Five fixes I keep reusing, from almost three weeks of arguing politely with an AI news briefing.
6:50 am. My phone buzzes. News briefing.
11:56 am. Buzz again. Another one.
I didn't write a single word of either. I wrote the instructions. Then I spent almost three weeks changing my mind about them.
That second part is the bit worth sharing. Most posts about AI agents show you what the thing can do, usually with a lot of exclamation marks. I'd rather show you what I did when it did something slightly wrong. Not broken. Just not quite what I meant. Because honestly, that's where most of us get stuck.
My day job is program management. When a project disappoints someone, we don't throw it out. We go back to what the requirement actually said, find the gap between what was asked and what was wanted, and fix the spec.
A prompt, it turns out, is a requirements document. And the same discipline works on it.
So here are five fixes. None of them is really about news. You could use them on almost any agent you set up.
Version one: the spec I wrote
My first request was short. Embarrassingly short:
I want 3-4 pieces that are interesting, original and have depth, with real-life scenarios, not the usual marketing fluffs. Keep it not too technical please.
Oh, and it had to land as a phone notification. A briefing I have to go and fetch is a briefing I won't read. I know myself.
Now, if someone handed me that as a project requirement, I'd hand it straight back. Interesting. Original. Depth. Those are adjectives, not requirements. Nobody can test them. Not me, not a delivery team, and not the AI.
And yet it worked well enough that I enjoyed it. Which is the trap, really. It was fine. Then I started noticing what I wanted that it wasn't giving me. Each time I asked for a change, and each change taught me something I could reuse.
The five fixes
Fix 1: Say what it's for, not just what it's about
What I noticed: I was very well briefed on tech, and had absolutely nothing to say about anything else.
What I asked for: a second briefing at midday, plus a few pieces beyond tech, "just to enable me to have a conversation with leaders and peers."
The rule: a topic is not a purpose. Add one sentence that starts with "so that I can..." and suddenly the agent can make better choices. Mine says the briefing should help me hold an informed conversation with senior people outside tech. That one sentence is why every non-tech piece now ends with a question I could actually raise with a leader. I didn't ask for that format directly. The purpose did the work.
Fix 2: Give the unexpected a seat, then fence it in
What I noticed: everything agreed with everything else. Very polite. Very samey.
What I asked for: one piece that's unique or insightful and doesn't follow the usual narrative.
The rule: an agent that summarises the news will, by default, summarise the consensus. That's not a flaw, it's what summarising is. So reserve one slot for the thing you'd otherwise miss, and write the guardrails in the same breath. Mine has to be backed by evidence, avoid conspiracy theories and contrarianism for its own sake, and say plainly when the evidence is one study or one expert's view.
Skip the fence and "against the grain" becomes "against the facts" faster than you'd think.
Fix 3: Turn adjectives into acceptance criteria
What I noticed: headlines and one-liners with nothing behind them. Sort of a news tweet with a title.
What I asked for: "I don't like a one-liner without context, background and any substance."
The rule: "depth" was an adjective, so I swapped it for things I could check:
The agent must open and read the actual article. Never summarise from a search snippet.
Each piece runs 120 to 180 words.
Each piece uses the same labels: In short, What happened, Background, Why it matters.
If a story has no substance, pick another one.
The one line I never touched was Why it matters. I'd already told it I liked that one. Some requirements you just leave alone.
Fix 4: Use quotas, not feelings
What I noticed: I was bored of AI news. (Yes, I see the irony.)
What I asked for: widen the search.
The rule: "less of X" is a feeling. An agent needs a number. So: at most one AI piece per briefing, and zero is fine. Then I named the areas to draw from (economy, health, energy, Asia, workplace trends and a few more) and added a rule against two pieces from the same area.
Notice I capped AI rather than banning it. Some days the AI story really is the biggest story of the day, and I don't want to miss it just because I'd had my fill of it that week.
Fix 5: Run a sample first, and tell each run what the others cover
What I noticed: my first sample leaned on older sources. And my midday briefing was at real risk of just repeating the morning one.
What I changed: I told it to check today's date first and focus on the last day or two. And I told the midday run to favour whatever has developed since the morning, and skip the morning's top stories.
The rule: ask for a sample before you schedule anything. You'd never go live without a test run on a project, so don't do it here either.
