Wake up to a news briefing your AI agent wrote overnight

Most mornings start the same way: you open five or six tabs, skim a couple of news sites, a subreddit, maybe a newsletter or two, and twenty minutes later you've read a lot and retained almost nothing. The thing you actually wanted — what moved in the two or three areas you follow — is buried in everything you didn't. That whole rounds-making chore is something an AI agent can do while you sleep, and hand you one short briefing instead of ten open tabs.

This is a use of a coding agent most people never reach for. Tools like Codex and Claude Code can search the web, run on a schedule, and write a file — which is all a morning news digest really needs. You write the instructions once, put them on a timer, and from then on a fresh briefing is waiting when you wake up. Here's the whole setup, kept to things a non-developer can follow.

What you're building

The end state: every morning at, say, 7 a.m., an agent searches the web for news in the topics you care about, throws out the noise and the duplicates, and writes a short, dated briefing — five or six items, each a one-line headline and two sentences of context, with a link so you can read more or check the claim. It lands in a file (or your notes app, or your inbox) and you read it over coffee in three minutes.

What you need

• An AI agent that can read the web — Codex or Claude Code both work. This recipe uses Codex's command line; a one-line note at the end covers the Claude Code equivalent.
• A way to run a command on a schedule — your computer already has one: cron on macOS/Linux, Task Scheduler on Windows. No extra service to sign up for.
• Five minutes to write the briefing instructions, plus a week of light tuning afterward.

Step 1 — Write the briefing (this is the whole game)

The quality of your morning digest is almost entirely the quality of the instructions you give. Be specific about four things: which topics, what kind of sources, how long, and what to leave out. Vague in, vague out. Here's a starting prompt you can paste and adapt:

Search the web for news from the last 24 hours on: [your topics — e.g. AI model releases, the semiconductor industry, and my city's local politics]. Pick the 5–6 most important items. For each, write one bold headline line and two sentences of plain context — what happened and why it matters — and include the source link. Drop anything older than a day, drop duplicates and rewrites of the same story, and skip celebrity gossip and clickbait. If there's genuinely nothing important in a topic today, say so in one line instead of padding. Put today's date at the top. Keep the whole thing under 400 words.

Notice what that prompt does: it caps the length (so you actually read it), it tells the agent to admit a slow news day instead of inventing filler, and it demands source links — which matters, because an agent can get things wrong, and a link lets you check anything that sounds off before you repeat it.

Step 2 — Run it once by hand

Before you automate anything, run it once and read the result like a tough editor. Save your prompt to a file — call it news-brief.txt — and run it in the terminal:

codex exec "$(cat news-brief.txt)"

codex exec runs the agent once, non-interactively, and prints the result — no chat window, no follow-ups. Read what comes back. Too long? Tighten the word cap. Covering the wrong things? Name your topics more precisely. Missing links? Say "every item must have a link" more firmly. Spend a few rounds here; every fix you make now is one you get every morning for free.

Step 3 — Put it on a timer

Once the prompt gives you something you'd actually want to read, schedule it. On macOS or Linux, cron is the built-in tool for "run this command at this time every day." Open your schedule with crontab -e and add one line to run at 7 a.m. and save the briefing to a dated file:

0 7 * * * codex exec "$(cat ~/news-brief.txt)" > ~/Briefings/$(date +\%Y-\%m-\%d).md

Reading the line left to right: 0 7 * * * means 7:00 every day; the rest runs your prompt and writes the output into a Briefings folder, one dated Markdown file per day. On Windows, Task Scheduler does the same job through a wizard — "create a basic task," set it daily at 7 a.m., and point it at the same command. Either way, your machine has to be on (or set to wake) at that hour for it to run.

Step 4 — Decide where it should land

A dated file in a folder is the simplest target and the easiest to skim back through later. But the briefing is just text, so it can go almost anywhere you already look in the morning:

• A notes app — point the output at a synced notes folder (Obsidian, or a plain iCloud/Drive folder) and it shows up on your phone.
• Your inbox — pipe the text into a mail command so the briefing arrives as an email; then it's on every device without you opening anything.
• A file you keep open — overwrite the same digest.md every morning and keep it pinned in your editor.

Start with the dated file. Move it somewhere fancier only once you trust the output enough to read it without checking the terminal.

Tuning it over the first week

The first few briefings will be a little off — too long, too broad, missing your niche, over-covering one topic. That's expected; treat the first week as calibration. A few adjustments that tend to help:

• If it repeats the same story across days, add: "don't repeat items you'd have reported in the last three days."
• If it's too credulous, add: "flag anything that's a single unconfirmed report rather than confirmed news."
• If one topic drowns the others, give each a quota: "at most two items per topic."
• If mornings are quiet, that's fine — a three-line "slow day" briefing is a feature, not a failure. The goal is signal, not volume.

A word on trust

An agent's briefing is a starting point, not a wire service. It can misread a source, blend two stories, or quote a number wrong — the same way a hurried human skimming ten tabs would. That's exactly why the prompt insists on a source link for every item: the briefing tells you what to look at, and the link lets you confirm anything before you act on it or repeat it. Read it the way you'd read a sharp friend's text summary — fast, useful, and worth double-checking when it matters.

If you use Claude Code instead

The recipe is identical; only the run command changes. Where this used codex exec "...", Claude Code runs a one-shot task from the command line the same way, and the cron line and folder setup are exactly the same. The instructions you wrote in Step 1 are the portable part — they work whichever agent you point them at.

The shift here is small but real: you stop being the one who does the morning rounds. The repetitive part — visiting the same sites, skimming, discarding, summarizing — is exactly the kind of bounded, repeatable chore an agent handles well on a schedule. Write the brief once, put it on a timer, and spend the twenty minutes you used to lose to tabs on the three things that actually mattered.