🧭 THIS WEEK AT BuildProven
Howdy, bit rushed, so short one this week -
something AI is very very good at - digesting and summarising information
bit of stream of thought summary of things i have learned since using AI
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
Introducing Grok Bot — Grok Bot is your team of always-on agents. They have their own computer, work inside tools and apps like you do, and keep working a reported figure/a reported figure. Pity it’s in some stupid $200/mth plan.
crazy hype on X. It’s like openclaw but apparently easier. Thing is though these things use agents… and using openclaw, setting up agents to do what you want and to not fail and not spend big API $$ is really hard - I’m still learning that. Think I need to learn and write about getting agents right.
also Grok 4.6 - apparently very good but i don’t use the tools it’s available with [cursor i think], so won’t use it.
GenOffice - Free, opensource AI office tools - probably would like it if i liked / used office stuff at home.
Claude Code common workflows - v. useful, definitely bookmark this!
🗺️ FEATURED INSIGHT
200 meeting notes in, one action-item table out
Last week you pointed a bulk workflow at a folder of holiday photos, because nothing bad happens if it mislabels a beach. This week, point the same skill at the folder that eats your Friday afternoons: your meeting notes.
The job, stated as an outcome
You are not writing a parser. You are telling a coding agent what you want and where it is allowed to look. In plain English, to Claude Code (Codex runs the same saved workflow):
Use this exact prompt:
Read every meeting note in this one folder.
Return a single table with one row per action item: owner, what was committed, the date it was committed, the due date if stated, and status.
Only use text that is actually in the notes. If an owner or date is not stated, write UNSTATED rather than guessing.
Do not edit or move any note.
Show me the first 10 rows before writing the file, and wait for me to approve.Five things are doing the work there, and none of them are clever prompting. One folder, so the blast radius is known. One output file, so nothing gets overwritten. UNSTATED instead of a guess, so the gaps are visible instead of invented. A preview before the write. And an approval point that is you.
The agent handles the file reading and the writing. That is its job, not yours. Save the whole instruction as a reusable workflow so next quarter is one sentence, not a rebuild.
For a recurring review, schedule that workflow weekly against the same approved folder. The agent produces the table; the named operations owner reviews the preview and approves the output before anyone acts on it.
Try it yourself
You can try this on 10 notes from one meeting week in about 15 minutes. Make a copy of that folder, use the exact prompt, compare the table with the commitments you already know, then record the caught and missed rows before you point it at a quarter of work.
Some Random Thoughts on AI
I’ve been using AI a few years now and coding/building for about 1 yr. Some thoughts & lessons -
General
Give it context - in the prompt or preferably in project instructions, tell it what you are working on and why. This will massively improve output.
Be specific - more specific = better output.
Group things - use projects or folders or both [whatever the UI] to group different sets of work / projects.
Experiment! - use different AI providers, and compare results - see what seems to work best for you.
Make AI compete! - feed output from ie Claude into ChatGPT and say ‘Claude said this, what do you think!?’. Interesting getting different view points and outputs from different AI models. One might show up as best or maybe the best is a combination.
Some best use cases
summarising large sets of information or documents
brainstorming, ideas, planning …
research
one thing missing is images and video - I do not yet use AI for that much - my attempt at images in these articles is not very good :(.
Building
be specific and follow basics:
i want xyz - write a requirements doc for that
review that doc, think and create the best design for that
[switch to lower model] - build that design! including creating tests that automatically run as part of our release process
or better, build part 1 of that design —> then test it
repeat
[if u don’t know - ask!]
agents are really hard - i need to learn these more.
the last 10% of building something is hard, so try not like i do have 10+ projects open at 90% finished. :(.
Get the Field Guide to AI: three practical builds with copy-paste prompts — https://buildproven.ai/subscribe
Weekly build logs from a 25-year program manager who codes with AI.
— Brett
👉 Hit “Reply” and share your experience — I read every one!