And remember this one, because it caught me out: a scheduled agent starts fresh every time. It has no memory of earlier runs. None. If you have more than one, the prompt has to carry that context for them.
The prompt, condensed
The real one is longer (of course it is). This version keeps every rule above and swaps my details for brackets, so you can copy it and make it yours.
You are my personal assistant. I'm a [your role], based in [your city]. Use plain English and define any jargon the first time it appears. Task: use web search to find what is happening right now, then send me my [MORNING] news briefing. Check today's date first and focus on the last day or two. Purpose: help me hold an informed conversation with senior leaders and peers across industries. Depth: open and read each article before summarising it. Never summarise from a search snippet. If a story has no substance, pick another one. Structure: 5 to 6 pieces. A) Tech and science (1 piece). B) Beyond tech (3 to 4 pieces), each from a different area: economy and markets, trade and geopolitics, health, energy and climate, workplace and leadership, education, culture, Australian stories. C) Against the grain (1 piece): challenges the usual story, backed by evidence. No conspiracy theories, outrage bait or contrarianism for its own sake. Say plainly if the evidence is one study or one expert's view. AI limit: at most one AI piece in the whole briefing. Zero is fine. Skip: product launch hype, press-release rewrites, funding news with no substance, gossip, listicles. Format for each piece (120 to 180 words): bold headline; In short; What happened; Background; Why it matters; Conversation starter (for Beyond tech); The usual story and Strength of evidence (for Against the grain); source link.
For a second briefing later in the day, add one line: "I already got a briefing this morning, so favour what has developed since, and don't repeat the morning's top stories."
The checklist to reuse
Before you schedule any agent, check your prompt has:
A purpose. One sentence starting "so that I can..."
Acceptance criteria. Every adjective swapped for something you could check.
A seat for the unexpected, with the guardrails written in.
Quotas. Numbers and named areas, not "more of this, less of that".
A sample run, plus a line on what any other runs cover.
Then change one thing at a time. Change five at once and you'll never know which one helped.
Where this leaves me
The briefing is on its fifth version, and I fully expect a sixth. That's not failure. That's just how any spec behaves once a real person starts using the result. Requirements don't finish. They settle.
But here's the bit I didn't see coming.
In my last post, Network Segmentation: Four Hurdles to Expect, I wrote about distributed sites: hundreds of locations, no local IT staff, and a teller workstation, an ATM, a security camera and the heating, ventilation and air conditioning (HVAC) system all quietly sharing one network. True. Also a bit abstract. You nod, you move on.
Then a few days ago my briefing served up a piece about supermarket fridges.
In October, hackers got into the refrigeration systems of hundreds of supermarkets in France, remotely switching units into defrost mode. Fresh and frozen stock went in the bin, and Paris prosecutors said the attack sparked fires. And it didn't stop at France. In early November, stores in Belgium, Luxembourg, Sweden and the Netherlands reported the same kind of thing.
What a coincidence. Suddenly my distributed sites weren't a line in a blog post any more. They were a freezer full of spoiled food.
But the detail that really stopped me was this one. According to the manufacturer, Danfoss, most of the controllers hit in France were older models it no longer updates. One had been declared end-of-life in 2021, another was phased out back in 2016, and a year ago customers had been urged to plan replacements. Some of the affected units were newer. They just hadn't had their firmware updated.
And that's a risk I overlooked in my own post. I wrote about how the devices at distributed sites are connected.
I didn't write about whether anyone is keeping them up to date. A fridge controller doesn't nag you to restart for updates like your laptop does. It just sits in the back room for a decade, doing its job, until somebody else finds a way to exploit it first. Then suddenly, all eyes are on the updates.
Most of these briefings feel pretty mundane. A trade story here, a health study there. But they quietly widen my lens, and every so often two dots from completely different places join up. That's a real skill, and honestly, it's one program management needs a lot more of.
So here's my question for you. Think of one small thing in your week you'd happily hand to an agent. Now write its requirements without using the word "interesting".
What does it say?
References
CBS News / AFP, "Hackers turn up the heat in refrigerators at hundreds of supermarkets in France, sparking fires and food losses", 12 October 2026.
The Hacker News, "Attackers Heat French Supermarket Fridges Up to 50°C by Hijacking Danfoss Controllers", 13 October 2026.
CBS News, "More grocery stores in Europe report fridge and freezer temperature spikes as apparent cyberattacks continue", 2 November 2026.
