Spinzu

AI builder | testing new models & agents so you don't have to

[Spinzu @Spinzu]

A PROMPT LIBRARY IS ONLY AS GOOD AS WHAT YOU STEAL FROM IT.

been making motion pieces with claude all week. the prompt was never the hard part.

what actually made them work:
→ first frame already full. no slow intro
→ the "wow" moment before second 3
→ one idea per video, not five
→ short. 6–10 seconds beats 30

so don't copy the prompt from the top 10.
open it, watch the first 2 seconds, and ask why it stopped the scroll.

the structure is free. the idea has to be yours.

[Movez @0xMovez]

this is impressive…

i made Pinterest for Opus 5.5 motion designers:

• 2,286 real builds - 600+ with exact Opus 5.5 prompt
• 995 full workflows (prompt, code, or breakdown)
• sorted from most viral to least viral
• motion design, 3D & WebGL, animation, games

Send it to your Claude and never run out of ideas:

https://t.co/U0jaXW4TWe

[Spinzu @Spinzu]

@KurmaRitish rewrite every time. i keep the beats (hook, what changes, the payoff, the length) and throw away the wording. copied phrasing reads like a copy even when the idea is yours

[Spinzu @Spinzu]

A PROMPT LIBRARY IS ONLY AS GOOD AS WHAT YOU STEAL FROM IT.

been making motion pieces with claude all week. the prompt was never the hard part.

what actually made them work:
→ first frame already full. no slow intro
→ the "wow" moment before second 3
→ one idea per video, not five
→ short. 6–10 seconds beats 30

so don't copy the prompt from the top 10.
open it, watch the first 2 seconds, and ask why it stopped the scroll.

the structure is free. the idea has to be yours.

[Movez @0xMovez]

this is impressive…

i made Pinterest for Opus 5.5 motion designers:

• 2,286 real builds - 600+ with exact Opus 5.5 prompt
• 995 full workflows (prompt, code, or breakdown)
• sorted from most viral to least viral
• motion design, 3D & WebGL, animation, games

Send it to your Claude and never run out of ideas:

https://t.co/U0jaXW4TWe

[Spinzu @Spinzu]

@fawnunwrap wild that kimi used 8x fewer tokens and still cost 2x more. the price per token is doing all the damage here

[Spinzu @Spinzu]

@Sprytixl what happens when two night runs touch the same flaky test from different angles? do the PRs pile up, or does the next run see the open one and skip it?

[Spinzu @Spinzu]

THE BEST LINE IN THIS POST IS THE LAST ONE.

"45 minutes buys distribution attempts, not demand."

ignore the $70K. nobody can promise you a number.
the useful part is the boring part:

→ 3 lines no model writes for you: who buys, what you sell, at what price
→ get those wrong and the agents automate the wrong business perfectly
→ every morning read 3 numbers: conversations, replies, paid
→ conversations grow but paid doesn't? the offer is the problem, not the bots

agents can run the hunt.
they can't make anyone want it.

[Noisy @noisyb0y1]

ANTHROPIC LEAKED A FILE THAT BUILDS A BUSINESS IN 45 MINUTES - IT HUNTS CUSTOMERS ON ITS OWN AND BRINGS $70K IN THE FIRST MONTH

in month one the system pulls 140+ customers, in month two the profit number goes past $180K

root → offer → surface → hunt → clock → back into the root

the first 9 minutes you write 3 lines no model can write for you: who buys, what we sell, at what price.

get those wrong and the next 36 minutes automate the wrong business perfectly.

the offer is one page, not a deck: the pain, the objection, the price.

the page, the payment link and the replies get built by 4 files in parallel in 5 minutes.

the first chain listens where your buyer complains - thousands of signals scored cheaply, a handful get a real letter.

a personal letter to someone who was never buying is a paid guess.

the second chain publishes 5 variants a day - one post is a guess, five a day for a month is a distribution.

the third chain is the cheapest: every closed deal describes the next one, and tomorrow the hunt runs a narrower list.

in the morning you read 3 numbers: conversations, replies, paid - if the first grows and the third doesn't, the offer is the problem.

and the honest part: 45 minutes buys distribution attempts, not demand - the machine runs whether or not anyone wants it.

save this and paste it into Claude Code - an hour builds the machine, the market decides the number ↓

[Milo @m1lo_exe]

THIS IS F*CKING INSANE

Someone turned Claude Code into a full quant research team. five agents instead of one chat, each with one job, and one of them exists only to kill your results

