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.
A weekly brief takes three kinds of work. Someone has to find the updates, choose which ones matter, and check the draft against the sources. In Claude Code, you can give each job to a separate agent with its own instructions and tools.
This guide builds a researcher, a writer, and a checker for a topic you follow, tests them together, then schedules the brief for Monday morning. You’ll need Claude Code, the Claude desktop app, and a Claude Pro or Max plan. Each run counts toward your plan’s usage limits.
Step 1: Set up the folder. Update Claude Code and create a home for the agents and briefs.
Step 2: Write the house rules. Define the topic, sources, audience, and format.
Step 3: Build the researcher. Create one agent and check the notes it produces.
Step 4: Add the writer and checker. Run all three agents in order and review the first brief.
Step 5: Schedule it weekly. Set up a local task for Monday morning.
Step 1: Set up the folder
Open Terminal and check your Claude Code version.
claude --versionUpdate it before setting up the agents.
claude update
Create a folder for the project, move into it, and launch Claude Code.
mkdir ~/weekly-briefcd ~/weekly-briefclaudeClaude Code asks whether you trust the folder you just created. Choose Yes. The project rules, agent files, and briefs will live there.
Step 5 uses the Claude desktop app. Open Claude, then About Claude, from the app’s menu bar to check the version. This walkthrough uses version 1.1.5368 or later for local scheduled tasks.
Step 2: Write the house rules
CLAUDE.md holds the project instructions Claude Code reads when a session starts. Put the rules for the brief here so the agents have the same topic, source list, and format.
Paste this into Claude Code.
Create a CLAUDE.md in this folder for a weekly brief. Ask me these questions one at a time, then write the file: 1. What topic should the brief cover? 2. Which 5 to 10 sites or accounts count as good sources? 3. Who reads the brief, and what do they need from it? 4. Which words or phrases should never appear? Add these fixed rules: every item links to its source, items must be from the last 7 days, the brief has 5 items and stays under 400 words, and briefs are saved in a folder called briefs.
Answer each question in a sentence or two. For the source list, start with places you already trust, such as official blogs and product changelogs.
Open the file and confirm that it reflects your answers. Anthropic recommends keeping CLAUDE.md under 200 lines so the instructions are easier to follow. To edit it later, run /memory and open the file.
Step 3: Build the researcher
Each agent lives in a Markdown file that specifies its job, model, and tools. Ask Claude Code to create the researcher’s file for you.
Create a project subagent in .claude/agents/ called researcher.
Its job: find 8 to 10 updates on the topic in CLAUDE.md from the last 7 days, using the sources listed there first. For each update, save the headline, the date, one sentence on what happened, and the link.
Save the notes to briefs/[today's date]-research.md. Never write the brief itself.
Give it only these tools: WebSearch, WebFetch, Read, Write. Use the sonnet model.
Open .claude/agents/researcher.md. Check that tools lists the four requested tools and model is set to sonnet.
Since you just created the agents folder, restart Claude Code so it picks up the new agent. Type /exit, then run claude again from ~/weekly-brief.
Call the researcher directly for its first test.
@agent-researcher find this week's updates and save the notes
Open the research file in briefs/ and follow two links. Check that each page supports the note and that its date falls within the last seven days. The researcher may also list sources it couldn’t access.
Step 4: Add the writer and checker
Ask Claude Code to create both agents in the same project.
Create two more project subagents in .claude/agents/. writer: reads the newest research notes in briefs/ and writes the weekly brief following CLAUDE.md. It picks the 5 strongest items, writes 2 to 3 sentences on each with the link, and saves the brief to briefs/[today's date]-brief.md. Tools: Read, Write. Model: sonnet. checker: reads the newest brief and the research notes and runs 5 checks: every item has a link from the research notes, every item is from the last 7 days, the brief has 5 items, it stays under 400 words, and it uses none of the banned words in CLAUDE.md. It replies with a table of the 5 checks, PASS or FAIL for each, and one line on what to fix. It never edits files. Tools: Read, Grep, Glob. Model: opus.
The writer can save the brief; the checker uses Opus and read-only tools, so any edits go back to the writer.
Run the team once from Claude Code.
The researcher saves its notes, the writer turns them into a five-item brief, and the checker reports which requirements passed.
