By name: kai-tech-prefs description: Kai's technical preferences - Go CLI defaults (urfave/cli), Charm TUI stack, dead-repo check, doc conventions, default license. Triggers - cli, tui, library, dependency, recommend tool.
Technical preferences
Stack
Python, AWS, Kubernetes, Terraform. Distributed systems, platform engineering, o11y-heavy. LLM APIs and production-grade AI systems.
Go CLI
Kai is/was a maintainer of urfave/cli. Default to urfave/cli over cobra/kong. When in doubt about Go CLI tooling, check what urfave/cli does first.
Charm (preferred TUI/CLI styling family)
Charm (github.com/charmbracelet) is Kai's preferred TUI/CLI-styling family. Reach for:
- bubbletea (TUI framework)
- gum (shell-script prompts/styling)
- glow (markdown render)
- lipgloss (styling)
- huh (forms)
- vhs (terminal recording)
- soft-serve (self-hosted git)
- wish (SSH apps)
- mods (LLM CLI)
- freeze, melt
When a coily/repo-recall/eco tool needs interactive prompts, fancy output, or a TUI, bias toward this stack over hand-rolled ANSI or rivals like tview/promptui. Cotton-candy aesthetic, MIT-licensed, very actively maintained, plays nicely with urfave/cli for the imperative-CLI-with-occasional-TUI shape.
- Don't suggest dead or dormant repos - 12-month-commit bright line before recommending any OSS dependency.
- Aliases and pagers - no renamed-command aliases, pagers configured off.
Docs
No "Repo layout" / "Project structure" sections in README. Filesystem is self-documenting. If a dir layout needs explanation (non-obvious separation, unusual build output), brief prose under a purpose-focused heading, not an ASCII tree.
Licensing
Three tiers by repo intent - MIT for shareable, AGPL-3.0 for deployment-of-one, proprietary All-Rights-Reserved for private personal repos. Default to MIT when intent is unclear. Full policy: coding-licenses.
More preferences
- JSON-twin discoverability for dashboards - three discovery mechanisms so cold-start agents find a JSON variant.
- Configs go in SSM, not in skills or code - stash rotatable config-shaped values in AWS SSM, never hardcode.
- Batch job rules - no parallelism for rate-limited jobs, plus coily wrapper rules.