Ask a single chat if a trading rule works and it answers in 4 seconds. it never saw the chart. it never ran the test

Here’s what you’re looking at in the video:

→ left column: the whole pipeline. question → scout → backtest → verdict → sweep → skeptic → ledger → paper → runner

→ orange box: Opus 5.5 as the lead. it plans and hands work to the team

→ green box: the actual backtest. 200-day moving average rule on the S&P 500, 1999 to 2026, vs buy and hold

→ blue cards: the three attacks the skeptic runs before you ever see a result

→ bottom: everything gets written to research.md so the team never tests a dead idea twice

the team:
lead (Opus 5.5): turns a question into a plan
scout (Sonnet 5.5): checks the data before anything runs
sweeper (Sonnet 5.5): runs every variation, never just one
skeptic (Opus 5.5): tries to break every finding
librarian (Haiku 5.5): pulls strategies with evidence behind them

the numbers from the published run:
same growth. 6.8% a year vs 6.8%
worst fall -24.4% vs -56.8%
sharpe 0.61 vs 0.44
$10K → $61,526 vs $61,009
42 trades, gross of costs

Same money. less than half the crash

The strongest model goes to the skeptic, not the idea guy. ideas are cheap. killing bad ones is where the money is save and every study ends in one of three fixed verdicts: DATA VALIDATED, TRADE-OFF or NO EDGE. tested on years the rules never saw (after 2015 by default). no agent can invent a sharpe ratio it never calculated

NO EDGE still counts as a result. log it, never retest it without a new reason

this is not a money printer. no costs, no taxes, no slippage in the backtest. paper first, always

full setup with every agent file in the article below ↓

[Spinzu @Spinzu]

@KurmaRitish exactly, reading the structure beats copying the wording. the first two seconds tell you almost everything

[Spinzu @Spinzu]

@DonatelloXPro the 0.85 and 0.95 thresholds are the part i'd want to see more of. how did you pick them, and do you re-check how well jev's scores match reality after a week?

[Spinzu @Spinzu]

@andreysuperior made a smaller 58-part version of the same loop. every part placed only after clearance + clash checks pass

[Spinzu @Spinzu]

NOTHING GETS PLACED UNTIL IT PASSES.

58 parts. every one written as code. placed one at a time.

the loop behind it:
→ list every part with dimensions before writing anything
→ each part is its own function
→ after every part: check clearance, clashes, cable bends
→ one fails, stop. fix it before the next part goes in
→ moving parts last. the fold only gets tested when everything else is green

most people prompt "build me a phone" and get one big guess.
this way you get 58 small checked steps.

works the same for a watch, a drone or a car.

prompt ↓
"Build [device] in code, part by part. List all parts with dimensions first. One function per part, place them one at a time. After each part, check clearance and clashes, stop and fix anything that fails. Animate moving parts last. Show me the parts list before any code."

[Spinzu @Spinzu]

@Brandel0r does the label get marked INFERRED until the test passes, or does the model write it as OBSERVED from the start?

[Spinzu @Spinzu]

THE COOLEST DEMO THIS WEEK IS ALSO THE ONE I'D BE MOST CAREFUL WITH.

an agent that maps every device on your wi-fi and can change their settings is basically a new admin on your network.

the post gets the important parts right:
→ backup before every change
→ passwords in the keychain, never in the chat
→ every change has an undo

what i'd add before trying it:
→ start in read-only: scan and report, change nothing
→ approve one device at a time
→ keep it off anything security-related: cameras, locks, router admin

let it look first. let it touch later.

[codila @0xCodila]

F*ck YES... Claude Opus 5.5 just reverse-engineered my home

one prompt found every device on my Wi-Fi - now one chat runs them all

try with yours, here's step by step guide:

step 1 → copy prompt below and paste it into Claude Code on a computer on your home Wi-Fi

step 2 → It scans your whole network and names everey device it finds, brand and model where it can

step 3 → for each one it tells you what's wrong and what it can fix or improve

step 4 → pick the devices to take over, it backs up each one first

step 5 → passwords go into your keychain with one command, never into the chat

step 6 + the result → say yes and Claude fixes your slow Wi-Fi, cuts ads and bloat on your TV, sets schedules on your AC and much more (every change has an undo)

Change your life today. Then read article below on full reverse-engineering:

[Spinzu @Spinzu]

