

Four releases so far; each one made the list shorter.

Numeris shows you four numbers. Find the hidden mathematical rule and enter the fifth before the timer runs out.

The real work isn’t the puzzles, it’s the engine that generates them. Hand-write puzzles and you write a hundred and stop. Write a procedural rule engine and it generates forever, but then you inherit a difficulty problem: a randomly generated rule can be trivial or impossible.

The fix was to score rules by difficulty and draw from a band tied to player performance. Play well and the band shifts up. This is the most transferable thing game design taught me: difficulty isn’t a fixed number, it’s a function of the player.

Miqat is a prayer times and qibla app. What makes it different is technical: there is no networking code in it. Not a disabled tracker, not an SDK behind a switch; no code that can open a connection at all. Which means there is nothing to send anywhere, and nothing to sell.

Everything is computed on the device from the position of the sun, using the same astronomy a printed timetable uses. That’s why it works on a plane, in the desert, and with roaming switched off.

I validated the engine against published timetables: Istanbul, 19 August 2026, Diyanet method: 04:37, 06:16, 13:14, 16:57, 20:01, 21:31.

No package dependencies; the solar calculation is written directly in Shared/Core/Engine.

Poligon is an iPad and Mac asset studio for people making retro-styled games. The thesis is one sentence: the 2D canvas is the 3D texture. Paint a texel on the canvas and it appears on the mesh; paint on the mesh and it lands in the canvas. There is exactly one compositing path, so the two views cannot disagree.

Getting that right was the hard part. UV islands have to pack onto whole texel boundaries, otherwise an island edge lands halfway across a pixel and the model looks blurry. Without automatic seam bleeding you get a black line at every island edge.

PSX mode is entirely optional and my favourite piece: it deliberately puts back the three artefacts that made 1994 hardware look the way it did: vertex snapping, affine texture warping, and 15-bit ordered dither.

Two custom Metal surfaces, a 120 Hz canvas, and a file-based document format that diffs cleanly in git. Layers are stored as raw palette indices.

No account, no subscription, no cloud, no AI generation. One-time purchase, and it works on a plane.

GuessFour is a logic and deduction game. You try to guess a hidden number, and the feedback after each guess narrows the probability space.

The only design question was how much information to give the player. Too much and the puzzle solves itself; too little and it stops being a logic game and turns into gambling. Finding the right setting meant running a few hundred simulations on paper, measuring the average number of moves to a solution and tuning the feedback against it.

Unlike Numeris there’s no procedural generation here; the puzzle space is fixed and closed. That makes it harder as a design problem: you can’t hide behind a generator.

Sequences is the same game as Numeris on Android, but not the same code. This is the original, written in Expo and React Native; Numeris on iOS is the native rewrite I did afterwards.

Running two platforms from two codebases wasn’t a deliberate choice, it was the result of a lesson. The React Native version shipped fast, but timer precision and haptic feedback never felt right on iOS; in a speed game, 30 milliseconds of latency is something you can feel. The native rewrite fixed it.

The product lesson is clean: a cross-platform framework is the right call as long as the most critical feel in the product isn’t platform-specific. Here it was.

There are seven sequence types and questions are generated with random parameters every game, so there’s no memorizable question pool. A repeat-prevention memory keeps the same question from coming back-to-back. At game over the last question’s solution is shown step by step; that’s the part aimed at learning while you play.

Datum is a 3D scanning app for LiDAR-equipped iPhones and iPads. Two scenarios: scan rooms indoors and get a floor plan you can export to CAD, or take a photorealistic capture outdoors.

I applied the same constraint as my other apps: processing stays on the device. The standard approach in this category is to ship the scan to a cloud and process it there. The interior plan of a building is sensitive data; for a contractor or an architect, keeping it off the network is a reason to buy.

The name was a deliberate choice. “Datum” is a foundational term in geodesy and CAD: the reference surface measurements are tied to. Architects, surveyors and contractors, the target audience, already know the word.

Status: in preparation on App Store Connect, build uploaded and validated. It goes to review once the screenshot set and listing copy are finished.