Problem

Coding agents compile. They still ship the same generated look: system fonts, violet accents, shadow on every card. The harness grades its own homework.

Approach

An independent critic beside Cursor. Playwright captures the rendered page. Fourteen detectors name genericness and drift. Direction becomes a reviewable patch. Proof runs in a disposable checkout.

Shipped

Next.js monorepo, MCP tools, Studio skill graph, and a captioned product demo. Open source. Local dogfood. No public production host claimed here.

Problem

Most writing about agents starts at the product surface. The actual job is smaller. The model proposes. The harness allows, writes, and proves. People skip that boundary.

Approach

One task people can hold: create hello.txt containing exactly ok. FakeModel first, no API key. Optional live guest on the same Allow, write, and byte-check path. The model never writes the disk and never decides that the run succeeded.

Shipped

Runnable proof script, tests for success and failure, JSONL transcripts, interactive Allow, argv-only run_command, teaching memory, and a journal TUI. A tutorial, not a product.

Problem

Agent chats go sideways when every session is a new personality, the expensive model does the busywork, MCP schemas eat the window, and green means a screenshot.

Approach

A harness, not a second IDE. Playbooks compile into commands you can rerun. Always-on rules stay a few short files. Skills load when you type them. Install refuses to rewrite your home folder without --i-consent.

Shipped

Skills, three specialist agents, apply/verify/hygiene CLIs, and the same kit in the editor and the terminal. Marketplace listing is a manual next step. A clone works today.

Problem

Notices are dense. People get lost, then hand a chatbot a password. A civic copilot that files, pays, or scrapes a portal is a hazard, not a feature.

Approach

Help a person understand a document. Extract what matters. Build a checklist. Look up a reviewed official channel. Stop. No Aadhaar, PAN, OTP, or banking details. No model-supplied URLs.

Shipped

Local Next.js app, eight WebMCP tools, on-device OCR, EN/HI/MR and more explain languages, and a reviewed portal directory. Intentionally no GitHub Actions CI. No public host claimed here.