Short episodes showing exactly how the systems get made — the prompts, the wins, the glitches. Watch how it's built, steal the code, or hire the operator behind it.
Each episode is one real thing I built with Claude Code — shown end to end, with the exact prompts and the free code to run it yourself.

A scheduled Claude Code run checked my other AI sessions every 15 minutes and re-read the same idle tabs each time: 374 runs and 9.9M Fable output tokens in one week, the weekly limit gone two weeks running. The fix is a content fingerprint per tab and an exit-code gate that only starts the model when something changed.

My NAS said the disk was nearly full and could not say why. I had Claude Code list every file on it, 12.2 million rows into one SQLite table, with only ls, find and du allowed to run on the NAS. It built a page that sums what I tick. The biggest gap turned out to be Synology Drive version history, 4.2 TB of it, invisible to a normal file listing. It deleted nothing; the delete key stays with me.

AI agents merge their own tickets into my dashboard. I wrote a deterministic tripwire that judges every merge after it lands, ran it silently over the last 50 merges, fixed three false fires with a test each, and wired it in as a shadow step that logs and never blocks. Then it flagged its own merge.

I asked my local voice assistant 24 questions about my own day and it made things up in 7 answers. A test that counts lies plus four strict lines in the system prompt took that to zero, and the reply now starts in a third of a second.

My AI agents work on branches and one button merges their work to main. It died half way in 44 of 65 presses. A read-only preflight, a gate earlier in the chain and a Fix line on every stop turned the failures into one-line answers, and a tally keeps the score.

Upload audio, get a transcript, nothing lingers. With Claude Code I built a whole transcription product in 20 commits: audio deleted the moment the job finishes, the transcript on first download, everything else gone within 24 hours. The rules came before the features.

My Pocket Operator gateway had one tidy line: if no billing secret is set, reuse the connect secret. It meant the key for the account-linking door also opened the billing door, and nothing would ever warn me. Now unset means billing is off, and equal secrets make the gateway refuse to start.

I gave a free local model a 179-rule rulebook and one web page. It proposes one small patch at a time, a measuring tool decides keep or revert, git remembers every keep, and the tool has no publish button.

Every prompt in my AIOS went to the big cloud model by habit. I tested 12 free local models on a real job, wrote every score to a database, and built a router that reads it: it names the cheapest model that measured good enough, and when it has no data it says so instead of guessing.

I pointed my AI at five live role postings and it built a competency matrix against my own profile — Have, Partial, Missing, ranked by fit — before I'd read a single line. A skills radar, not a scroll.

The Telegram bot has been the AI's mouth for months, but Telegram is a channel between me and it - Slack is where a decision gets made in a thread on a Tuesday and never gets written down anywhere else. So it got the same read access I have: channels, threads, DMs, full-text search. scripts/slack_client.py wraps the Slack Web API in eight subcommands; the first six are read-only and unguarded, because reading a workspace is a private act. Posting into one is not - it lands in front of other people, under my name, with no undo. So send refuses by default: it composes the message, prints it, and waits for four characters of confirmation. Getting there cost two token traps and a scope typo - a bot token can't see DMs at all (it needs a user token), usergroups:read is not groups:read, and a token pasted where it shouldn't have been took the whole encrypted secrets file down with it.

My AIOS redesign forked into three parallel prototypes with no clear winner. I brought in Fable 5 — a brand-new Claude model with no history — to scout all three plus production and pick the one that deserved to survive. It did, and then it told me not to build a fourth.

My overnight AI audit flagged a consulting proposal I'd let go cold for five weeks. By morning it had workshopped the pitch three ways and generated the full styled proposal — and a human just has to hit send.

I said "on screen 2, tab 3: implement the plan" out loud — and the right Claude Code session on the right monitor picked it up. A free spacedesk screen fabric, a Win32 window router, and a regex intercept that never spends a token.

I push one commit to GitHub and thirty seconds later Cloudflare Pages has rebuilt and redeployed the live site — no server, no FTP, no deploy button. The pipeline that used to be a whole DevOps role is a box you tick once.

