This week extends Workshop 3’s visual-analysis work into automated, local, and accessibility-focused territory. Try scheduling an agentic task in Claude Cowork, set up the Hugging Face connector and run a local model for alt-text, correct auto-captions on a recorded talk, close the loop from image analysis to image generation across several tools, draft a slide deck with Claude Design, or build a voice interface as an AI Artifact.

There is no in-person meeting.

Asynchronous expectations. Read the Required items in the menu below and complete at least one of the six exercises. Post the result of that exercise back in the cohort Discord — a screenshot, a spreadsheet, a folder of alt-text, a corrected caption file, a set of generated images, or a short reflection on what surprised you. The Discord post is the deliverable. Also do the small Workshop 4 setup at the bottom (GitHub account), because the next session depends on it.

Where This Week Fits

Workshop 3 framed image generation and image analysis as two sides of the same coin, and ended on two threads we set aside for later: Claude Cowork (Claude working in a local folder with read/write access) and local models (running the model on your own machine instead of sending images to someone else’s server). This week picks both of them up. Several of the exercises live in Cowork; one closes the loop back to image generation and asks the W5 question again — every generated image is a picture of its training data — but now with your own analyzed set as the input; and two stay in Claude proper, drafting a slide deck and building a working voice tool.

A reminder that carries over from Workshop 3: always point Cowork at a new, empty folder with only the materials you want it to touch — never a broad personal directory.

The Tools This Week

You don’t need all of these — pick what the exercise you choose calls for:

Reading Menu

Optional:

Exercise Menu — Pick One, Post to Discord

A. Schedule an agentic task in Claude Cowork (~60–90 min setup, then it runs itself)

This is the agentic move we’ll name formally in Workshop 4: instead of asking Claude to do one thing once, you give it a standing instruction and a schedule, and it runs on its own. Cowork can hold a scheduled task against a local folder.

  1. Make a dedicated folder. Create a new, empty folder (e.g. ~/cowork-agentic-watch/). This is the only place the task will read from and write to.
  2. Open Cowork and point it at that folder.

  3. Write a detailed instruction — detail is the whole exercise. A vague task (“keep me updated on AI”) produces noise. Build out the what, where, format, and when. A worked example:

    “Every morning at 8 AM, search the web for new articles, papers, and tool announcements about agentic AI applications in [your field — e.g. archival description, language pedagogy, museum interpretation], published in roughly the last 24 hours. For each item, capture: title, source, URL, publication date, a 2–3 sentence summary, and a relevance note tied to my work. Append new rows to agentic-watch.xlsx in this folder — do not overwrite existing rows, and skip anything already listed. Highlight in yellow any item that looks directly usable in my teaching. At the end of each run, write a one-line digest at the top of a digest.md file.”

    Notice what makes this useful: a named output file, an append-don’t-overwrite rule, a dedup rule, an explicit schema for each entry, and a relevance filter in your voice.

  4. A more ambitious variant. Ask Cowork to store the same data in a web-friendly form — articles.json with one object per item — and then to build a small front-end (a single index.html) to browse, filter, and search it. Heads up: Cowork will likely suggest you use Claude Code instead for the front-end. That suggestion is itself the lesson — note when the chat/Cowork interface hands off to a code-first tool, because that’s exactly the boundary Workshop 4 is about.
  5. Let it run at least once (or trigger it manually), then read the output critically. Did it follow the schema? Did it dedup? Are the “relevant” items actually relevant, or did it pad the list? Tighten the instruction and run again.

Post to Discord: a screenshot of the spreadsheet (or the front-end), plus one or two sentences on what you had to add to the instruction to make the output trustworthy.

B. Run a local model for alt-text via the Hugging Face connector (~45–60 min)

Workshop 3 ended on the local-models question: web tools like Claude send your images to someone else’s server; a local model keeps them on your machine. This exercise wires Claude up to Hugging Face so it can find and reason about an open model for the job.

  1. Add the connector. Go to claude.ai/settings/connectors and add “Hugging Face” from the gallery. (For higher rate limits, set an HF_TOKEN — see hf.co/settings/mcp — or create a free account.)
  2. Make a new folder and select it in Cowork. Move a handful of test images into it (reuse your Workshop 3 set if you like).
  3. Confirm the Hugging Face connector is turned on for this conversation.
  4. Prompt Claude to find a local model for the task: “Search Hugging Face for a local (open-weights) model suited to writing alt-text / image captions for these images. Compare two or three candidates — size, license, what they’re trained on — and recommend one I could run on my own machine. Then draft alt-text for each image in this folder.”
  5. Read the recommendation as a research result, not just an answer. What did it suggest, and why? What’s the license? How does its alt-text compare to what Claude itself produced in Workshop 3 — does the smaller, open model flatten more, hallucinate more, or handle your domain’s specifics worse? This comparison is the local-vs-cloud trade-off made concrete.

