Tech with Emily

Portfolio

Selected work.

A living index of what I'm building. Each entry links to a longer write-up with the tradeoffs that shaped it.

  • Personal 2026

    Trousseau, an autonomous wedding coordinator

    Planning my own wedding turned into eight months of small decisions with long lead times, and no tool held the context between them. I built a coordinator that holds it.

    Outcome
    One skill, twenty-three expertise packs, ten scripts, four subagents, and thirteen memory files.

    Claude Agent Skills · Python · Scheduled agents · Progressive disclosure · HTML

    Read the full write-up →
  • Personal 2026

    Plasmid, an autonomous dev crew for a one-person business

    A portable crew of ten Claude agents that maintains, tests, and ships my sites on a schedule, whether my laptop is open or not. The crew brings its own craft; each repository it opens teaches it that repository's conventions. Built verification-first, so autonomy is earned rather than granted.

    Status
    First autonomous drafts open on this site's own repository. Phase 1 of 6 in flight.

    Python · Claude Agent SDK · Scheduled cloud jobs · Astro · TypeScript · Cloudflare Workers

    Read the full write-up →
  • Personal 2026

    Skill Library, a portable knowledge base my agents inherit

    I kept teaching the same procedures to a new assistant every few months and losing the work each time the tool changed. Skill Library is one repository that validates and packages twelve of those procedures, published at the path every major coding agent already scans.

    Outcome
    Nine drifting copies of the same validator collapsed to one. Twelve skills now pass a single portability gate on every commit, and a new project inherits the whole library with one command.

    Python · GitHub Actions · Agent Skills spec · Claude Code plugin API

    Read the full write-up →
  • Personal 2026

    The Keeper, a self-auditing skill library

    My skill library had grown to thirteen skills nobody had measured, eval files in a format the real test runner couldn't read, and no way to know when the next skill I built would push the whole listing past the limit where the model quietly stops finding it. I built the librarian that checks.

    Outcome
    13 skills converted, audited and CI-gated. 72 native eval cases recovered from a dead format. A 9-skill work order triaged to 3 by its own evidence gate before any were built.

    Claude Agent Skills · Python · GitHub Actions · Claude Code CLI

    Read the full write-up →