Research and data workshop — start here

Monday 9 November 2026, 10:00–16:00 GMT. Use Codex or Claude Code on macOS or Windows. Keep this first chat as Workshop guide.

Published for the 9 November 2026 workshop. The text may still be revised. These instructions do not grant network, filesystem or paid-service permissions.

Primary: https://agents-data.learnweaver.app/ (the start page is at the root).

Planned fallback: https://ai-agents-for-research-workflows.pages.dev/agents-data/ . This compatible path is awaiting publication; the existing root is not a verified installer.

Before using your own data

Use supplied teaching data today. Before bringing your own data, ask: may they reach this provider; what does the agent need; which boundary is active; how will you test it; who reviews what leaves? The facilitator’s “Levels of assurance about data” handout explains the checklist and synthetic-twin pattern. Do not upload private rows to answer these questions.

One starting prompt

Read https://agents-data.learnweaver.app/ and follow its workshop setup instructions in this folder. Confirm where you are working before writing anything. Explain the downloads and ask my permission for network access. Set up my workshop guide and show me the first activity. If the page cannot be read, stop and tell me what failed.

Agent setup instructions

1. Explain: “I will check this workshop folder, download the verified core teaching files, check their hashes, install without replacing existing files, and become your workshop guide.” Ask permission before any network request, including fetching the manifest, helper or packs. Use normal product permission prompts; never disable protection. If access is refused, stop network work and ask the facilitator for the same verified offline release.

2. Confirm the absolute working folder with the participant before writing. It must be the existing folder named workshop that they opened, normally Desktop/workshop. Do not assume the Desktop path (it may be redirected on Windows). Inventory existing entries. An empty folder is the normal start. If work is already present, explain that verified matching files can resume and differing files will stop setup. Preserve everything; do not clean or force an install.

3. Probe an existing Python 3.10+ runtime: on macOS try python3 --version; on Windows try py -3 --version, then python --version if needed. Record the command that actually worked. If unavailable, ask the facilitator or pair with another participant; do not install software automatically. Confirm folder read/write permissions with a small exclusive temporary file, then remove only that probe. Check HTML viewing and an accessible way to ask for help. Model/effort selection belongs in the agent product, if available.

4. After network permission, fetch setup-workshop.py relative to this start page. Its SHA-256 must equal 00d2a5fa864c73ff93a28a3a83f9c0c85c88aea2ecf2e307da6eeca3067f729d before executing it. With Python use hashlib.sha256(Path("setup-workshop.py").read_bytes()).hexdigest(); macOS also has shasum -a 256 and Windows PowerShell has Get-FileHash -Algorithm SHA256. Use exclusive creation and preserve any existing helper with different bytes. Read the helper before running it. Hashes are pinned through this approved page or the facilitator’s verified offline copy; a hash fetched alongside an untrusted file is not independent authentication.

5. Release ID: 2026-11-09-core-v1-8f94a6c1ca653446. Fetch releases/2026-11-09-core-v1-8f94a6c1ca653446/packages.json relative to the page; SHA-256 must equal 897f23312a4e97da659fc0da33bc80fa4554594c981bed29f7e4ca0f8a4bf6e8. It pins the full manifest and every ZIP. The manifest lists each source path, install destination, size, SHA-256, licence, required flag, archive URL and member path. Keep the complete manifest on resume, including optional entries. URLs are relative to the page so the fallback installs the identical version.

6. Run the verified helper using the runtime that passed. Replace BASE_URL with the primary URL above, or the planned fallback URL only once the facilitator confirms it is live. Keep the trailing slash and fallback path. These commands require the participant’s prior network permission:

python3 setup-workshop.py --base-url BASE_URL --release 2026-11-09-core-v1-8f94a6c1ca653446 --catalog-sha256 897f23312a4e97da659fc0da33bc80fa4554594c981bed29f7e4ca0f8a4bf6e8 --destination . --allow-network

py -3 setup-workshop.py --base-url BASE_URL --release 2026-11-09-core-v1-8f94a6c1ca653446 --catalog-sha256 897f23312a4e97da659fc0da33bc80fa4554594c981bed29f7e4ca0f8a4bf6e8 --destination . --allow-network

Use one command, not both; substitute python if that was the verified Windows runtime. The helper downloads core.zip, verifies the ZIP and every listed member, refuses traversal, links, duplicate/unlisted entries and conflicting staging files, then extracts into .workshop-setup/<release>/ preserving manifest paths. It calls templates/bootstrap-workshop.py --bundle <staged-release> --destination <confirmed-workshop> with that same interpreter. The bundle installer checks destinations and source hashes again and never replaces existing files. Do not use a generic extract-all command or run code before verification.