Set expectations: a local, open model writing alt-text through Cowork won’t give you much detail. Expect short, generic descriptions — “a black and white photograph of a building” rather than the situated, context-aware reading Claude produces in the chat window. That gap is the point: it shows you concretely what you trade away when you keep the images on your own machine.

Post to Discord: the model it recommended (with one line on the license/trade-off it surfaced) and a side-by-side of one image’s alt-text from the local model vs. from Claude.

C. Correct auto-captions with Claude (~45–60 min)

Auto-captions on recorded talks are reliably wrong in the exact places that matter: names, jargon, your discipline’s terms-of-art. Claude can clean them up — with an important limitation to understand.

  1. Gather a video and its captions. A Zoom recording of one of your lectures plus its auto-generated .vtt/.srt caption file is ideal. Put both in a new project/Cowork folder.
  2. Prompt with context and a clear instruction: “This is a recording of [course / topic / who’s speaking]. The caption file was auto-generated and has errors, especially in names and specialized terms. Use the video to make determinations about likely errors in the captions, fix them, and return a corrected caption file with the original timing preserved. Flag anything you’re unsure about rather than guessing. Key terms to watch for: [list names, jargon, titles].”
  3. Know what’s actually happening. Claude won’t review the audio. What it can do is look at screenshots of the slides at given timestamps — so it corrects captions by reading on-screen text and context, not by listening. That’s powerful for slide-heavy talks and weak for discussion-heavy ones.
  4. For long or audio-only recordings, this chat-based approach gets unwieldy. The better path — which you’ll be equipped for after Workshop 4 — is Claude Code, using a model directly for transcription from the audio. Note where the chat tool’s ceiling is.

Post to Discord: a before/after snippet of a few corrected lines (especially a fixed name or term), and a note on where the “reads slides, not audio” limitation showed.

D. Close the loop: image analysis → generation → reading the dataset (~60–90 min)

This is the exercise that completes the Workshop 3 argument. You analyzed a set; now you’ll ask the model to generate into that set, and read the generated image as a picture of the training data behind it.

  1. Start from a related image set in a Claude Project — your own photos, public-domain art by one artist, a run of similar historical documents, a set of comic covers, whatever shares a connection. (Your W5/Workshop 3 set works.)
  2. Run an analysis across the set: shared visual features, period markers, composition, what they have in common.
  3. Ask for a generation prompt: “Based on this set, write me a text-to-image prompt that would generate a new image fitting naturally into this collection.”
  4. Run that one prompt through several generators — Microsoft Copilot, Google Gemini, and ChatGPT Images (use whatever you have access to; two is enough). Don’t tweak the prompt between tools — the point is to hold it constant.
  5. Read the results against the set and against each other. Where each generator “fills in” beyond your prompt — the faces, the bodies, the settings, the defaults it reaches for — tells you what its training data assumes. Connect it back to the Washington Post interactive and “Excavating AI”: what does the divergence tell you about the dataset itself, both yours and theirs?

Post to Discord: the prompt, the images from each generator side by side, and your read on what the differences reveal.

E. Make a slide deck with Claude Design (~60–90 min)

Claude Design generates editable, PowerPoint-style decks from a prompt — a fast way to draft a talk and a good test of how far you can push past the default template.

  1. Open Claude Design and select “New Slide Deck.”
  2. Feed it real content — notes from a lecture, a paper abstract, or a workshop you’ve given.
  3. Prompt: “Build a slide deck for a 15-minute talk based on this content. Make it visually distinctive — not the default corporate template — and match it to [my discipline / the tone of the talk].”
  4. Iterate at least three rounds — color, layout, typography, visual hierarchy. Notice what’s faster than building it by hand, and what’s worse.

Post to Discord: a couple of slides (or the exported deck) and one line on what Design got right and what you had to fight it on.

F. Build a voice interface as an AI Artifact (~2–3 hr, advanced)

You can prompt Claude to build a working voice interface — input and output — entirely inside an Artifact, no code on your part. This is a direct bridge to the tool-building in Workshop 4.

  1. Turn on AI Artifacts in Settings first (the Artifacts capability that lets an Artifact call Claude). Without it, the prompt below won’t produce a working interface.
  2. Open a new conversation and prompt: “Make me an AI artifact that uses a voice interface for [your goal — e.g. asking questions about a primary source in my course, quizzing me on key terms, talking through an idea from my research]. Use the Web Speech API for voice input and output, and show a clear visual indication of when it’s listening versus responding.”
  3. Iterate and talk to it. Tighten the goal, the prompts it asks back, the personality. Compare it to Claude’s built-in voice mode — what does building your own give you that the default doesn’t?

Post to Discord: a screenshot or short screen recording of your voice artifact, plus a note on what you’d change before putting it in front of students.

What to Carry Into Workshop 4

Workshop 4 (June 24) is the first session in Stage 3 — Claude Code Web. To prepare:

If you did Exercise A or C and felt Cowork hand you off toward Claude Code, hold onto that feeling — Workshop 4 is where we cross that line on purpose.

Cross-references