You don’t need a different AI subscription for every task. One local model behind Ollama covers the everyday list — privately, offline, free after the download. This guide is the practical playbook; the deeper setup lives in the brain guide.

The one-time setup

Install Ollama, pull one good general model sized to your hardware (the brain guide covers picking by VRAM — as a rough rule of thumb: 8 GB cards run 7–9B models, 16 GB runs ~14B, 24 GB+ runs 27B-class comfortably). Then every use case below is the same model wearing different instructions.

1. Chat with your PDFs

The zero-infrastructure version: extract the text, paste it in, ask questions.

pip install pypdf
python -c "from pypdf import PdfReader; print('\n'.join(p.extract_text() for p in PdfReader('doc.pdf').pages))" > doc.txt

Then: “Here’s a document: [paste]. Answer only from it, and say ’not in the document’ when it isn’t.” That last clause is the anti-hallucination seatbelt — use it every time. For whole libraries of documents, graduate to a RAG tool; for the contract, manual, or paper in front of you, paste-and-ask wins on simplicity.

2. Long video → key takeaways

Transcribe locally with Whisper (works on any video or audio file), then: “Summarize into key claims, then list anything presented as fact that deserves verification.” A 40-minute video becomes 2 minutes of reading — and that second clause makes the model flag the shaky parts instead of laundering them.

3. Voice memos → organized notes

Same Whisper pipeline, different prompt: “Turn this rambling transcript into: decisions made, action items with owners, open questions.” Ramble on your commute; arrive with minutes.

4. Any document → personalized study guide

“Create a study guide from this: key concepts in plain language, then 10 quiz questions from easy to hard, answers at the end. I’m at [your level] — pitch it there.” Then — the real move — answer the quiz and paste your answers back for grading. Retrieval practice beats re-reading; the model is a tireless quizmaster.

5. Interview prep with teeth

“You’re interviewing me for [role, paste the job listing]. Ask one question at a time, follow up on weak answers, don’t move on until I’ve answered well. Debrief at the end.” The one-at-a-time instruction is what makes it a rehearsal instead of a listicle.

6. Fact-check the AI with the AI

For anything that matters: “List the factual claims in your last answer and rate your confidence in each; flag which ones I should verify.” Models are much better at spotting their own weak claims when asked directly than at not making them — this catches a surprising share.

Why local for all this

These are your contracts, your health documents, your voice, your job hunt. On your GPU, nothing leaves the room, there’s no per-question fee, and it works when the internet doesn’t. Wire the same brain to voice, a face, and remote access, and it stops being a chatbot and starts being infrastructure.