26 hacks, one work session

I thought I knew my way around this thing

July 29, 2026

In partnership with

Daily user. Comfortable. Zero curiosity about the parts I wasn't using.

Then I hit a post with 26 hacks in it, and somewhere around the eighth one the author dropped a note saying most people quit reading right about here. I kept going, partly out of spite, and by the end I felt like I'd been driving a race car in first gear for a year.

What sold me wasn't the tips. It was that one of the hacks is the author deleting their own past advice. They used to tell everyone to build .md files and folders. Now they say skip it, because folders leak old outputs into your context and nobody maintains those files past week two. Watching someone publicly retire their own popular advice is rare enough that I trusted the other 25 more.

So instead of listing them in order, I want to walk you through one normal work session and drop each hack where it actually bites.

Win AI Search Without a Big Team

92% of VCs use AI to find companies. 58% of buyers start there, too. If you're not showing up in AI answers, you're invisible before the conversation even starts. Join HubSpot for Startups, Anthropic, and Marketing Against the Grain on July 30 (11 am ET) for a live AEO teardown. Real startup. Real recs. Register and unlock the free Startup Visibility Bundle.

*Ad

Before you type a single word

Three of these are setup, and they're the ones with the highest ceiling.

Get off the browser. The author is blunt about it: the browser is the demo. The desktop app is where a feature called Cowork lives, and Cowork is the thing you're supposed to be using instead of a plain chat box. If you're still typing into a bare text field, that's habit, not a workflow.

Empty your global instructions. Turn every setting off. Leave Connectors on. That's the entire recommended config, which sounds wrong until you remember the context problem below. A stuffed global instruction field is a tax you pay on every single message, whether it's relevant or not.

Learn the difference between a Skill and a Project. This is the line that made the whole system click for me: a skill is a capability you teach once, a project is a place you go back to. Skills travel with you. Projects hold state, remember between chats, and share with your whole team in two clicks.

The best move in the whole list follows from that. Take your about-me context, the paragraph you keep re-pasting into every conversation, and rebuild it as a skill instead of a file. Same information, except now it shows up automatically instead of sitting in a folder you forgot about. And skills fire themselves. You don't prompt one, you type a slash command and it runs.

Stop typing what you could say in 10 seconds.

Wispr Flow turns your voice into clean, professional text inside any app. Emails, Slack, client updates — speak once, send without editing. 4x faster than typing.

*Ad

The first prompt is where most of the gain hides

Give the goal, not the steps. Handing over a rigid checklist drags the output down. Describe the outcome you want and get out of the way. Every step you specify is a step it can't improve.

Add use AskUserQuestion to your prompt. This one I tested the same afternoon. It flips the conversation: instead of answering immediately, it interviews you first. The author's framing is that it prompts itself better than you ever will, and honestly, the questions it asked me were the ones I'd been avoiding.

Use Research mode. It's the plus button in the chat box, and the author says almost nobody knows it's there. Save it for something genuinely hard, not for looking up a date.

Talk instead of typing. Cowork reads 100,000 words in about 15 seconds, then sits there waiting on you at 60 words per minute. You speak at 150. The bottleneck in your workflow is your hands.

The hack where everybody quits reading

Long chats get dumber.

That's it. That's the whole hack, and it's the one that reorganized how I work. The longer a conversation runs, the worse the output gets, and the author points out that Anthropic's own engineers have said this out loud.

Which means the thing you've been doing, typing "no, I meant" and trying to steer it back, is the worst available move. Every correction makes the context messier and the next answer weaker.

The fix is a button most people have never clicked. Go up to an earlier message and hit "Restart the conversation from here." Clean slate from a good point, instead of dragging a swamp behind you.

There's a money version of this too. A token is roughly a word, and every word you've sent and received rides along in each new message. The example in the post is ugly: message 30 costs you about 31 times what message 1 did. Extra turns are the most expensive habit you have, and you're paying that bill in quality as well as cash.

Same root cause as the .md folder advice they deleted. Too much context backfires. It loops, it repeats itself, it gets less creative. Keep the durable stuff in the project and keep the chat clean.

Own AI deployment, grow your career

Making AI actually work day to day is becoming its own job. Hear from three people doing it: Simone Santiago Broad (Yoco), Yelva Espinoza (Zumba Fitness), and Fin's Dave Lynch. They share what the role really looks like, how it came to exist, the skills worth hiring for, and the challenges they're tackling right now. Watch the full conversation on demand.

*Ad

Pick your engine, then leave it alone

Two lines on models, and they're refreshingly specific.

Point Fable 5 at your single hardest problem. Two turns, then switch. Run Opus 5 High for everything else.

Note the word High, not Max. The author is deliberate about that. You rarely need the heaviest setting, and reaching for it by default is the same instinct that makes people leave a chat running for 40 messages.

Nothing it says is true until you check

Four hacks here and they're all the same warning wearing different clothes.

It sounds completely sure when it's wrong. Confidence is a writing style, not an accuracy signal.

It also agrees with you too easily. The author calls it a sycophant, which is harsh and fair. Pleasing you and being right are two separate jobs, and only one of them is being optimized.

So make it audit its own answer before you trust a line of it. And ask for five very different versions instead of one, then pick. That final pick is taste, and taste is the part that isn't being automated.

Which leads to the line I keep repeating: the first draft is yours to fix, not to ship.

One rule with no exceptions, and it belongs right here. Never paste secrets. No logins, no passwords, no keys. Ever.

Three to run tomorrow morning

If 26 feels like a lot, it is. Here's the short version I'd actually start with.

Move to the desktop app so Cowork is within reach. Restart long conversations instead of arguing with them. Make it audit its own answer before you believe anything it produced.

That's a fifteen minute change to your setup and it lands on the same day.

What I keep coming back to

The last hack on the list is the one I can't put down: outsource the thinking, never the understanding.

Almost none of these are prompt tricks. They're about treating this thing like a fast, overconfident collaborator that needs a clean desk and a human with judgment sitting next to it. Tight context. Short conversations. Permanent skepticism. You own the final call.

Here's the one thing worth doing tonight, and it takes ten minutes. Open your last long chat, the one that went sideways and you gave up on. Find the message where it was still good. Restart from there and ask the same question once, cleanly.

I'd bet money the second answer is better than everything that came after your first correction.