Want to access all premium content?
Become a paying subscriber to get access to all courses, guides like this, prompts and skills library, and $500+ worth of AI tools.
Following a topic usually means checking the same sites, opening the same tabs, and sorting new stories from things you've already read. Claude Code cloud sessions can do that on a schedule. Opus 5.5 checks your sources, skips stories it has covered, and saves a brief you can read when you wake up.
We'll use AI news as the example, but you can follow any topic. You'll need a Claude Pro or Max plan, a GitHub account, and a browser. No Terminal required.
Step 1: Set up the desk. Create a private repository and connect GitHub.
Step 2: Get your first brief. Save your sources and research rules.
Step 3: Schedule it. Choose when the brief runs.
Step 4: Test it now. Check the files, links, and repeat stories.
Step 5: Keep it useful. Review the results and watch your usage.
Step 1: Set up the desk
Cloud sessions run on Anthropic's servers, so your laptop can stay closed. They use your plan's allowance.
On github.com, create a repository called research-desk. Set it to Private and select Add a README file. This gives Claude somewhere to save its work.
Open claude.ai/code, connect GitHub, and give the Claude GitHub App access to research-desk.
If the repository doesn't appear, check its access under GitHub's installed apps.
Next, open the environment selector beneath the session's message box. Open the settings for Default, set Network access to Full, and save.
Full access lets Claude reach news sites outside the default list. You can use Custom instead if you want to allow only specific source domains.
Step 2: Get your first brief
Start a cloud session with research-desk, the Default environment, and Opus 5.5 if it appears in your model selector.
Paste this prompt and replace the topic:
Set up this repository as my personal research desk.
Topic: AI tools, models, and product launches.
Create three files:
1. CLAUDE.md
Rules: check sources.md, include only developments published in the last 24 hours, and skip stories already covered in seen.md. Choose up to five useful stories. For each, give a headline, two sentences explaining what changed, the publication date, and a source link.
Don't fill the brief with older stories just to reach five. List sources you couldn't access.
2. sources.md
Suggest five reliable sources for this topic, favoring official blogs and changelogs. Give each URL and a short description.
3. seen.md
Record each covered story with its date, headline, and link. Treat different articles about the same announcement as one story.
Run the first brief now. Save it in briefs/ using today's date as the filename, then update seen.md.
If supported, publish a private artifact page called Research Desk that I can open on my phone. If that isn't available, keep the Markdown brief and explain how to open it.
Commit and push the files to main.
Read the first brief and open its links. Check that the dates refer to the developments themselves, rather than an old page updated today.
Review sources.md too. Replace weak sources before scheduling the next run.
If a source fails, check the error. A blocked host may need a network-setting change; a site's own access restrictions may need a different source.
Copy the Research Desk page link if publishing worked. That feature still needs testing in your account.
Step 3: Schedule it
Open claude.ai/code/routines, select New task, and choose manual setup if prompted.
Name it Morning brief and paste:
Run today's research brief using CLAUDE.md and sources.md.
Read seen.md first. Skip previously covered developments, even when another site publishes a new article about them.
Save the brief in briefs/ with today's date. Update seen.md, then commit and push both to main.
If this session can update my existing Research Desk page, put today's brief at the top while keeping the same link:
[PAGE LINK]
Otherwise, save the Markdown brief and report where it is.
List inaccessible sources. Don't guess what they reported.Choose Opus 5.5 if available, then set a daily schedule at 6:07 AM in your timezone. Check the timezone before saving.
Under Advanced settings, select research-desk and the Default environment. Review the Connectors tab and remove anything this task doesn't need. Scheduled runs may use connected tools without stopping for approval.
Create the task and check its next scheduled run.
Step 4: Test it now
Select Run now, then open the session from History.
Check these three things:
Check | What should happen |
Repeat stories | Previously covered developments stay out |
Saved files | The brief and updated seen.md reach main |
Research Desk | The existing page updates, if supported |
An immediate test may find no new stories. That's fine. It should say so rather than repeat the first brief.
Each run starts fresh. The saved files carry its research history forward.
If changes land on a claude/ branch, check which branch the next run reads. Make sure the updated files reach that branch before testing again. A prompt naming main doesn't override repository permissions.
If page updates fail or create duplicate pages, keep using the Markdown briefs until that part works reliably.
Step 5: Keep it useful
A successful run status doesn't tell you whether the research was good. Open a brief once or twice a week and check its sources.
Watch your usage after the first few runs. If the task uses too much, reduce the source list, run it less often, or choose a smaller available model.
Use the task's on/off switch to pause it without rebuilding the setup.
Keep GitHub connected. Anthropic's documentation describes skipped runs after a disconnection and automatic disabling after 72 hours, so check the task after reconnecting.
You can also review cloud sessions from the Claude app's Code tab, where available.
Use cases
Once the daily brief works, add these four jobs as separate tasks. Give each task an explicit schedule in the task settings.
Use case #1: Watch a competitor
Check one company's announcements and pricing each week.
Check [COMPANY]'s official blog, changelog, and pricing page.
Compare them with the previous saved check. List new announcements with dated links. For price changes, show the old and new figures only when both are supported by saved evidence.
Save the report in competitors/ using today's date, then push to main.Before quoting a change, check the live pricing page.
Use case #2: Track tool prices
Keep a dated comparison instead of rebuilding one every time you need it.
Update prices.md for [TOOLS].
Include the plan, currency, monthly price, billing period, source link, and date checked. Separate monthly billing from annual plans advertised at a monthly equivalent.
Mark changes since the previous check. Leave unavailable prices unconfirmed, then push to main.Check two entries yourself after the first run.
Use case #3: Get a Monday recap
Turn the week's briefs into a shorter reading list.
Read the briefs from the last seven days.
Choose up to five important developments, remove duplicates, and explain each one's relevance to [MY WORK] in one sentence.
Include the original source links. Save the recap in recaps/ using today's date, then push to main.Every item should trace back to a saved brief.
Use case #4: Check a release
Set a one-time task for an announcement you're waiting on.
Check whether [PRODUCT] has released [FEATURE].
Use the official blog, changelog, and documentation. Distinguish a public release from a preview or waitlist.
Answer released, not released, or unconfirmed, with a dated source link. Save the result in the repository.Choose a one-time trigger if available, then confirm its date and timezone.
Confirmed limitations
Scheduled cloud tasks are described as a research preview for eligible Pro, Max, Team, and Enterprise accounts. Availability, model choices, and limits may change. Anthropic's documentation lists an hourly minimum interval.
Opus 5.5 can miss stories, misread dates, or summarize a page incorrectly. Read the original source before relying on a claim.
Runs use your account permissions and plan allowance. Keep unnecessary connectors off, and check repository access before allowing unattended changes.
Artifact publishing and updating the same page from scheduled sessions remain unverified in this setup. Saved Markdown briefs give you a fallback.
The original guide was checked on October 22, 2026. Exact menus, model availability, and scheduling behavior still need checking in your account.
Closing
Start with one topic and five sources. Read the first three briefs closely, check the links, and adjust the research rules when something repeats or doesn't belong.
Then add a Monday recap or one competitor check. Keep the tasks that save you reading time, and pause anything you find yourself ignoring.
See you in the next one :)