Research

        epic
   ┌──────┼──────┐
   ○      ○      ○    wave 1
   │      │      │
   ○──────○      ○    wave 2
          │
      [ FAIL ] ──► PASS

Agent orchestration · Claude Code plugin

completely

An orchestrator for autonomous coding agents, built on one rule: done has to be proven, not claimed. Work is split into tasks in the Beads tracker — epic, tasks, waves of dependencies — and each task goes to a fresh worker with a clean context. A task closes only after deterministic checks and an independent evaluator that defaults to FAIL.

Read the field report → github.com/23ag1/completely ↗

┌─ layout ─┬─ states ─┬─ perf ─┐
│    ✓     │    ✓     │   ✓    │
└──────────┴──────────┴────────┘

  ├────────── 16.7 ms ──────────┤
   ▁▃▁▁▅▃▁▁▃▁▁▁▃▅▃▁▁▁▃▁▁▅▃▁▁▃▁
        60 fps · cpu 20×

Frontend standard · Claude Code plugin

frontend-quality

A frontend standard: taste, mechanical layout checks and performance analysis. Three skills — process and a quality rubric, visual decisions, performance — plus a ui-verifier acceptance agent. Layout is checked in a real browser against a catalogue of failure modes, and performance is traced down to the function that ate the frame budget: 60 fps with the CPU throttled 20×.

Read the case study → github.com/23ag1/frontend-quality ↗

 url ──► html · css · js
            │
   ┌────────┼────────┐
 stack   effects   tokens
   └────────┼────────┘
            ▼
       build plan

Claude Code skill

site-teardown

Reverse-engineers any website into a complete build blueprint: the tech stack, every visual effect and how it is implemented, the full design system, the assets needed and a section-by-section build plan. It fetches the page source itself, so all it needs is a URL.

Read the case study → github.com/23ag1/site-teardown-skill ↗

 edit ──► check ──► review ──► commit
   ▲                   │
   └─────── fail ──────┘

 Claude Code  ◄──── config ────►  Codex

Skills · Hooks · Subagents

Agent tooling

Infrastructure for agent-driven development. Own skills, hooks and subagents: automatic checks after every edit and before every commit, protection against conflicts between parallel branches, scenario and regression checks. Plus a sync that compiles the same instructions for Codex CLI.

Read the case study → Personal tools — not published

 12B  ▁▃▃▁
 26B  ▁▃▅▅▃▁

 ctx  ├────────────────┤ 64K
 kv   q8_0   ·   w 4–5 bpw

Runtime · Quantisation · Benchmarks

Local LLM stack

Running and measuring open-weights models on a MacBook. An isolated MLX runtime for 12–26B models — weights stay out of the system cache, served over HTTP on localhost. Experiments with weight and KV-cache quantisation (4–5 bpw, q8_0), measurements at up to 64K context, quality scored on BFCL (tool calling) and long-context benchmarks. The working profile is picked by the numbers.

R&D — not published

 voice
   │
   ├─ phrases ─── sure? ──► do
   ├─ app menu ── sure? ──► do
   ├─ meaning ─── sure? ──► do
   └─ log of misses

Voice control · macOS

macvoice

Voice control for macOS, local, without the cloud. A tiered router: on-device speech recognition, then a phrase dictionary, then the active app's menu through the Accessibility API, then meaning matched by an embedding model, then a log of misses. Each tier runs only when the one before it isn't sure.

Personal project — not published

 ┌───────┐
 │   ·   │   image
 │  ∘○∘  │   + camera move
 │   ·   │   + voice-over
 │       │   + analogue grade
 └───────┘   1080 × 1920

Video · Speech synthesis

AI video pipeline

Short vertical videos from AI images, fully local. Builds 1080×1920 clips from still images with camera movement added in the edit, a voice-over and an analogue grade. Speech synthesis quality is checked automatically. No paid APIs.

R&D — not published

        ▁▃▅▃▁
     ▁▃▅▅▅▅▅▃▁         ▁▃▁
  ▁▃▅▅▅▅▅▅▅▅▅▅▃▁   ▁▃▅▅▅▅▅▃▁
 ────────────────────────────
    ∘    ·     ∘    ·    ∘

Claude skill · WebGL2

ink-wash-painting

Chinese ink-wash painting (水墨) of any subject, as code. A public Claude skill: for each subject Claude writes a procedural renderer on WebGL2 shaders — how paper, water and ink behave, the grammar of the brush stroke. It doesn't reuse one scene; it builds a new one for every request.

Read the research → github.com/23ag1/ink-wash-painting-skill ↗