AutoAdy

The AI media buyer engineered to conquer every Meta account l

[KNOX @knoxtwts]

https://t.co/rB9EOgDeNL

[KNOX @knoxtwts]

i found a way to run a full agency-grade marketing audit on any website in 20 minutes...

example below...

agencies charge four figures for this and take two weeks to deliver. most of that time is a junior filling out a template

the actual audit is 6 questions asked in the right order. so i turned the one we run on real clients into 6 claude prompts

it reads your page like a stranger with money, finds the exact line where the message loses them, and hands back the one fix to make first

the full prompt stack is in a free guide, every prompt paste-ready

like + comment "AUDIT" and i'll send over in dm's

(following + rt for priority access)

[KNOX @knoxtwts]

so you're telling me your whole team can now...

> dump the design system, brand voice & seo rules into skill files
> prompt Bolt like normal
> have it auto-pick & stack the right ones on every single build
> get everything out the door to the same standard

without anyone invoking anything?

we're not going back

[KNOX @knoxtwts]

token prices dropped all year and somehow every enterprise ai bill went up. nobody finds this weird?

chamath says token costs are doubling every 45 days for a 5% productivity lift btw…

the glean founder’s take is that everyone’s measuring the wrong number. cost per token means nothing. cost per successful task is the metric. a cheap model that produces rework isnt cheap, you just pay for it twice

[KNOX @knoxtwts]

you ask an AI for deep research, hit go, and come back an hour later hoping it actually finished.

Octen just ran one of these, 129 searches into a full sourced report, in 2 minutes 35 seconds.

and it scored higher than OpenAI, Gemini, Grok and Perplexity on the benchmark, not lower.

we all quietly accepted that good research means waiting.

turns out we were just waiting on slow search.

[KNOX @knoxtwts]

for every "you're right. i made a mistake" from claude or gpt, users should get a full token reset

[KNOX @knoxtwts]

the most interesting job signal i've seen this year is one company recruiting substation electricians and backend engineers on the same careers page.

anthropic committed $50B to compute buildout. fluidstack is the one actually building it, and they are hiring across basically every function to get it done.

situational awareness, blackrock and google all sitting on the same cap table is not a combination that lines up by accident.

if you have been waiting for the part of ai where the work is physical and the deadline is real, this is that.

the interesting infra jobs stopped being at the labs.

[KNOX @knoxtwts]

Ramp pulled the token bills of 1,000+ companies and almost everyone is reading the wrong number off it…

the teams overpaying the worst were barely touching frontier models AT ALL

> inefficient teams: 48% of spend on frontier
> efficient teams: 51%
> the inefficient ones still pay 4.15x more per million tokens

the separator was cache reuse. 28% hit rate against 88%

so every team downgrading to a cheaper model to fix the bill is solving the wrong problem...

the same context gets re-sent uncached on every run either way

what that actually means: you cannot shop your way out of this. the bill only moves when you can see what each individual request cost you

most teams are flying blind on the fastest growing line item they have

[KNOX @knoxtwts]

https://t.co/9fdOsrK3ci

[KNOX @knoxtwts]

i found a way to steal the structure of any winning meta ad...

example below...

the ad library shows every ad your competitors run and how long it's been live. an ad that's been printing for months is a proven structure sitting there for free

everyone screenshots it and copies the words. wrong move. the words belong to their offer, and meta clocks near-copies anyway

the structure is what transfers. so i built a claude workflow that pulls out the skeleton, locks the character across 9 pillars, and writes the prompt that re-shoots the same structure with an ai actor for your offer

in the video: left is an ad running since june, right is the rebuild. nobody wrote a script

the full workflow is in a free guide, every prompt paste-ready

like + comment "SWIPE" and i'll send over in dm's

(following + rt for priority access)

[KNOX @knoxtwts]

your next month of content is sitting in reddit search, pre-validated.

the method: search the subreddits where your buyers live ("site:https://t.co/ql2erMDsbu]") with pain language stapled on: "struggling", "how do you", "anyone else", "frustrated", crossed with your topic words. every hit is a question real people upvoted, which means the demand for the answer is proven before you write a word.

then the rule that keeps it honest: mine the QUESTION, never the answers. the thread gives you the pain in the reader's exact language, your stack gives the fix. answering with the thread's own top comment makes you a plagiarist with a scheduling tool.

one search this morning surfaced the same ui complaint hitting a claude subreddit weekly for months. that's a content calendar someone else wrote for you, sorted by upvotes.

[KNOX @knoxtwts]

if you wanted reach in 2015 you needed a content team & a budget.

by 2020 it was an editor & a posting schedule.

now the machines generate infinite everything & reach is the single thing they cannot print...

own the pipe or rent it forever

[KNOX @knoxtwts]

the system our agency uses to produce ad creatives for paying b2b clients is now written down in a free doc.

it's a claude code build. one intake form takes the icp, pain, offer and cta, and it returns 100+ finished meta statics, each one a different concept with the message designed inside the artwork.

the doc is the full build, not a summary:

> all 20 prompts in build order, each a complete brief you paste as written
> the 80-entry template and visual metaphor library, with the seed entries claude code expands for you
> the fire-and-forget layer, so batches finish with the laptop closed and archive themselves before provider links expire
> a worked example from a real ad copy to a folder your media buyer can upload straight to ads manager

the wall in this video came off one run. the whole build takes about 2 hours.

like + comment "ENGINE" and i'll dm you the doc. (following + rt for priority access)

[KNOX @knoxtwts]

so ai in hospitals now works like this:

> doctor talks to the patient, nothing else
> the notes, referrals and admin draft themselves in the background
> doctor reviews, signs, done
> 2.5 million patient visits a week already run on this

the most regulated industry on earth figured out background agents before most marketing teams did.

damn.

[KNOX @knoxtwts]

https://t.co/5SYgA0dC0S

[KNOX @knoxtwts]

pov: anthropic to the government

[KNOX @knoxtwts]

i default to ignoring anything labeled "legal ai" because it's almost always a gpt wrapper that reads one contract and hands back a summary.

this one reads every contract your company has already signed -> then scores the deal in front of you against your own history.

shows you which clauses your team has quietly caved on before and which ones you always hold.

so the review gets sharper every time you close a deal. vertical ai that actually builds its own memory instead of renting one from openai.

been around this space a while. this is a real one.

[KNOX @knoxtwts]

casinos that are banned from running meta ads just did 15 billion views in a couple of months with the cheapest distribution method

[KNOX @knoxtwts]

the self driving guys spent a decade getting a machine to predict the real world, turns out that was the groundwork for simulating any world.

odyssey just raised 310m to go general with it, and the demos run in real time.

honestly did not expect to see this so soon.

[KNOX @knoxtwts]

prompt engineering, 2022–2026.

cause of death: somebody finally opened the box.

clarity lets you set the concepts driving an output instead of guessing at words.

rest in peace to the magic phrases.