🧭 THIS WEEK AT BuildProven
Howdy, next couple of months will be quite short editions, with work travel coming up, family vacay and just general summer/don’t want to be in front of a computer type of time….
That said, this little/tiny newsletter is growing very slowly honestly so i’ll have to think a bit if to continue in the long run. As usual, pls reply and tell me what you need, questions, problems etc. All feedback welcome :). Best actually is if you can forward to 500 friends, would be awesome.
btw, big game Friday - US vs Australia, world cup. Come on Aussies! Win this and boasting rights for probably decades. Thankfully / luckily [i assume], I won’t be drinking as much as the Scots in Boston.
Read on a for a bit more advanced way of summarising large documents.
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:
open source GLM-5.2 launched, and looks pretty great - in top 10! Gotta see if this runs on the new macbook! Very strong at coding it seems in benchmarks.
update; it will not run :(. frontier MoE this big needs a 512GB Mac Studio cluster or a GPU server. No laptop runs it!
Fable 5 launched by Anthropic but just as quickly was removed due to security concerns with the US government. I had access to it but only very briefly and insufficient to give a verdict.
🗺️ FEATURED INSIGHT
One of the most obvious and time saving use cases of AI is summarising documents. Huge documents, contracts can be read and summarise in minutes & it’s amazing. You can improve the output though!
How the summary actually lies
Paste a 50-page vendor contract into Claude and ask "summarise this", and you get back something fluent and plausible. The trap is that it reads exactly the same whether it caught the auto-renewal clause or skipped it. Fluency is not coverage.
Three ways a naive summary quietly misleads you:
It drops the exception. The contract is standard, except for one liability carve-out on page 31. A summary optimised for "the gist" treats the carve-out as noise. The gist is the standard part. The money is in the exception.
It flattens hedged language. "The vendor may, at its discretion, adjust pricing" becomes "pricing is fixed for the term" because the model rounded toward the cleaner reading.
It can't tell you what it didn't read carefully. Ask for a summary and the model has no reason to admit "I skimmed the appendix." So it doesn't.
You can't fix any of this by asking for a "better" summary. You fix it by changing what you ask for.
The fix: make it show its cuts
Don't ask for a summary. Ask for three things: the summary, the list of what it deliberately left out and why, and the list of what you should read in full yourself.
The second section is the whole point. A model that has to declare its cuts can't hide them behind fluent prose. You read the cut list in about a minute, and that minute is where you catch the page-31 carve-out it decided wasn't important.
Here's the prompt. Paste it into claude.ai chat, then attach or paste the document underneath it.
You are reading this document on my behalf. I am a director and
I will act on what you tell me, so be honest about the limits of
your read.
Return exactly three sections:
1. SUMMARY
The key points, decisions, numbers, dates, and obligations.
Quote exact figures and clause/section references. Do not round
or smooth over hedged language. If something is conditional or
discretionary, say so in those words.
2. WHAT I LEFT OUT (AND WHY)
List what you chose not to put in the summary, and the reason
for each cut (too minor, repetitive, boilerplate, or you were
not confident you understood it). Flag anything you skimmed
rather than read closely.
3. READ THESE IN FULL YOURSELF
The 3-5 specific pages, clauses, or sections where the cost of
my misreading is highest, or where you are least confident your
summary is complete. Tell me what to look for in each.
If you cannot find something I would expect in a document like
this (renewal terms, liability, pricing changes, dissent in the
minutes), say so explicitly under section 2.A worked example you can sanity-check
Run that prompt on a real vendor agreement and section 2 might come back with something like this:
What I left out (and why): Skipped the boilerplate definitions on pages 2-4 (standard). Did not include the support SLA table on page 18 (relevant but not a decision point). I was not confident about the indemnity language on page 31, so I left it out of the summary rather than guess at it. Read that one yourself.
That last line is the entire value of the method. A normal summary would have either ignored page 31 or, worse, paraphrased it wrong with full confidence. This one hands you the exact page where its read is weakest. You spend your twelve minutes on the three pages that actually carry risk, instead of pretending you read all fifty.
Completely random Claude Code or Codex Prompt
try this in claude code or codex and see what u get.
Build "Constellation" — a single HTML file that turns any text I paste into a living, interactive star map of ideas.
- I paste in text (notes, an article, a brain-dump). It extracts the key concepts and lays them out as glowing stars connected by faint lines where ideas relate — like a constellation.
- Bigger, brighter stars = more important ideas. Drag them around; the connections follow with smooth physics.
- Hover a star to see the idea; click it to gently pull its related ideas closer and dim the rest.
- Beautiful deep-space aesthetic: subtle parallax, soft glow, stars that twinkle. It should feel calm and alive.
- A "shuffle" button re-arranges the layout, and an "export" button saves it as an image.
- One file, no setup. Open it in a browser and it just works.
Build it, run it, screenshot it, and refine the look until it feels genuinely beautiful before calling it done.Weekly build logs from a 25-year program manager who codes with AI.
— Brett
👉 Hit “Reply” and share your experience — I read every one!