5 things to build in Claude this weekend (the prompts)

Build me a daily morning briefing that sends me one message at 07:30 Europe/London every day, by Telegram (bot token and chat id from a .env file in ~/briefing/). Inputs: my Gmail inbox from the last 24 hours (read-only), my Google Calendar for today and tomorrow, and a plain text file ~/briefing/watchlist.md where I list deals, clients and deadlines I care about. The message has exactly three sections. MOVED: what changed since yesterday (new replies, status changes, meetings booked). NEEDS YOU: anything waiting on a decision or reply from me, oldest first. AT RISK: anything past its deadline or with no reply for more than three days. Keep it under 200 words, plain text, no emojis, no tables. If a section is empty, write 'Nothing.' Do not make anything up: every line must trace back to an email, an event or a line in watchlist.md. Put the code in ~/briefing/briefing.py, write each day's message to ~/briefing/sent/YYYY-MM-DD.txt, add a cron entry for 07:30, and add a --dry-run flag that prints the message instead of sending it. Run a dry run and show me the output before you schedule anything.

Set up a second brain for me as an Obsidian vault at ~/brain/. Create these folders: People, Clients, Projects, Daily, Inbox, Templates. Every page is a markdown file with YAML frontmatter (type, name, status, last_touched) and these sections. People: who they are, company, how we know each other, last contact, open threads. Clients: what we do for them, contacts, current state, next step. Projects: goal, status, decisions made and why, next action. Put one template per type in Templates/. Link pages with [[wikilinks]] so a client page links to its people and projects. Create a CLAUDE.md in the vault root that tells you how the vault is organised, how to name files (Firstname-Lastname.md, client-name.md, project-name.md), and the rule that you never delete a page, you only mark it archived. Write ~/brain/scripts/daily_note.py to run every evening at 21:00 via cron. It reads the day's sent and received emails (read-only) and my calendar, then writes Daily/YYYY-MM-DD.md with: what happened, who I spoke to, decisions made, open loops. For any person or project mentioned, append a dated one-line entry to their page and update last_touched. If the page doesn't exist, create it from the template. Start by creating the structure and five example pages using made-up data, show me the tree, and wait for my go before connecting any real email or calendar.

Build me a review agent in ~/review/ that runs entirely from cron and never talks to me unless I ask. Daily log, every evening at 21:30: collect my git commits across the repos listed in ~/review/repos.txt, the files modified today under ~/work/, and the entries in my vault's Daily/ note for today. Write ~/review/logs/YYYY-MM-DD.md with factual bullet points under three headings: Done, In progress, Blocked. State what happened. No advice, no praise, no suggestions. Weekly roll-up, Sunday at 18:00: read the seven daily logs and write ~/review/weekly/YYYY-WNN.md with what shipped, what carried over from last week, and anything that appeared in the logs three or more days running without moving. Sunday health check at 18:30: write ~/review/health/YYYY-MM-DD.md that checks the machine and the system itself: disk space, the last run time and exit status of every cron job in this folder, whether each daily log exists for the last seven days, and any script that errored. Report PASS or FAIL per check with the evidence. Use plain Python and the Claude Code CLI in headless mode for the summarising. Log every run to ~/review/run.log. Test each job once by hand and show me the output files before you add the cron lines.

Build me a brand kit in ~/brand/ and a script that generates reel covers and captions from it. Folder layout: colours.json (primary, secondary, background, text, each as a hex value; I'll give you mine, ask me if you don't have them), fonts/ (the font files I put there, plus a fonts.json naming the heading and body fonts), templates/ (three cover layouts as HTML files at 1080 by 1920 pixels: big title, title plus subtitle, numbered list), captions/style.md (my caption rules: sentence length, tone, words I never use, how I end a caption, the call to action format), and out/. Write ~/brand/make_cover.py. Usage: python make_cover.py --template big-title --title 'Your title' --subtitle 'optional'. It fills the template from colours.json and fonts.json, renders it to a PNG with Playwright, and saves it to out/ with a slug of the title as the file name. Write ~/brand/make_caption.py. Usage: python make_caption.py --transcript path/to/transcript.txt. It reads captions/style.md and the transcript, calls Claude in headless mode, and writes three caption variants to out/ as plain text. Make the three templates, generate one test cover from each using a sample title, and show me the PNGs. Keep the code simple enough that I can change a colour in one file and regenerate everything.

Build me a life library in ~/library/ that stores everything I read and watch and lets me search it. Inputs: a folder ~/library/inbox/ where I drop files (PDF, EPUB, markdown, text), a file ~/library/links.txt where I paste article and YouTube URLs one per line, and a notes.txt for anything I add by hand. A script ingest.py runs every night at 02:00 via cron. For each new item it extracts the text (trafilatura for web pages, yt-dlp for YouTube captions, pypdf for PDFs), saves the clean text to ~/library/items/ID.txt, and asks Claude in headless mode for a short summary, five tags and the three ideas most worth remembering. It stores title, source, date added, summary, tags and ideas in a SQLite database at ~/library/library.db with full-text search on the text and the summary. A script search.py takes a plain English query, finds the best matching items with SQLite full text search, and prints title, source, date and the matching passage. Add a --ask flag that sends the top five passages to Claude and returns an answer with the item it came from for each claim. Never invent a source: if nothing matches, say so. Start with five sample items, show me a search working end to end, and tell me what you couldn't test.

Help me set up a small Linux server for running scheduled Claude Code jobs. I have just rented a basic Ubuntu VPS and I can SSH in as root. Walk me through it one step at a time and run each command with me. 1) Create a non-root user with sudo, disable root login over SSH, and switch to key-only login, but confirm with me before you restart SSH so I don't lock myself out. 2) Turn on the firewall allowing only SSH. 3) Install Node.js LTS, Python 3, git and cron. 4) Install Claude Code with npm and walk me through logging in. 5) Create ~/jobs/ with a logs/ subfolder and a run_job.sh wrapper that runs a command, appends the output with a timestamp to logs/, and exits non-zero on failure. 6) Add a crontab with one example job at 07:30 that writes the date to a log, wait for it to fire, and show me the log. 7) Set the server timezone to Europe/London. Print a checklist at the end of what is installed and what is scheduled. Do not install anything I haven't approved and never print my keys.