If a check still fails after the revision, read the checker’s fix and open the draft. A recurring date failure, for example, may mean you need to define the seven-day window more clearly in CLAUDE.md.
Step 5: Schedule it weekly
In the Claude desktop app, open the Code tab, select Scheduled in the sidebar, and click New task. Choose Set up manually if prompted.
Field | What to enter |
Name | weekly-brief |
Instructions | Paste the run prompt from Step 4 |
Advanced settings, then General | Turn on Only on this computer |
Local folder | Select ~/weekly-brief |
Trigger | Weekly, Monday, 8:00 AM |
Only on this computer switches the toolbar to Manual permission mode. Leave it there, choose the model, and click Create.
On the task page, click Run now and open the run from History. Review the permission requests as they appear. For tools you want the task to reuse on Mondays, select Always allow when prompted. Confirm that the run finishes and saves the brief to briefs/.
Keep the desktop app open and your Mac awake at the scheduled time. The app’s Keep computer awake option is under Settings, then This computer, then System.
Use cases
Once the weekly brief works, try these three additions.
Use case #1: Watch your competitors
Add a short section on what competitors launched, changed, or published that week.
Create a project subagent called competitor-watch. Each run, it checks these 3 sites for new launches, prices or posts from the last 7 days: [site 1], [site 2], [site 3]. It adds a Competitors section with up to 3 lines to the newest brief, each with a link. Tools: WebSearch, WebFetch, Read, Edit. Model: sonnet.
Also add these 3 sites to CLAUDE.md as competitor sources, and change the checker so its link check skips the Competitors section.
The exception lets the section use links gathered by competitor-watch, which won’t appear in the researcher’s notes. Check those links and dates yourself.
Run the agent on the latest brief.
@agent-competitor-watch add this week's Competitors section to the newest brief
Look for up to three linked updates or a line saying there were none that week.
Use case #2: Turn the brief into posts
Create an agent that drafts three posts from the strongest items.
Create a project subagent called post-drafter. It reads the newest brief and writes 3 short posts in a plain, factual voice that follows the banned words in CLAUDE.md, one for each of the 3 strongest items, each under 280 characters with the link. It saves them to briefs/[today's date]-posts.md. Tools: Read, Write. Model: sonnet.
Run it on the latest brief.
@agent-post-drafter write this week's postsCompare the names, numbers, and dates in each post with the brief before using the drafts.
Use case #3: Give the checker a memory
Ask the checker to keep a record of mistakes it sees more than once.
Edit the checker subagent: add memory: project to its frontmatter, and add this line to its instructions: "After each check, save any mistake the writer made twice to your memory, and check for those mistakes first next time."
Restart Claude Code with /exit, then claude, so the session loads the updated checker. Ask it to save its first note.
@agent-checker check the newest brief again and save one note to your memory about what the writer should watch for next time.
Look in .claude/agent-memory/checker/. MEMORY.md is the index. Open the note beside it to see what the checker recorded and what it plans to look for on the next run.
To include a new agent in Monday’s run, add it to the scheduled task’s instructions in the order you want it to run.
Limitations
Local tasks need the desktop app open and your Mac awake. If the Mac sleeps through a scheduled time, the app runs the task once for the latest missed time when the computer wakes. Closing the lid puts the Mac to sleep even with Keep computer awake enabled.
Some sites block automated visits, require sign-in, or omit publication dates. Check the researcher’s notes for unavailable sources and swap in accessible sources when needed. Each agent makes its own requests, so the team uses more of your plan allowance than a single prompt.
The checker uses the research notes as its reference. A wrong date in those notes can pass the checks. Open every source link before sending the brief.
A run in Manual mode can pause if it needs permission for a new tool. Use the manual test in Step 5 to see which permissions the task requests.
Cloud tasks can run while your Mac is off. They start from a fresh copy of a GitHub repository and can’t access the local folder used in this walkthrough. Anthropic’s docs list cloud tasks as a research preview.
The original guide was checked on November 13, 2026. Features, menus, and plan limits may change..
Closing
Run the team by hand first. Open the source links, review the checker’s table, and fix any project rule that keeps causing problems.
Once the Monday brief is reliable, add the competitor watcher or post drafter if you want more from the same research.
See you in the next one :)