This week, we will be looking at how classic digital humanities pedagogical tasks (such as the creation of an eportfolio or the annotation and presentation of an image collection) can change with the addition of agentic AI tools. We’ll define what agentic AI currently means and how it works, introduce GitHub and GitHub pages as a free method for students to build web-based projects; and introduce the Claude Code Web interface, comparing it with the local Claude Code Desktop workflow.

By the end of the session, every participant will have a live URL to either a deployed ePortfolio or a deployed, annotated image slideshow, in their voice, in roughly 90 minutes.

NEH Workshop 4 — Wednesday, June 24, 10 AM – noon, CHDR

Streamed and recorded. Open to UCF faculty, graduate students, and the larger arts and humanities community.

Open the slide deck →

What to Bring

Pre-Workshop Reading

Optional:

Session Outline (120 minutes)

  1. What “agentic AI” actually means. Tool use (the model can call functions you give it), planning (the model decides what to do next, in what order, when to stop), and subagents (the model can spawn helper instances of itself). An agent is the loop these compose into. Why this is a meaningful shift from chat-only Claude — and why everything we did in W1–W6 is still the substrate. Agentic behavior is no longer confined to one place: Cowork is also agentic, and tool use is now possible throughout the Claude web interface — but Claude Code is the more powerful agentic harness, which is why it anchors this stage.
  2. GitHub and GitHub Pages, from scratch. Walk through, together: creating an account, what a repository is, what a commit is in plain language, what main means, and what GitHub Pages does (a free public URL pointed at the contents of your repo). Anyone who needs to sign up does it now; everyone leaves this section with a working account and one empty repository.
  3. From Artifact to Code Web. What persists, what changes. Why a real GitHub repository matters even for a one-page site (version history, attribution, sharing). What Claude Code Web can and cannot do compared to Claude Code Desktop — the local app I demo alongside it today and that we tour fully in Workshop 6 (W11).
  4. Live demo: one build per interface. I run one demo in Claude Code Web (the browser) and one in Claude Code Desktop (the version that lives on your own machine), so we can compare the two interfaces and talk through the advantages and challenges of each:
    • CV → ePortfolio, in Claude Code Web: plan, build, and deploy a single-page personal site.
    • Image set → annotated slideshow, in Claude Code Desktop: generate alt text for each image, rename the files to match their contents, and build a slideshow. Both ship to GitHub Pages within the demo. We watch the URLs go live, we iterate visibly, and we name what’s different between the browser (Code Web) and local (Desktop) workflows.
  5. Hands-on: your document or your images. Pick a path. Each participant follows the demo on their own materials, in Claude Code Web. Roving help, pair support encouraged.
  6. Discussion: assignment design. Open Q&A on how AI fits into the assignments and projects you are building. What changes when “build a website” becomes a 90-minute exercise instead of a semester project? What gets added to the assignment to keep it pedagogical?
  7. Wrap. What broke today? What surprised you?
  8. ELC 5 team: creativity and AI (11:30 – noon). The Electronic Literature Collection, Volume 5 team (Electronic Literature Organization) joins us to discuss creativity and AI.

The ELC 5 team continues with a longer session and Q&A from 1 – 2:30 PM — all participants are encouraged to attend. More information: eliterature.github.io/elc5-info.

Core Exercise

Both paths build and deploy agentically in Claude Code Web — plan, iterate, deploy to GitHub Pages, straight from the browser with no local setup. Pick one. In the live demo I run one build in Claude Code Web and one in Claude Code Desktop (the local app) so we can compare the two interfaces and the advantages and challenges of each — the Desktop workflow is the one we tour more fully in Workshop 6.

If the GitHub connector asks for a token: Code Web’s GitHub connection sometimes needs a personal access token (PAT) rather than one-click authorize. If it does: GitHub → Settings → Developer settings → Personal access tokens → Fine-grained tokens, give it read/write on your repository, and paste it when Code Web prompts. We’ll walk through this together if it comes up.

Path A — CV to ePortfolio (deployed)

  1. Create a fresh GitHub repository. From your GitHub home page, click the green New button. Name it descriptively (my-eportfolio, firstname-lastname-cv). Set it to Public. Check Add a README.
  2. Upload your document. From the repo page, click Add file → Upload files and drag the CV / syllabus / description in. Click Commit changes.
  3. Open Claude Code Web. Connect the GitHub repository.
  4. Plan before you build. Prompt: “Enter plan mode. This repository has my [CV / syllabus / project description]. Read it, then propose a plan for a fun-but-professional single-page website appropriate to the work it represents. Don’t build until I confirm the plan.”
  5. Confirm, then build. Once the plan looks right: “Build it using the content from the document. Put the website in a subfolder called /site. Make it deploy-ready for GitHub Pages.”
  6. Iterate. Color, layout, sections. The conversation is the workflow.
  7. Deploy via GitHub Pages. Settings → Pages, source main branch, folder /site (or /(root)). Wait two minutes; the live URL appears.
  8. Save the URL. Bring it to W9.

Path B — Image set to annotated slideshow (deployed)

Adapted from the Image Metadata with Claude Code demo in DHSI 2026.

  1. Create a fresh GitHub repository. Same as above — Public, Add a README.
  2. Upload your images. Add file → Upload files, drag in the 5–15 images you have rights to use, Commit changes.
  3. Open Claude Code Web. Connect the repository.
  4. Plan before you build. Prompt: “Enter plan mode. This repository has a set of images. Plan a workflow to (1) generate descriptive alt text for each image, (2) rename each file to match its contents, and (3) build a single-page slideshow that presents the images with their alt text as captions. Put the slideshow in a /site subfolder and make it deploy-ready for GitHub Pages. Don’t build until I confirm the plan.”
  5. Confirm, then build and iterate. Let Claude do the renames and write the slideshow. Read the alt text — fix anything it got wrong or thin; that judgment is the point. Ask for navigation, captions, a credits line.
  6. Deploy via GitHub Pages. Settings → Pages, source main branch, folder /site (or /(root)). Wait two minutes; the live URL appears.
  7. Save the URL. Bring it to W9.

While you build, watch the comparison. We run one demo in Claude Code Web and one in Claude Code Desktop, comparing the two interfaces and discussing the advantages and challenges of each — same agentic loop (plan → build → deploy) and same Pages target, but Desktop keeps the files on my machine and commits through GitHub Desktop instead of the browser. Notice where the two workflows diverge: setup, where your files live, how you connect to GitHub, what persists between sessions. That contrast is the on-ramp to the CLI tour in Workshop 6.

Pedagogical Note

This is the workshop where some participants will feel the strongest dissonance: I just made a website without writing any code. It is worth naming what did happen. Describing what you want in prose and steering what the model builds is a kind of natural language programming — the prompt is the source, the running site is the output. And it is continuous with distant coding: just as distant reading lets us work with a corpus too large to read line by line, agentic AI lets us work with code without authoring every line — reading it, directing it, and judging it instead. You still made the design decisions, reviewed the output, and committed the changes. That is a literacy, even if it is not the one you grew up calling “programming.” Read Martin’s PMLA essay before the session if you want a frame for the conversation it will provoke.

For your own teaching: this is a high-leverage assignment design, but it is also one where students can use the tool poorly. Scaffold by requiring (a) at least three iteration cycles documented in the conversation history, (b) a reflection on at least one design decision the student overrode, and (c) attribution of AI use in the README.

Cross-references