By Sunday night you can have a working Claude Code agent that reads a folder of articles and writes a 700-word brief with numbered citations. No code, no separate no-code builder, no second subscription.
You need three things before Saturday: a terminal, a paid Claude plan, and six public articles saved as text files. That is the whole shopping list, and the second item is the one most people get wrong.
This is not a browser tab that builds apps for you. Claude Code runs in a terminal window and edits real files. That sounds worse than it is, and it is the reason the result keeps working next week.
The schedule below totals 204 minutes across two days. Most of it is you, not the AI.
What is safe to put in./sources
- 🌐 Public articles and blog posts — Anything anyone can read on the open web belongs here.
- 📊 Published reports and open datasets — Fine, provided the file itself is public and under 1,500 words.
- ✂️ Redacted client documents — Strip names, numbers and anything identifying, or swap in a public sample instead.
- 💬 Your own draft notes — Safe only after you remove anything you would not send to a vendor.
- 🔑 Passwords and API keys — Never. The agent reads every file in the folder.
- 🏥 Medical or financial records — Never, not even a small excerpt with the name removed.
- 📁 Client files under contract — Never without written permission from the client first.
The working agent you finish
By Sunday evening you have one folder called weekend-agent holding one new file, brief.md, and a project subagent named brief-writer that produced it.
The agent reads every markdown file in the sources subfolder, checks the claims in them against the web, and saves a 700-word brief with 8-12 numbered citations.
Anyone who can fill a folder and paste text can run this. What it asks for is a terminal – the plain text window that comes with your computer – plus a paid Claude account and a set of text files.
A browser tab alone will not do it.
The input side is six public articles saved as markdown, each under 1,500 words. The output side is the single brief.md, and that is the only file the agent is allowed to touch. That pairing is what makes the folder safe to reuse.
Book roughly 204 minutes across the weekend. That splits into 79 minutes of clicking and typing, 20 minutes while Claude Code works, and 105 minutes reading its output and fixing sources. The review is the largest slice, and it is yours.
Leave the folder in place on Monday. Next Friday you swap the six source notes, ask Claude Code to run brief-writer, and the same setup writes the next brief.
| Phase | Best tool | Hands-on min | Claude Code min | Review/edit min |
|---|---|---|---|---|
| Install and sign in | Claude Code install page | 20 | 0 | 0 |
| Write the brief prompt | Claude Code | 15 | 1 | 15 |
| Generate the agent files | Claude Code | 5 | 2 | 20 |
| Add six sources | Finder or Explorer | 15 | 0 | 5 |
| First one-source test | Claude Code | 2 | 3 | 15 |
| Fix citations | Claude Code | 10 | 4 | 20 |
| Full six-source run | Claude Code | 2 | 8 | 20 |
| Permission and privacy pass | Claude Code settings | 10 | 2 | 10 |
What Claude Code costs you each month
You cannot run Claude Code on the free Claude tier, and there is no separate Claude Code bill to pay. Access rides on a paid Claude subscription or on metered billing through an API key.
An API is the connection that lets software talk to Claude directly. An API key is the long password that authorizes that connection and pays for each use.
Claude Pro opens it for $20 per month, or $17 per month if you pay for a year up front. That is the amount to budget.
Max 5x costs $100 per month and Max 20x costs $200, and both plans include Claude Code. All figures are as of November 2026.
Company seats split in a way worth knowing. A Team Premium seat carries Claude Code, and a Team Standard seat leaves it out, so check the seat type before you assume.
One free path exists: point Claude Code at a model running on your own machine through Ollama. The catch is that Ollama runs open models, so Claude itself is gone, and the brief quality changes with it.
For this weekend, spend the $20 on Pro and keep everything on a single plan. It is the route Anthropic supports, and sign-in takes a minute.
The words in that prompt, decoded
The install and sign-in check
Get Claude Code from Anthropic's own install page. Pick the instructions for your operating system, Windows or Mac, and run what that page gives you.
Copies from blog posts or package registries are not the supported source, and a bad build shows up later as strange errors.
The install itself takes about 20 minutes on a home connection. When it finishes, sign in with the account that holds your Pro, Max or Team Premium plan, or the key set up for API billing.
A session tied to a different account will not find your plan.
Then type one small request, such as asking what day of the week a date falls on, and wait for a plain answer. An answer means the session is live and paid for.
A refusal is the useful failure at this point. It means the login or the billing is wrong, and you fix that now rather than after building the project folder.
⚠️ COMMON MISTAKE
Do not open a second account
Claude Code shares one usage pool with the Claude apps, so coding work draws down the same rolling five-hour limit and weekly limit as your chats. If a session stops, wait for the window to reset or turn on usage credits. Never sign up for a second account to get around the limit.
The weekend-agent folder and six sources
Make a folder called weekend-agent, and inside it one subfolder called sources. Put nothing else in the folder yet. No scripts, no theme notes, no old drafts.
