GitHub Copilot Pro vs Pro+ vs Claude Code
A practical decision guide for model access, premium-request headroom, Claude Opus availability, and which coding subscription fits different workflows.
Engineering Notes
Build notes & architecture breakdowns.
Deep dives into how Ora and htmlctl are designed, shipped, and iterated — from system prompts to deployment pipelines.
MacTalk vs Apple Dictation
Apple Dictation is built in. MacTalk is for local Whisper and Parakeet control, app-audio capture, auto-paste, and faster dictation into AI workflows.
MacTalk Was My ASR Playground — and It Led to Ora
MacTalk started as my sandbox for trying Whisper, Parakeet, and different ASR interaction models. That experimentation helped shape Ora — and now MacTalk is back in my daily workflow for talking to coding agents.
Run Coding Agents From Your Phone With Telegram
Why Telegram works as a mobile control surface for coding agents, and how TelePi and TeleCodex turn voice notes, screenshots, and short prompts into real coding workflows.
TelePi vs TeleCodex
A practical comparison of the two Telegram coding-agent bridges: context model, mobile UX, ASR, attachments, launch profiles, and when each one fits better.
TeleCodex: Telegram Bridge for OpenAI Codex CLI
TeleCodex keeps OpenAI Codex CLI reachable from Telegram with launch profiles, live plan visibility, file exchange, voice input, and instant handback to the terminal.
AI, the UI Anywhere
The broader essay behind TelePi and TeleCodex: why steering coding agents from your phone makes sense once most of the job becomes supervision instead of typing.
TelePi: A Telegram Remote Control for the Pi Coding Agent
After canceling Claude Code, I doubled down on the Pi coding harness and built a Telegram bridge — bi-directional session hand-off, cross-workspace switching, and model selection, all from a chat interface I was already using.
Why I Switched from Claude Code and Google AI Pro to GitHub Copilot Pro+
Claude Code's shared usage limits were a bad fit for my architecture-first workflow, Google AI Pro felt siloed, and GitHub Copilot Pro+ gave me a broader multi-model bench for planner, coder, and reviewer roles.
How a Newsletter Became htmlctl's Extension System
Restoring the newsletter forced more than a backend. It led to a reusable extension model with same-origin routing, compatibility validation, safer runtime boundaries, and static releases left intact.
We Don't Just Design for Humans Anymore
I tried to share an htmlctl update on social media. The preview was blank. That moment led to shipping automatic OG image generation, favicon support, robots.txt, and a sitemap generator — making discoverability effortless by default.
The Bar Has Been Raised: Consumer Expectations in the Age of AI Software
Ora is a good reminder of how fast the baseline is moving. If a solo developer can build a voice-first tool that understands speech, uses local or cloud models, remembers context, and acts across Mac apps, users will soon expect that kind of flexibility everywhere.
Dynamic Tool Discovery: How We Reduced Prompt Tokens by 60%
Ora registers 41 tools. Sending all their schemas to the LLM on every turn was expensive and noisy. Dynamic tool discovery introduced lazy loading via a client-side discovery index — shrinking the initial prompt tool block from ~1,200 tokens to ~450.
More posts coming soon · follow on GitHub