Blog

Blog.

What I built and how: the stack, the architecture and what I'd change next time.

20 posts · RSS feed

Latest · · AI

Same agent, rebuilt workspace, a quarter of the bill: what EnvOpt shows

A new arXiv paper leaves the model and the harness alone and lets another AI rebuild the workspace the agent works in. On 200 test tasks, the same Claude Haiku 4.5 agent cost $42.36 before and $11.55 after, and passed more tasks.

7 min read · Read the post →

2026

  1. CLAUDE.md vs AGENTS.md: which one Claude Code reads, and why I keep only one

    When a repository has both files, Claude Code reads CLAUDE.md and leaves AGENTS.md alone unless you import it. I found five of my repositories set up the wrong way round: the real instructions in AGENTS.md, and a CLAUDE.md that only pointed at it.

    How I buildAI
    5 min read
  2. Running several Claude Code agents at once: worktrees, handoffs and one check per phase

    What broke when I put many Claude Code agents on the same repositories, and the four rules that stayed: one worktree per agent, small file-disjoint jobs, a rewritten handoff, and one full check per phase.

    How I buildAI
    5 min read
  3. Your agent finished the task. Did every step work? What prefix executability shows

    A new arXiv paper replays every tool call an agent made and marks the first one that could not run. On 200 function-calling tasks, one agent finished 94 % of them, yet only 47 % of its runs ran clean from start to end.

    AIpapers
    5 min read
  4. CLAUDE.md, hooks, skills or the task: where each rule for Claude Code belongs

    Claude Code gives you four places to put a rule, and they differ in when they load, what they cost and whether they are enforced. How I decide, with the rules in my own setup that moved from one place to another.

    How I buildAI
    6 min read
  5. Same local model, twice the right answers: what the MuninAI paper changes

    A new arXiv paper puts the open-weight model a research group already runs inside an agent loop over its own papers. Accuracy on 199 literature questions went from 38.7 % to 87.4 %, and the usual top-5 retrieval scored worse than no retrieval at all.

    AIpapers
    5 min read
  6. What Claude Code really costs: tokens per session across thousands of runs

    Tokens and list-price cost per finished Claude Code session, from thousands of my own sessions: what fills the context window, why most of the bill is re-reading, and the changes that took one tier from $32.86 to $1.75 a session.

    How I buildAI
    6 min read
  7. Running many coding agents at once: worktrees, handoffs, one check per phase

    The two parts of a many-agents setup that break quietly: the worktree each session works in, and the handoff file a session leaves for the next one. Setup, traps, cleanup and a template.

    How I buildAI
    7 min read
  8. Claude Code on a server: one copy of the work, and you attach to it

    Closing the laptop killed 27 % of my long agent runs. Syncing files to a server is the wrong fix. Here is the one that works, and an open-source setup you can run.

    How I buildAI
    5 min read
  9. Claude Code hooks: run each one in warn mode for a week first

    A hook that blocks the wrong thing stops good work. Mine logged for a week before any of them blocked. One of the rules fired 307 times and was right zero times in the sample I checked.

    How I buildAI
    5 min read
  10. My CLAUDE.md, and the failure behind each rule

    Every line in my CLAUDE.md is there because something went wrong. Here are the rules, what broke, and what an instruction file costs you on every step.

    How I buildAI
    6 min read
  11. Rows of small terminal windows, each with a thin line running out, all the lines merging into one.

    How I run thousands of Claude Code sessions without losing the plot

    One task per session, written briefs, handoffs, hooks, what it costs and what went wrong.

    How I buildAI
    12 min read
  12. A cable plugged into a switchboard with several possible paths, only one of them lit.

    Ask for a role, not a model: the small AI gateway behind my sites

    Code asks for “summarise” or “translate”, never for a model name. One table decides the rest.

    AISvelteKit
    14 min read
  13. Sitely — AI website generation platform for Czech businesses

    Why AI-generated websites all look the same, and what I built to stop it

    A hundred prompts, the same page a hundred times. The seeds and checks I used against it.

    AISvelteKit
    3 min read
  14. Columns of news text falling toward a globe and turning into map pins.

    Meridia: getting places out of news text

    News says "the region" and "local authorities". A map needs a place.

    ProjectsAI
    1 min read
  15. A car on an empty road at night, its signal fading above it while a screen in its window stays lit.

    AutoPlakáty: keeping ads playing with no signal

    Tablets in moving cars lose signal all the time. The screen still can't go blank.

    ProjectsSvelteKit
    1 min read

2025

  1. A vase being 3D printed layer by layer next to an identical outline of the same vase.

    Niro3D: one price in the browser and on the server

    The price you see while configuring has to be the price at checkout. One package does both.

    ProjectsGoSvelteKit
    1 min read
  2. A car wheel and two spanners beside a tidy grid of small squares.

    Lucky Autoservis: local pages in two languages

    One matrix of services, car makes and places, in Czech and Russian.

    ProjectsSvelteKit
    1 min read
  3. A sheet of paper sliding onto a neat stack inside a round vault door.

    ReadVault: a Chrome extension

    A reading list that stays on your machine.

    Tools
    3 min read
  4. A storm of small red shapes stopping at a brick wall, with calm empty space behind it.

    BlockAid: a website blocker for Chrome

    Blocking distracting sites while you work.

    Tools
    3 min read