Every working file goes in sources, and the only thing that will ever come out sits at the top level. The agent reads one place and writes one place.
Save six public articles into sources as .md files, each under 1,500 words. Name them so you can tell them apart at a glance, like pension-fees-2026.md.
Keep the plain text of each article, not a screenshot or a PDF link, because Claude Code reads text files fastest. The six originals stay exactly as saved; brief.md never overwrites them.
The output is brief.md at the top of weekend-agent, sitting beside sources, not inside it. That split is what lets you check the agent later: six inputs on one side, one file on the other. If anything else in the folder changes, something went wrong.
Public articles only. This set is safe because every file is already published and none of it belongs to anyone. The folder now holds six clean inputs and one named output, and the prompt has exact paths to work with.
The exact prompt you paste
Paste one prompt into a Claude Code session opened in weekend-agent, and it writes the whole agent for you. That prompt is the job description and the rulebook at once.
It names the input folder, the output file, the web checks, the allowed tools, the write limit, and the persistent rules in CLAUDE.md. Everything the agent does later comes from these lines.
Prompt to copy
Create a project subagent called brief-writer in this folder. It reads every .md file in ./sources, extracts claims, checks each claim against the web, and writes ./brief.md as a 700-word brief with 8-12 numbered citations. It may read files and search the web, but it may only write brief.md. Put its persistent rules in CLAUDE.md and make it ask me before running any shell command. Then show me the files you created.
Send it as one block, in the folder itself. Do not trim the 700-word target or the 8-12 citation count, and do not soften the only-write-brief.md rule. Those three numbers and restrictions are what make the agent predictable a week from now.
The payoff is a session that stops asking what you want and starts building. Claude Code reads the paths, writes the rules, and comes back with files rather than a plan.
The files Claude Code should generate
A chat reply is not the build; the build is files you can open. Expect three things back: CLAUDE.md, a project subagent file for brief-writer, and a short README or run note telling you how to call it.
If you only get a plan in the chat, tell Claude Code to create the files now, in this folder.
Open CLAUDE.md first. Look for four lines of substance: the folder it works in, the tools it may use, the rule that it may write only brief.md, and the rule that it asks you before running any shell command.
Each one is a limit, and each limit is what keeps the agent on one job.
Then check the agent's own file. It should show that brief-writer can read files and search the web, but cannot write anywhere except ./brief.md. If the write limit is missing, add it before you run anything.
Once the write limit holds, Claude add-ons worth installing extend what your agent can reach without widening that boundary.
You now hold a project you can read line by line. That inspection is the whole point: you spent five minutes checking files instead of finding out the hard way during a full run.
The one-source test before the full run
Run the first real test on one source, not six, so a citation mistake costs you 15 minutes instead of a weekend. Generated files can look correct and still fail on evidence. A full run buries four bad citations under 100 good sentences.
Move five of the .md files out of ./sources and leave one in place. In the project folder, type "Use brief-writer" and let Claude Code work.
Expect about 2 minutes of hands-on time, 3 minutes of Claude Code time and 15 minutes of review and edit.
Open ./brief.md and confirm it exists. Then check the folder: the source file and CLAUDE.md should be untouched, and nothing new written besides the brief.
Read each sentence beside the source and mark two things: claims the article does not support, and citation numbers that go nowhere. Time check: 2 minutes hands-on, 3 minutes Claude Code, 15 minutes review.
That one-file run catches a bad rule cheaply. Fix the small failure now, and the six-source run starts from something that already works.
What the folder should look like when the test passes
Correct
- brief.md created at the top level of weekend-agent
- Six source files byte-for-byte unchanged
- Every citation number opens a real page
- CLAUDE.md still lists only brief.md as writable
Wrong
- A new draft.md or notes.txt appears beside the brief
- A source file was tidied or reformatted
- Citation 7 points to a page that does not exist
- The agent ran a shell command without asking
The citation fixes that earn trust
The first output usually carries invented sources, links that do not open, quotes the source never says, and factual sentences with no citation number at all. Polished prose hides all four, which is why the brief reads better than it checks out.
Add one line to CLAUDE.md: every factual sentence needs a source line and a quote or page reference, and the agent may cite only links that open. This turns trust from a feeling into a rule the agent can follow.
Rerun the one-file test and set the new brief beside the old one. Where the new version cites a link you can open and quotes the source word for word, the rule worked.
Keep it. Where it did not, tighten the wording once and run again.
The result is a brief you can edit instead of a confident draft you must fact-check from nothing. For 10 minutes of hands-on work, 4 minutes of Claude Code time and 20 minutes of review, that is the difference between a draft and a document.
The full six-source run
With the citation rule in place and the one-file test passing, put the five sources back so ./sources holds all six .md files. Then run the agent again in the same folder.
