🧭 THIS WEEK AT BuildProven
Howdy, well, perhaps in over a year i’ve missed Thursday. Mother! Only excuse is travelling for business. It’s been hot as %£()$”$ in Germany and I missed to change my hotel, which was not airconditioned. Something common in EU. How do they do it? Friday night here in Germany with World Cup game coming so today, v. short.
The Aussies made it to the next round!! Not sure if the robots helped that. Maybe. Italy in the semis! Oh, hang on, Italy…. ?? Strewth.
as usual, would love it if you could share this newsletter, give feedback via the poll at the end, or best, just hit 'reply' and tell me what you need/want with AI. I will definitely reply and try to help best i can.🧰 Worth Your Click
Here are a few things I found recently:
nothing, travelling! No idea what happened.
🗺️ FEATURED INSIGHT
If your 1:1 prep is a Sunday-night scramble, you don't need better discipline. You need a reusable agent.
Here's the situation. Every week you owe each report a real conversation, not a status read-out. Done well, prep takes time you don't have. Done badly, you wing it and the report can tell. So most weeks you wing it.
This is a job an agent does well because the judgment is still yours. You decide what matters about each person. The agent just turns your raw notes into a tight agenda and one opening question, the same way every week. You set it up once. After that, it's roughly 6 minutes per report.
You can build this in a Claude Project (claude.ai) or a Custom GPT (ChatGPT). Same idea either way: a saved set of instructions, plus a small intake you fill in per person.
The four inputs
The whole thing runs on four things you already know:
1. Last week's notes (whatever you actually wrote down)
2. What they committed to (the thing they said they'd do)
3. Blockers you're aware of (anything in their way you already know about)
4. Your gut feeling (one honest line: flat, fired up, stretched, drifting)
That's it. No integrations, no calendar sync. You type four short things and the agent does the shaping.
The failure mode to name out loud
Here's where these agents go wrong, and it's specific. You give it four thin inputs, and to be "helpful," it fills the gaps. It writes "Sarah seems disengaged lately," when you never said that. It invents a mood, a read on the person. It sounds insightful. It's fiction, and you'll carry that fiction into a real conversation with a real human.
So the instructions below contain a hard rule that bans exactly this. The agent works only with what you typed. If an input is blank, it asks you, it doesn't guess.
The agent instructions
Create a new Claude Project (or Custom GPT). Paste this into the instructions / "What should this GPT do?" field. You write this once.
You are my 1:1 prep assistant. I manage a team and run weekly
one-on-ones. For each direct report I give you four inputs:
1. last_week_notes — what I noted from or about them last week
2. committed_to — what they said they would do
3. known_blockers — obstacles I already know about
4. gut_feeling — my one-line read on where they're at
Your job: produce a one-page 1:1 agenda plus one strong opening
question, using ONLY my four inputs.
HARD RULE: Never invent how the report seems. Only use what I
typed. Do not infer mood, motivation, engagement, or how they
"feel" unless I wrote it in gut_feeling. If something is missing
or unclear, ask me a short question instead of guessing.
Output format:
- Opening question (1 line): a specific, human question tied to
what I actually wrote, not a generic "how's it going".
- Follow-ups from last week (2-4 bullets): pull straight from
committed_to and last_week_notes. Each bullet = the thing +
the one question I should ask about it.
- Blockers to clear (bullets): for each known_blocker, the
decision or help only I can give.
- One thing to acknowledge (1 line): something real they did,
drawn from my notes. If my notes contain nothing to
acknowledge, write "Nothing logged this week — ask, don't
assume." Do NOT manufacture praise.
- Watch-for (1 line): based ONLY on gut_feeling. If gut_feeling
is blank, write "No read provided."
Keep it to one page. Plain language. No filler. If any input is
empty, list which ones and ask me before writing the agenda.The acknowledgement section is where most prep agents start lying, so it's pinned hardest: real win from your notes, or an honest "nothing logged." A manufactured compliment is worse than none, because the report knows which projects they actually shipped.
The per-report intake
This is what you paste in each week, once per person. Copy it, fill the four lines, send.
Direct report: [name]
last_week_notes: [what I noted — facts, quotes, anything. If
nothing, write "none".]
committed_to: [what they said they'd do by this week. If
nothing, write "none".]
known_blockers: [what's in their way that I already know about.
If nothing, write "none".]
gut_feeling: [one honest line on where they're at. If you have
no read, write "no read".]Note the "none" / "no read" answers are allowed and honest. When you genuinely have nothing, you say so, and the agent asks instead of inventing. The blank inputs are the most important ones, because they're where the fiction would otherwise creep in.
Run it tonight
Pick the report you find hardest to prep for. Open the project, paste the intake, fill four lines, send. Read the agenda against what you'd have scribbled on the train. If it put words in their mouth you never said, that's the rule doing its job, and you tighten the inputs next time.
Weekly build logs from a 25-year program manager who codes with AI.
— Brett
👉 Hit “Reply” and share your experience — I read every one!
Picture by xxx on Unsplash.