What’s inside

One SKILL.md (the playbook) + one starter template. A preview:

Digital Handyman

When to use

• You have a pile of small chores (rename files, tidy a folder, build a slide deck, clean a spreadsheet) and no desire to do them
• Your downloads folder is full of old installers eating gigabytes and you want them gone safely
• You need a quick fetch (a video transcript, a file moved, a wiki page started) without learning a new tool
• You want recurring digital chores (calendar scheduling, reward-balance logging) handled semi-automatically
• A browser step is blocked (login wall, CAPTCHA) and you need a fallback plan

Workflow

Describe what needs doing and what "done" looks like. Vague asks get a plan back for confirmation first — a coding agent with a shell has infinite tools, which is exactly why the target needs to be clear.

1. Say the chore in plain language (input: your words → output: task + acceptance criteria)

For anything touching files, accounts, or the system, the agent shows the plan — and the full file list for renames and deletes — before acting. Plan-first is the price of admission for chores with side effects.

1. Show the plan first (input: task → output: execution plan)

Deletions, overwrites, and irreversible actions always need your explicit yes. One real case scanned a downloads folder, listed 31 leftover installers totaling 4.4GB — and deleted nothing until the list was approved.

1. Approve the risky bits (input: plan → output: approval)

Execute step by step and show results along the way so you can catch mistakes early. A blocked browser step gets a fallback rather than a dead end — in one real case the agent installed an open-source tool to fetch a YouTube transcript when the browser was blocked. Tell the user what was installed.

1. Do the work in stages (input: approved plan → output: partial results)

A short log: what changed, what was deleted, what was installed, and what needs your follow-up. No summary, no trust — the user can't verify what they can't see.

1. Report what changed (input: done → output: summary)

Output spec

• Task list with acceptance criteria, execution plan shown before acting, completion summary (changed / deleted / installed)
• Approval log: what was approved and when
• Fallback notes: what failed and what replaced it
• Companion template.md skeleton — fill in the blanks and use it

Good vs. bad examples

• ❌ Bad cleanup: "Clean my downloads" → agent deletes 31 installers (4.4GB) without showing the list — one of them was a license-key archive you needed
• ✅ Good cleanup: agent scans, lists all 31 installers with sizes and the 4.4GB total, you approve, then it deletes and reports the freed space
• ❌ Bad browser task: agent keeps hammering a login wall with retries and guesses at your password
• ✅ Good browser task: b

Download the full skill pack (ZIP)