Skills, commands and agents pile up in your global config, inside every project, and in plugin caches — no single view, and duplicates drift out of sync in silence. I had my dashboard build a registry that walks every location, catalogs all 99, and surfaces the strays in one search box.

The proper fix for tracking loose ends is a dashboard heatmap (todo #34) — but that needs a desktop session. So I built the usable-today version first: inbox/open-threads.md, one line per open thread, closed by texting 'close 3' from any device. Writing it down became a free audit — 11 loose threads surfaced the moment it existed, two I'd have sworn were done. The scrappy text file that beat the fancy dashboard. Channel A, on camera.

A master-architect agent sweeps every project folder in the workspace each night to decide what's alive and what's stale. Two folders with perfectly written plan files passed as real work while their content folders held nothing but a .gitkeep — so the discovery filter learned to look inside.

My AIOS rewrite grew a slick new Todo board — then every row was leaking raw **markdown**: literal asterisks and backticks in a UI that's supposed to look done. The fix: a RichText renderer for the drawer, a plainTitle helper that strips markdown from every row title.

An overnight render died halfway and my Telegram AI said nothing — a per-message claude -p process exits after replying, so anything it backgrounds dies untracked. I built run-detached-notify.ps1 (one guaranteed Done/FAILED message on finish), then taught the same bot to read my 👍/👎 and mail a weekly report card (weekly-feedback-digest.ps1). The AI's feedback loop, closed: it reports back and it listens back. Channel A, on camera.

One push to GitHub, and /release fans out lint, test, commit, push and deploy across roughly 20 repos — the DevOps job that used to be a manual per-repo chore is a single command now.

All the tests passed, I pushed, and two minutes later GitHub mailed a Run failed notice with seven red tests on a public repo. The tests weren't wrong - this laptop is too rich. It has every library and API key I've ever installed, while the CI runner installs exactly three things: pytest, detect-secrets, bandit. So two test files quietly leaned on what only exists here and died on a clean machine. Which makes the obvious fix useless: running them again locally reproduces the same blind spot, because green on my machine is not evidence of anything. The gate is a pre-push hook that runs a throwaway environment holding exactly what the runner holds - same three checks, same order - and if any fail the push never leaves the laptop. Two details make or break it: the hook is pinned to Unix line endings (one stray carriage return and it stops running silently, which is worse than no gate at all) and it's committed as executable, or git just skips it.

One card reads every running process, maps it to a service and a port, and flags collisions. It caught two services fighting over port 8080 the instant it loaded.

Most portfolios list tools. I built a proof page that shows the delta. Five real wins, each one the same shape: the thinking constraint that was blocking me, the KPI before and after, and the AI system that moved it — 2-4 h to 0 h, 20 min to 3 min, scraping to one command, silent context deaths to zero. Every number links to the episode that proves it. The story of building showcase/cases.html.

My phone voice AI needed a trusted HTTPS URL. I was using Tailscale — account, device enrollment, DNS. Then I found cloudflared quick tunnels: zero account, one command, a live URL with a real cert in seconds. Wrapped it in run-tunnel.ps1 and ripped Tailscale out.

One sentence in, a full Seedance 2.0 shot list out. I built a Claude Code skill that writes the script for any AI-video idea — copy-paste prompts, CapCut steps, a credit budget. Here's the anatomy.

I built a dashboard button that opens any past Claude Code session in the browser, shows the full parsed conversation — tool calls and all — then resumes it in the terminal with one click.

My outboard died twice at sea with no fault code and restarted instantly. I fed the real manuals to Claude Code — it built a cited knowledge base of the whole boat, ranked six suspects (kill-switch stop-circuit #1), and handed me a free dockside test plan before any mechanic. Off-brand domain, on-brand method: AI debugs a real-world problem from the manuals up. Channel A, on camera.

Running AI at home burns real money. So I had Claude build a two-layer router: one flag picks the cheapest model that's still good enough, then the cheapest host to serve it. Same prompts, same results, a fraction of the bill.

One cron job at 01:00 fans out 20 AI agents that audit every repo on my machine for health and security, then write a single ranked morning report — done before I pour the coffee.

Fifteen daily Claude Code dares scripted in one night build — 9 built-in commands + 6 anonymized power-user shortcuts, one caption-first 9:16 spine.

