5 AI projects you can complete this weekend (vol. 3)

A job-application agent

Intermediate, about 8 to 12 hours. A browser agent that reads your resume and autonomously fills those forty-minute multi-page application forms while you watch and approve. Build it once, reuse it for every application, and screen-record it doing your most hated chore.

1. 🧰 Set up Browser Use. It's open source and free; Claude can walk you through the install and connect an LLM key (a few dollars covers a weekend of runs).
2. 📄 Give it your facts. A single profile file: resume, work history, standard answers (visa status, notice period, salary range). The agent pulls from this, never invents.
3. 👀 Run supervised. Watch it fill each form, approve before submit, and keep a log of what was sent where. You stay the applicant; it does the typing.

I'm setting up Browser Use to fill job application forms from my
resume while I watch. Walk me through it as a non-engineer:

1. The exact install steps for my machine, and how to connect
   the LLM key.
2. Design my applicant profile file: every field job forms ask
   for, so the agent never has to guess.
3. Write the agent's instructions: fill from my profile only,
   never invent answers, pause for my approval before every
   submit, and log each application to a CSV.

A job-fit scout

Intermediate, about 6 to 9 hours. An n8n agent that pulls new listings every morning from job feeds, scores each against your profile with written reasons, and drafts tailored outreach for the top matches. Using an agent to run your own search is proof of skill and a portfolio piece at the same time.

1. 📡 Pull the listings. RSS feeds from job boards, or an Apify scraper on your target companies' career pages, into n8n on a morning schedule.
2. ⚖️ Score with reasons. Claude compares each listing to your profile and returns a fit score out of ten plus two lines on why, so you trust the ranking.
3. ✍️ Draft the outreach. For everything scoring eight or higher, Claude drafts a short, specific note referencing the actual role. Drafts only; you send.

Design my job-fit scout in n8n. Every morning: pull new listings
from [feeds/boards], score each against my profile, draft outreach
for the top matches into Gmail drafts.

1. List the n8n nodes in order with key settings.
2. Write the scoring prompt: fit score /10 against my profile
   [paste resume summary], with two lines of reasoning, strict
   on must-haves I actually lack.
3. Write the outreach-draft prompt: 4 sentences, references the
   specific role, zero generic enthusiasm.

A meeting-to-action pipeline

Beginner friendly, about 4 to 6 hours. Every call recording becomes action items, decisions, and follow-ups, filed into Notion automatically before you're back at your desk. It kills the most hated task in every office and demos perfectly in thirty seconds.

1. 🎙️ Record with Fathom. Free tier records and transcribes your calls and produces the recap automatically.
2. 🔁 Pipe it through. Zapier or Make (free tier) catches each new recap and hands the transcript to Claude for extraction: actions with owners, decisions, follow-ups.
3. 🗂️ File it in Notion. Each meeting becomes a page in a database: attendees, decisions, action items with checkboxes. Nothing gets lost in a notebook again.

Design my meeting-to-action pipeline: Fathom records the call,
Zapier catches the transcript, you extract, Notion stores it.

1. List the Zapier steps in order with settings.
2. Write the extraction prompt: action items with owner and due
   date, decisions made, and open questions, strictly from the
   transcript, no invented owners.
3. Design the Notion database: properties, views, and a weekly
   "what's still open" view I check every Friday.

An AI red-team squad

Intermediate, about 6 to 8 hours. Five AI critics, each with a different lens (logic, evidence, finance, audience, risk), attack any pitch or strategy doc and return a severity-scored weakness report. Run it before every big meeting, and hand the pattern to your team like a tool.

1. 🧠 Define the critics. Five short persona briefs: what each one attacks, what evidence it demands, and how it scores severity from one to five.
2. ⚔️ Run them as subagents. In Claude Code, each critic runs as its own subagent over your doc, blind to the others, so you get five genuinely independent reads.
3. 📊 Merge the report. A final pass dedupes the findings, ranks by severity, and writes the one-page "fix these first" summary.

Build me a red-team system in this folder. I'll drop any strategy
doc or pitch into input/ and run one command.

1. Create 5 critic personas as separate prompt files: logic,
   evidence, finance, audience, risk. Each attacks only its lens
   and scores every finding 1-5 for severity.
2. Write the runner: each critic reviews input/ as an independent
   subagent, results merge into one deduped, severity-ranked
   report in output/report.md.
3. Run it on the sample doc I give you and show me the report.

Your own MCP server

The frontier one, about 6 to 10 hours. MCP is how Claude connects to outside tools and data, and you can build your own server that plugs YOUR data (a job tracker, a portfolio database, your notes) into Claude as live tools it can actually call. Almost nobody outside engineering can say they've built one, and Claude Code writes most of the code.

1. 🎯 Pick one data source. A spreadsheet you update, a folder of notes, a simple database. One source, two or three tools (read, search, add), nothing fancy.
2. 🏗️ Let Claude Code scaffold it. The official MCP SDKs are free; Claude Code sets up the project, writes the server, and explains every file as it goes.
3. 🔗 Connect and demo. Add your server to Claude's connectors, then ask Claude a question it can only answer by calling YOUR tool. Screen-record that moment; it's the demo.

I want to build my first MCP server this weekend, as a
non-engineer, with you writing the code.

My data source: [describe it, e.g. a Google Sheet where I track
X / a folder of markdown notes].

1. Propose the 3 tools my server should expose (read, search,
   add) and what each returns.
2. Scaffold the server with the official MCP SDK, explaining
   each file in one line as you create it.
3. Walk me through connecting it to Claude and give me 3 test
   questions that prove the tools work end to end.

How to actually finish by Sunday

Every tool in one place

• Browser Use (browser agent, open source): browser-use.com
• n8n (agent automation, free self-host): n8n.io
• Fathom (call recording + transcripts, free tier): fathom.video
• Claude Code (the build partner for projects 4 and 5): claude.com/claude-code
• MCP docs (official SDKs and guides): modelcontextprotocol.io