An operating system for working with AI agents

Why

For most of my career, “designer” meant I stopped at the handoff. In 2025 that stopped being true. With agents doing the typing, the bottleneck moved to the thing designers are good at: defining the problem precisely, structuring the work, and judging the result. So I treated my own workflow as a product.

The ICM method

ICM is the workspace methodology I use for every agent project. It treats a repo like a physical production shop:

• AGENTS.md is the front-door sign — what shop did you just walk into.
• CONTEXT.md is the reception desk — it routes a request to the right workbench.
• Each stage is a workbench with a small contract: allowed inputs, process, outputs, and checks.
• references/ is the rule binder; processes/ are SOPs; scripts are machines the agent invokes rather than improvises.
• output/ is the finished tray — humans review it before the next stage trusts it. Every job carries a durable run.json.

Core rules: load only the context for the current bench; write durable files, not chat conclusions; verify before handoff; when the same correction repeats, edit the source (contract, reference, or script) instead of the output; keep contracts small. It’s what lets a long pipeline run across many sessions and models without the context degrading.

What it has shipped

• ConvertSave — Tauri 2 / Rust / React desktop app; 398 commits; cross-platform CI; parallel Claude Code worktrees for features.
• ConvertSave for iPhone — 100% native Swift/SwiftUI media converter (AVFoundation, ImageIO, PDFKit, CoreImage), zero third-party dependencies, sharing the Nocturne design tokens with desktop.
• Screen Studio Redactor — native macOS SwiftUI/AppKit app plus CLI that finds and masks PII in tutorial recordings (Vision, Core Image/Metal, SQLite; non-destructive projects). Paired with a 265-job Antigravity computer-vision swarm that inspected video keyframes before anything went to YouTube.
• Linked Designs — a Canva App bringing Photoshop-style Smart Objects to Canva (Canva Apps SDK + Connect API, OAuth 2.0 PKCE, Node/TypeScript).
• YouTube operations — watch-folder auto-uploader with metadata rules, SEO dashboards, a competitor-velocity warehouse (Top10k), a thumbnail pipeline with tightly constrained copy rules, and a Remotion B-roll generator.
• Studio hardware tooling — Swift/AVFoundation utilities for zero-latency iPhone mirroring in device frames, camera pixel-format fixes, and headless control of Screen Studio through Chrome DevTools Protocol.

Voice & ambient workflows

A lot of my agent work happens by voice. A custom skill (top10k-voice) lets me review competitor videos and curate production queues hands-free. A Discord voice bridge runs an agent against a workspace and reads results back at 1.25× so I can supervise from the studio floor. Dictation (Wispr Flow, superwhisper) is the default input for prompts and PR descriptions.

Tooling I’ve built for agents

• Figma MCP server
• Top10k MCP + CLI
• Mangools / VidIQ MCP
• screenstudio-cli (CDP)
• youtube-auto-uploader skill
• convertsave-paper-design skill
• ios-simulator / xcodebuild skills
• Thumbnail Agent v6.2 rules
• ICM templates & conventions

Everything above runs on my own machines and pipelines; where a tool is public (ConvertSave, Linked Designs) it’s linked from the work page.