Lead with the stack.
State your framework, runtime, and hard constraints in the first line of the prompt. Models default to safe averages otherwise.
vibgineer · a working catalog
Curated by hand. Free forever. If you write code with AI, this is the bookmark you don't lose.
◆ start here
Cursor is the default. Tab-completion that understands your repo, agent mode for the multi-file moves.
see cursorAdd a CLI coding agent on top. Claude Code if you're starting fresh; Codex if you're on the OpenAI side.
see claude codeMods, skills, and MCPs that earn their place on a working machine. Install what fits your stack.
browse mods§01 / the toolkit
Six categories of pieces that compound. Browse what's already documented; the rest are coming as I get to them. Entries with write-ups are tools I run on my own machine.
The IDEs, CLIs, and agents you live in. Daily drivers and the products that round out the rest of the day.
Model Context Protocol servers. The tools your agent can actually call — files, browsers, databases, issue trackers.
Hooks, statuslines, memory layers, slash command packs. The smaller pieces that turn a stock agent into yours.
SKILL.md packs that teach an agent specific workflows or domains — design, brand, content, code review, generative work.
When you're building agents, not just using them. The SDKs and orchestrators worth knowing.
Run multiple agentic sessions in parallel without losing the plot. Terminal multiplexers that make multi-agent work bearable.
◆ from the manifesto
“The vibe coder is not replacing the engineer. They are the field condition — a noisy, parametric system that produces emergent software when tuned by someone with taste.”
— vibgineer · field notes / vol. 001
§02 / tactics
State your framework, runtime, and hard constraints in the first line of the prompt. Models default to safe averages otherwise.
Switch to plan / agent mode for any change touching more than two files. Cheaper than fixing a sprawling diff after the fact.
Make the model open and quote related files before letting it write new code. Hallucinated APIs vanish when the source is in context.
Prompt caching cuts cost ~10x on repeated context. Worth it for any agent loop that runs longer than a single turn.
Don't let an agent ship code AND write tests AND update docs in one breath. Sequence the work so each diff is reviewable.
Models silently improve and silently regress. Pin the exact snapshot for production tools and re-evaluate on a schedule.