Six parallel read-only Claude agents read 946 of my own prompts across 351 sessions, found my exact pain points, and handed me a five-pillar redesign. Here's the method, so you can audit your own AI usage.

Everyone opens the polished Claude app and desktop. I mostly use a Telegram bot and a bare terminal instead — same AI brain, but they sit next to my real files, inbox and scripts, so they can actually DO things. Page tag EP.11; built by the night build.

A client wanted to see how the automation actually works. So I described one real workflow out loud and Claude wired the whole n8n pipeline node by node, then fired it: draft episode to live URL in one click.

88% of companies use AI. Only 6% are good at it. 30% of AI projects get abandoned. That gap is the entire opportunity. Here is the consulting pitch, the doctor-vs-pharmacist positioning, the Constraint→KPI→Build method, and the 60-second Matrix-operator video script.

Siri can't reach my files and the Claude app has no desktop voice. So I built a push-to-talk PWA that installs to my phone's home screen, routes each question across two AI tiers, talks back out loud, and already knows every project I'm running.

The sharpest ideas happen mid-session and vanish when you close the tab. I taught an AI to read my saved sessions back, score each chunk for novelty and surprise, and hand me a pre-qualified backlog — each idea carrying its own source transcript.

I'm not a filmmaker. But I built a skill that gave Claude the script, Seedance 2.0 the shots, and CapCut the edit — 18 seconds, two generations, zero camera. The 6-frame gallery shows the exact prompt behind each shot.

A faceless daily series: one small Claude Code move a day. 21 dares in two tiers — built-in commands and skills I built myself — each one you can try in five minutes.

The n8n media agent plus paid poster, rebuilt inside Claude Code. Type a topic: it expands the prompt, fires kie.ai → Sora 2, burns in captions, trims dead air, and queues the Reel in self-hosted Postiz. $0/mo on the posting.

Long sessions die silently as context fills. I had the AI build its own cockpit status line — model, effort, and two live fuel bars — so I know exactly when to /compact.

Agents scrape a site built for human eyes badly. So I gave mine an llms.txt, a JSON index, a tiny CLI and an MCP server — now any Claude reads the whole site from one command.

A 20-minute tutorial I never played: one slash command cut it into 80 frames, read 565 transcript segments — then three research agents proved its money claims were ~4× inflated.

Half past eight I gave it the idea; it asked six questions. I answered two from the couch and fell asleep. By one in the morning it had read the brief and built this whole page alone — then left a note listing every guess it made.

The AI-built globe got personal: 1,782 LinkedIn contacts stream onto it as a living layer — thin pins, a terminal at 50 names/sec, and three real packet-trace routes. One toggle fades it all back to the plain globe.

A rotating, futuristic Earth for the brand, designed in one chat on claude.ai/design. Every prompt, the palette it guessed from a screenshot, and a live globe you can drag to spin.

Five art directions fought a style-off over the same screen. The winner became the house style — and sixteen screens later, the look is mine.

One prompt mapped my entire AI OS as a sketch — and the model quietly leaked raw colour codes into the labels. The glitch, and the 5-cent fix.

One skill turns a plain prompt into a hand-sketched diagram — no design tool, ~5 cents a render.

Six drawn candidate boards in two rounds, a winner called Calm Lagoon — and the moment the AI rejected my favourite shade for accessibility.
Each case: one thinking constraint that was blocking progress, the metric that proves it's gone, and the AI system that fixed it.
The client and enterprise cases: what was blocking, what moved, and the system that ran it.
I turn messy, real-world product operations into AI systems that actually run. Everything on this site, I built myself with Claude Code. I'm open to a Product Owner / AI role in the Malmö–Lund region and to select consulting work.
No noise — just the episode, the prompt, and the code, the moment it ships.
No inbox? On the videos, comment BUILD and the bot DMs you the link.
Every episode's free, open download in one place.
LiveHire the operator behind the builds — role or consulting.
LiveLearn the AI-agent-operator playbook alongside others.
Coming soonThe product: a private AI assistant per person, now Pocket Operator.
Invite-only beta