Kickoff
The first genuinely useful thing my AI setup ever did was tidy my downloaded music. Not a rule. Not a threat. A chore, quietly delegated.
August 2025, one day after starting my prompt library, there’s a commit: “wip: fixes for downloaded music idea.” I buy music, it lands in my downloads folder as ZIPs, and sorting it — extract, check the tags, import, archive to the NAS — was exactly the kind of chore I’d put off for weeks. So I taught the prompts to drive some shell scripts and made the machine do it.
That little pipeline outlived everything around it. The two-hundred-clause prompt library is a fossil. The music manager, in modern form, still runs today. It never once needed to be shouted at.
The Run
The ladder appears in the repo names
By September, a pattern was emerging in my repos, visible in the names alone: llm-scripts became llm-agent-local-music-manager and llm-agent-local-secrets-manager — the first time I’d used the word agent for anything I’d built. Then came llmt, an attempt at a reusable “LLM toolkit” template. Prompts had grown scripts. Scripts had grown agents. Agents were demanding a toolkit. I didn’t have the vocabulary yet. The ladder was already being climbed without my permission.
llm-scripts/
llm-agent-local-music-manager/llm-agent-local-secrets-manager/llmt/
The quiet months weren’t quiet
Then, in October and November, my home repos go quiet. Not because I’d lost interest — because I was spending that energy at work, on a brute-force prompt project I’d nicknamed “yournamehere”: helping a programme manager with their tasks, powered by nothing but hand-crafted prompts and my own persistence. [WORK-EVIDENCE: specifics and dates to confirm — kept at outcome level regardless.]
It worked. I want to be precise about that, because it’s about to sound like a complaint. It worked, and it was exhausting — every result bought with manual effort, every session started from scratch, nothing accumulated. My home experiments had shown me a ladder; work had just shown me, on a real problem with a real colleague depending on it, precisely where the bottom rung’s ceiling was.
Hold onto that name. Not because it’s a good one — it is almost aggressively not a good one — but because yournamehere comes back later in this story, transformed, filed under a heading you will not see coming.
The Debrief
The lesson from this stage is about picking first projects:
On The Training Ground
Your Own Music Manager. Build your first agent this weekend, on a chore that doesn’t matter — because the chore is the curriculum.
Info
You’ll need: an agentic CLI that can read and move files (Claude Code, or any tool with filesystem access), and one messy folder you keep meaning to sort — downloads, screenshots, receipts, photo dumps.
The ideal first-agent task is personal, recurring, file-touching, and survivable if it goes wrong. Then make it actually survivable:
cp -r ~/Downloads ~/agent-sandbox
cd ~/agent-sandboxEverything until the final step happens in the copy.
One markdown file, CHORE.md, four headings. This page — not the prompt you’ll type — is the actual engineering:
## Trigger
When I ask you to "sort downloads".
## Steps
1. Group files by type: installers, documents, images, archives.
2. Move each group to ~/agent-sandbox/sorted/<type>/.
3. Rename to YYYY-MM-DD-<original-name> using file modification date.
## Done means
Every file moved, a summary table of what went where, nothing renamed twice.
## Never
Never delete anything. Never touch a file over 1GB. Never leave
duplicates unresolved silently — list them and stop.Notice what you’re being forced to decide: what “done” means, what the failure modes are. You never decide those things when you’re prompting ad hoc.
Point the agent at the sandbox: Read CHORE.md and sort my downloads. It will do something wrong — misclassify a file, invent a category, choke on a weird filename. That’s the curriculum arriving.
This is the discipline that makes it an agent rather than a conversation: when it errs, don’t correct it in chat — improve CHORE.md so the next run can’t make that mistake. Reset the sandbox and re-run:
rm -rf ~/agent-sandbox && cp -r ~/Downloads ~/agent-sandboxTwo or three rounds of this and the page stops changing. That’s your signal.
Point it at the real folder. Run it weekly. That’s it — you own an agent now, in the only sense that matters: you’ve felt the difference between asking for an outcome and specifying a process. The rest of this series lives in that gap.
Tip
What you just proved: agents emerge from writing down structure, not from frameworks. Every failure taught you something about naming, edge cases, or “done” that no prompt guide contains — and it cost you a folder you were going to sort anyway. For a team: give every engineer their own version of this; it’s the cheapest agent training that exists.
For the record: the same folder-sorting instinct is the entire reason this site has a Mixes section and not just a Dropbox link nobody can find.