sokowirowka.dev

my daily tech stack

A short, opinionated tour of the tools I work with every day, Claude Code, Neovim, and a local-AI stack, plus a few notes on getting more out of Claude Code. If you take one thing from this page: run /insights first.

claude code

Start with /insights. It reports your token spend, tool usage, and session length across the last month, a useful baseline before changing skills, hooks, or your CLAUDE.md.
Skills I use most. superpowers:brainstorming for turning "build X" into a written spec before code is written. solana-dev as a reference playbook for Solana wallet and RPC patterns.
Local docs index before starting work. I use lv-local, my own MCP server (going public once I'm happy with it), to embed the relevant docs and books, then call search_code from inside Claude Code.
Three agents for non-trivial work. Executor writes the code, auditor reviews it against the spec, tester runs and extends the test suite. Each agent gets a narrower context.
Audit, cross-check, fix, re-grade. Run the audit, have Claude cross-check each finding against the source, fix in order CRITICAL → HIGH → MED → LOW with one commit per group, re-run tests after each tier, then re-grade the audit doc.
Eval-driven loops for LLM code. For voice agents, parsers, classifiers: write the eval harness first, then iterate against pass-rate.
Smoke-test the real boundary. cargo check and unit tests don't cover integration, run the actual HTTP / MCP / runtime path before marking a task done.
Lesser-known commands worth knowing. /rc (/remote-control) lets external scripts or scheduled agents push prompts into a running session. /compact manually compacts context before auto-truncation. /doctor when auth, MCP, or hooks feel off. /cost mid-session for token spend.
Hooks I keep around. UserPromptSubmit for project-specific reminders. PostToolUse matched on Edit|Write to run cargo check and cargo clippy after edits, so compile and lint errors surface before the next prompt. A Stop hook to run the test suite at end of session.
Loop. plan → execute → verify (cargo check && cargo clippy && cargo test) → commit.

neovim

How I landed here. VS Code → Cursor → JetBrains → Zed → Neovim. They all worked. Working in nvim is the one I find most satisfying day to day.
From scratch with lazy.nvim. No distro. init.lua + a lua/plugins/* tree, one file per concern. Easier to reason about than LazyVim/AstroNvim once you know what you want.
Plugins I actually use. telescope for fuzzy-find, nvim-treesitter for syntax, mason + nvim-lspconfig for LSP wrangling, nvim-cmp for completion, hop.nvim for motion, trouble.nvim for diagnostics.
Rust gets its own setup. rustaceanvim + crates.nvim, rust-analyzer with sane defaults and inline crate version hints in Cargo.toml.
AI lives inside the editor too. codecompanion.nvim wired to a local Qwen 3.6 27B over an OpenAI-compatible adapter. Inline prompts, chat buffer, no cloud round-trip.
Sits next to Claude Code, doesn't replace it. Terminal split for claude, nvim for the actual editing. Two tools, one job each.

local ai

Qwen 3.6 27B on llama-server. One process, OpenAI-compatible endpoint at 127.0.0.1:8081, 131k context. Used as the default model for everything that doesn't need frontier-tier reasoning.
lv-local for RAG over my own docs. My MCP server (going public once I'm happy with it) indexes Markdown, PDFs, EPUBs, and source trees into a local embeddings store, then exposes search_code over MCP. I point it at course notes, books like Clean Architecture, and project repos.
Same model in Neovim via codecompanion. codecompanion.nvim talks to the local llama-server over an OpenAI-compatible adapter, so chat and inline prompts hit the same process I'm already running, no duplicated weights, no second runtime.
Why local. No per-token cost, no data leaves the machine, and it's fast enough for short prompts and inline edits.