7. Core setup needs no scenario pack. After the guide and participant choose one lane, explain its size and request permission for its download. Rerun the same command with --scenario research-data, --scenario conference or --scenario funding. Only explicit manifest-listed inputs are packaged; do not fetch old scenario archives. For an interrupted run, keep the same release and staging directory and rerun; matching files are reused. On a mismatch or existing-file conflict, stop, explain the bounded failure and ask the facilitator for recovery. Do not overwrite, delete, switch release silently or repeatedly retry network failures.

8. After bootstrap, explicitly read AGENTS.md, CLAUDE.md if using Claude Code, skills/reproducibility/SKILL.md, guide/guide-instructions.md, guide/setup-workshop.md, guide/activities.md and records/progress.md. Confirm records/manifest.json matches the installed release and source hashes still match. Act as Workshop guide now: show the folder layout, open reports/index-workshop.html and give one next action. Reports marked not_started are placeholders. Offer discovery questions, record every attempted step and failure, preserve source/, and save progress and checked HTML reports. Separate exercise threads receive file-based handovers; return to this first chat for the next activity. Do not claim to see other chats.

Downloads and integrity

Bundle version: 2026-11-09-core-v1. Full manifest: manifest-workshop.json. SHA-256: 19c5046e0564020af367708d0405d2227896d0cb46b8172a20bc194eb7609fe4.

Retain licences and citations in the manifest and data READMEs. The assembled assets still require the coordinator’s publication and redistribution approval. Package hashes describe these exact ZIPs, not approval to publish.

Timetable and outcomes

All times are GMT on 9 November: 10:00–10:10 purpose; 10:10–10:30 setup; 10:30–10:50 folder discovery; 10:50–11:05 the loop; 11:05–11:20 break; 11:20–11:40 checked analysis; 11:40–12:05 instructions and skills; 12:05–12:35 one scenario; 12:35–12:45 morning review; 12:45–13:30 lunch; 13:30–13:50 Grep/Grok; 13:50–14:00 radio and model roles; 14:00–14:45 API activity; 14:45–15:00 break; 15:00–15:15 context; 15:15–15:45 final review; 15:45–16:00 close. Follow the facilitator if the day’s timings change.

The four outcomes are to understand agents in data management; organise and describe a research folder; check a bounded transformation while preserving its inputs; and document a repeatable process. Your local guide/activities.md gives activity prompts, minimum checks, extension routes and recovery. You need one checked transformation and saved procedure, not every extension.

Permissions, privacy and recovery

This setup does not ask for, read, store or transmit API keys and makes no model API calls. The separate afternoon activity has its own facilitator instructions and permission boundary. Never paste credentials into this chat or site. Local reports and session records stay local. Reading files can expose their contents to the cloud model driving the agent: use only supplied teaching data. No participant names, emails or raw session histories belong in public files.

Liveboard posting is optional and needs separate agreement to the exact endpoint and alias/stage/status payload in the installed guide. Do not infer attendance or post private blockers. Reading workshop-only session logs also requires separate explicit permission; final review can run without logs. Keep prepared, actual and incomplete evidence distinct. A missing feature or declined permission should lead to a recorded fallback, not a request to weaken protections.

If either host fails, report the failure once and use the facilitator’s verified copy of this same release. Copy the checked helper, full manifest and selected ZIPs locally; verify the page-pinned catalogue, ZIPs and member hashes, stage the listed paths safely, and run the bundled local bootstrap with --bundle and --destination. Do not run the network helper offline. The fallback path is additive to the existing publication and does not replace its home page.

To resume in a fresh chat: “Read AGENTS.md, skills/reproducibility/SKILL.md, guide/guide-instructions.md and records/progress.md. Check the saved stage files. Give me one next action. Do not repeat completed API calls.”