Write prompts you can check
Turn a vague request into a repeatable recipe for rewriting, extracting, or summarizing text.
Learning path
Build projects
Start with prompts and structured data. Then build document search, process images and audio, or connect a model to a tool you control.
Prompts, coding, and structured outputs
5 lessonsGet structured JSON from a model
Extract fields into a predictable shape and validate the result before your app uses it.
Stream a local model’s reply
Print model text as it arrives, measure time to first visible text, and detect an interrupted stream.
Ask a local coding model for a change you can test
Reproduce a bug, prompt a local coding model with the contract, and evaluate its proposed Python change with unit tests.
Build a short meeting brief with source lines
Break notes into sections, ask a local model to select decisions or actions, and audit each selection against the original text.
Document search and answers
12 lessonsAsk questions about local documents
Build a small retrieval-augmented generation example with local embeddings and citations.
Build a persistent text search index
Index local text files in SQLite, retrieve exact terms with source locations, and refresh the index when files change.
Store embeddings and search by meaning
Save local embeddings with their source text and model identity, then use them to find paraphrases across sessions.
Extract text and page numbers from a PDF
Read a sample handbook page by page and identify which pages need optical character recognition.
Build and update a hybrid PDF search index
Combine SQLite keyword results and local embeddings while keeping each hit linked to its PDF page.
Test PDF retrieval and verify cited quotes
Score retrieval against known source pages, then use a local model to quote evidence you can check in the PDF.
Find documents across languages
Compare two local embedding models on six English, Chinese, and Spanish query–document pairs.
Rerank retrieved passages with a local model
Score question–passage pairs, then measure whether reranking improves the first result on the sample handbook.
Extract PDF tables without losing the rows
Read a ruled equipment table and preserve its column labels, cell values, and page, table, and row numbers.
Find the policy that applied on a given date
Look up a table entry across archived PDF editions and check that removed entries do not reappear from an older source.
Index prose, scanned notices, and PDF tables together
Build one searchable collection with dated table rows and citations that identify the original page and OCR text.
Answer a question with evidence from several documents
Search separately for each part of a question and validate a cited quote for each part before displaying it.
Images and scanned text
4 lessonsRead images with a local model
Try a vision-capable model on a screenshot and check what it can and cannot read.
Extract text from a scan with OCR
Use Tesseract on a sample notice, check dates and room numbers, and make the result searchable.
Test whether a vision model reads the right fields
Score exact fields on a sample notice and see how the model handles an intentionally hidden value.
Generate and compare images on your computer
Use a local diffusion model to make an illustration and run a controlled comparison of seed, prompt, and device.
Speech and audio
3 lessonsTranscribe a recording on your computer
Run Faster Whisper on a short recording, save timestamped notes, and add them to a searchable collection.
Turn a local text file into speech
Generate a WAV file with Piper, compare speaking rates, and check the spoken details against the source text.
Build an offline voice guide to the handbook
Transcribe a recorded question, retrieve and check a handbook quote, then generate and time a WAV reply on your computer.
Assistants and tool use
3 lessonsExpose a read-only tool with MCP
Start a local MCP server, discover its tool and resource, and check how it handles an unsupported argument.
Connect a local model to the MCP tool
Translate an MCP tool schema into a model request and verify the proposed call before the host executes it.
Review a model’s proposal before saving it
Have a local model propose a fictional reservation, then test confirmation, cancellation, missing details, and duplicate requests.