Principles
- Don't add a tool/complexity unless I actually have the problem it solves
- Prefer GUIs over TUIs
- Token savings must be real
Models
Using
- Claude: $20 USD/month, $200 USD/year
- OpenCode free models
- OpenCode Go: $10 USD/month
Dropped
Felt slow compared to Claude and OpenCode free models
- freellmapi: Stack LLM free tiers
- turbo-fieldfare: Gemma 4 26B-A4B inference in ~2 GB of RAM on any M-series MacBook · GitHub
Text to speech (TTS)
Problem: I want to listen to Claude Code output. 2026 macOS built-in Siri voices are still too robotic.
Using
- ElevenLabs API key (free tier) for listening to Claude output
- Mac OS Shortcuts: global
- Kokoro for converting longer text files into audio
Dropped
- edge-tts
- API calls to Microsoft Edge TTS, frequently fails
- 2026: GitHub - KittenML/KittenTTS: State-of-the-art TTS model under 25MB 😻 · GitHub
- pretty good but more robotic than built-in 2026 macOS Siri voices
Code quality
Using
- Linters
- oxlint and oxfmt
- anti-slop: Oxlint rules for rejecting low-evidence TS patterns
- fallow: codebase intelligence for TypeScript and JavaScript
Workflow / methodology skills
Using
- Matt Pocock skills
- I like
/domain-modelingand the GLOSSARY.md to discover
- I like
Dropped
- superpowers
- I believe it uses way more tokens?
- ponytail
- write less & more elegant code
- really good for fighting back against scope creep
- uses a lot of tokens and I rarely used it
Code navigation, discovery
Dropped
- codegraph
- fewer grep calls
- Opus 5.5 and Sonnet 5.5 got cheaper and I don't want to add complexity if I don't need it
graphify: does too much (I don't need it to read PDFs)- GitNexus: modifies the
CLAUDE.md(lists how many files are indexed, needs to tell Claude to update the index)
Token / context savers
Dropped
All: unproven token savings. They also conflicted with codegraph (would caveman grep / rtk grep instead of using it), though that's moot now that I've dropped codegraph.
- caveman
- fewer output tokens, but might not save tokens overall
- rtk
- filters and compresses command outputs before they reach your LLM context
- might not actually save tokens
- context-mode
Code review
Using
- IDE
- playwright: create side by side annotated diffs
Dropped
- hunk: TUI, I prefer reviewing in an IDE
- 5 min video: https://www.youtube.com/watch?v=FFfz81XM57k
Terminals
Using
- cmux
- built-in browser
Considered
- herdr: tmux for agents
- LLM can launch and reference other instances and panes
- when you close the terminal, you don't lose your work
- essential if you SSH to your Mac
- but doesn't make sense in a terminal like cmux or Warp, which already have left side panels
Dropped
- Warp
- didn't really need much of its nice UX (code diff/review, code editor, their agent stuff you have to pay for)
Agent orchestrators (to evaluate)
What problems do these solve, and do I have those problems?