Type "Use brief-writer" and expect about 2 minutes of hands-on time, 8 minutes of Claude Code time and 20 minutes of review and edit. The full run gives the agent six times the material, so give its review the same weight.
Check three things when it stops. All six sources should appear in the brief, the numbered citations should land between 8 and 12, and ./brief.md should be the only file that changed since you started.
You finish with a complete 700-word draft that draws on every note in the folder. From here the work is ordinary editing: cut, reorder, and keep the citations that hold up.
🗓️ Your weekend, in order
Install and sign in
Install Claude Code from Anthropic's page and confirm a prompt answers on your paid plan.
Build the folder
Create weekend-agent with a sources subfolder holding six public markdown articles.
Paste the contract
Paste the full brief-writer prompt and read every generated file, starting with CLAUDE.md.
Test on one source
Run brief-writer on a single file and check the brief and the folder for damage.
Repair the citations
Add the source-line and live-link rules to CLAUDE.md, then rerun the one-file test.
Run all six
Restore all six sources and produce the full 700-word brief with 8 to 12 citations.
Lock it down
Confirm only brief.md is writable and no private material sits in the folder.
The permission and privacy pass
Three boundaries keep the finished brief from becoming a way to read or change anything else.
The write rule stays where it is, in CLAUDE.md, so the agent can only create or overwrite ./brief.md.
The ask-before-shell rule stays beside it, so any command runs only after you agree.
The read side is on you. Nothing private goes into ./sources: no passwords, no client files, no medical or financial records, no private chats. The folder you tested is public articles, and that is what makes an agent reading every .md file in it safe.
When a job needs real material, cut it down first or swap it for a public sample that carries the same structure. Then open CLAUDE.md and read the allowed-tools line before each new source set, not once for the whole weekend.
Boundary check: one output path, one input folder, no tool beyond reading files and searching the web. Keep those three intact and the folder stays reusable on work you would send to a vendor.
The reusable weekly brief agent
Next week's brief comes from the same folder, with the six old source files swapped for six new ones and CLAUDE.md left untouched unless a rule has to change. Weekend-agent becomes your template, and building it again is not part of the job.
Claude tricks power users use turn that standing setup into shortcuts you stop thinking about.
Before the run, move the old brief.md into an archive folder with the date in the filename, so the new one cannot overwrite it. Then ask Claude Code to use brief-writer and let it work through ./sources.
The catch is your plan's limits. Claude Code draws on the same 5-hour and weekly limits as your Claude chats. If a Pro session stops mid-run, wait out the 5-hour window or turn on usage credits in billing. A second account does not fix it.
Done that way, a weekend's setup turns into a weekly habit: drop in six notes, run one agent, check one file.
The go/no-go check before you call it done
It is working when three checks pass on a fresh session, not when the last brief looked good on your screen. First, open ./brief.md: about 700 words, 8 to 12 numbered citations, every citation link opens, and all six sources are quoted or named.
Second, look at the folder. Only brief.md has changed. The six source files and CLAUDE.md are exactly as you left them, with no stray drafts beside them.
Third, close Claude Code, reopen weekend-agent, and ask it to use brief-writer once more. A rerun that produces a brief without your help proves the agent lives in the files, not in one lucky session.
The verdict rule: a failed check stops new sources and new features until that one check passes. Fix it, rerun, recheck. When all three pass, stop building and start using the agent.
Frequently Asked Questions
Do I need to know how to code to build this?
No. You type one prompt into Claude Code and read the files it creates. The skill you actually need is careful reading, because your review time is 105 of the 204 minutes in this build.
Can I run Claude Code on a free Claude account?
No. Claude Code has no separate price but will not start a session without a paid Claude plan or an API key. The only zero-dollar route runs it against a local model through Ollama, which drops Claude models entirely.
Will this use up my Claude chat limits?
Yes. Claude Code shares one usage pool with the Claude apps, so coding sessions draw down the same rolling five-hour limit and weekly limit as your conversations.
Why test on one source before running all six?
One file makes citation errors and any overreach in write permissions obvious and cheap to fix. A six-file run buries the same mistakes in a longer output and costs you a full review cycle to find them.
Does the agent need my private documents to be useful?
No. Six public articles are enough to build and test it. If real material is needed later, redact it first or substitute a public sample of the same shape before it goes in ./sources.
How do I run it again next week without rebuilding?
Swap the six source files and ask Claude Code to use brief-writer. Rename the old brief.md with a date before the run so you keep a record of every version.
Three checks, then stop building
Run the verdict cleanly. If ./brief.md is the only file that changed, if it carries 8 to 12 numbered citations that all open, and if brief-writer reruns after a fresh session, the build is finished. Stop there.
If one check fails, fix that check before adding sources or features. A bigger agent with a broken write rule is worse than a small one you trust.
That is a full weekend of setup for a brief agent you reuse all year. The folder does not expire, and neither does the prompt.