@AndAIyou exactly, the leak happens in the spec, not the code. i'd add a check step before handoff: scan the spec for function names, addresses or anything that looks like structure. if it's there, room A rewrites it

[Spinzu @Spinzu]

2 AGENTS. 1 FILE. 0 LINES OF COPIED CODE.

that's the whole clean-room method, and almost nobody explains the wall.

clean-room is two rooms, not one:

→ room A reads the app and writes a spec. behavior only: inputs, outputs, edge cases
→ room B builds from that spec. it never sees the original code
→ the spec is the only thing allowed to cross the wall

with agents it maps cleanly:

> agent A: studies the feature, writes SPEC.md
> agent B: fresh session, gets SPEC.md only, builds + tests
> no copied code. no shared context

the rule that makes it work:
if agent B ever sees a line of the original, the room is dirty. start over.

prompt for room A ↓

"Study how [feature] behaves in [app]. Write SPEC.md describing only behavior: inputs, outputs, states and edge cases. Do not include any code, names or structure from the original. Mark anything you're unsure about as UNKNOWN."

full walkthrough in the article below 👇
not legal advice. check the license and your local law before you ship anything.

[Skunzu @skunzu]

https://t.co/9tFdtAZjqj

[Spinzu @Spinzu]

"OPUS GOT DOWNGRADED" — EXCEPT THE OLD RUN IS THE BROKEN ONE.

look at the clips: sep 30 is the empty page. oct 20 is the better result.

but even that proves nothing.
one run vs one run is a coin flip, the same prompt can give different output on the same day.

if you actually want to know if a model changed:
→ run the same prompt 5–10 times on day one
→ save every output
→ run it again later and compare the spread, not one lucky or unlucky shot

one clip is a vibe. ten clips is a test.

[0xMarioNawfal - The Roundtable Show @RoundtableSpace]

The same Opus 5.5 prompt run 20 days apart produced noticeably different results, raising questions about whether the model has been quietly downgraded between runs.

[Spinzu @Spinzu]

@cyrilXBT the skeptic is the real hire here. radar and writer just produce, skeptic is the only one that can say no

[Spinzu @Spinzu]

@Flinzu what do you do when a note fits two topics? split it, or link one from the other?

[Spinzu @Spinzu]

@Sprytixl how many loops does it usually take before it hits 240/240? curious if it ever gets stuck on the same shard

[Spinzu @Spinzu]

91% LOOKS GREAT ON A PROGRESS BAR.
IT ALSO MEANS 21 BROKEN ENDPOINTS IN PRODUCTION.

most agent setups give the swarm a task.
the good ones give it a stop condition.

his example: 240 endpoints, 12 agents.

one-pass swarm:
→ opus splits the job into 12 shards
→ 12 agents run once, in parallel
→ stops at 219/240. 21 still failing. you find out later

swarm + loop:
→ same plan, same 12 agents
→ every shard gets verified, not trusted
→ only the failed shards run again. green work stays frozen
→ each pass is smaller and cheaper than the last
→ it stops at 240/240, not when the agents feel done

the difference is one line in the prompt:

❌ "migrate api/ to v2"
✅ "migrate api/ to v2 until all tests pass. re-run only what failed. show me what changed each loop."

a swarm finishes the task.
a loop finishes the job.

save this for your next migration.

[Sprytix @Sprytixl]

THIS IS ACTUAL F*CKING GOLD

Anthropic just shared one of the best tricks for Claude Code.

The idea: a swarm alone is not enough. You wrap it in a loop and give it a stop condition instead of a task.

Same 240 endpoints, two ways: a one-pass swarm got to 219 and stopped with 21 still failing. The swarm plus loop came back 240 of 240 green in 8 minutes.

claude --swarm 12 --loop "migrate api/ to v2 until 240/240 green"

Opus 5.5 plans once and cuts the job into 12 shards, 12 agents take one module each in parallel, and every shard gets verified instead of trusted

the loop never re-runs green work, only the shards that failed, so each pass is cheaper than the last

91% looks great in a progress bar and still means 21 broken endpoints in production

one prompt from you, 3 loops from it, 0 edits by you

save it, you'll use this on every migration from now on

[Spinzu @Spinzu]

@AndAIyou haven't self-hosted, but curious: is the missing kv cache the hardest part to add yourself, or the load balancing?

[Spinzu @Spinzu]

@mojeskoqq which one fails most, the wrong constants or the skipped assumptions? my guess is assumptions, nobody checks those