vivi

Testing AI setups, sharing what actually works

[vivi @xbtvivi]

your agent lost its context between sessions, claimed broken code works, and ran three independent tasks one after another

those are three different bugs in three different places, and not one of them lives in your prompt

there are only three layers under an agent. the harness is the environment it works in, the loop is what happens after each model call, and the graph is who runs when. every failure you have had belongs to exactly one of them, and the symptom tells you which.

doesn't matter if you're building agents for code, research or ops, you need this triage

read the symptom, fix the layer:

> state lost between sessions - harness. add .claude/memory/progress.json and stop treating the chat thread as a database

> agent says the patch works, tests say otherwise - loop. terminate on a pytest exit code, never on the model announcing it is done

> three independent tasks ran sequentially - graph. fan them out with asyncio.gather and join them at one sync gate

> the same file read five times and billed five times - harness. sha-256 the file, cache the content, serve the hash on every repeat read

> retries that never stop - loop. an attempt cap plus an escalation rule, or it burns budget looping on confidence

> nothing catches the edge cases - graph. route the finished patch to an adversarial node that writes tests designed to break it before the PR opens

> string parsing, dedup and file filtering done by the model - none of that is reasoning. a 20-line python script does it for zero tokens

the whole harness is 7 files: CLAUDE.md for the rules, settings.json for timeouts and budget caps, two hooks for caching and audit, three memory files for progress, tool cache and a rollback log.

the article closes with a benchmark table - 6.7x faster, 80% cheaper, 98.6% pass rate. it gives no task set, no sample size and no method, so read it as the author's own figures rather than a measurement. the architecture stands without them.

the terminal in the clip is a separate build of mine and everything on it is a visual, not a live feed.

the full code for the hashing cache, the evidence loop, the fan-out and the red-team node is in the 4-minute article below.

[vivi @xbtvivi]

your agent lost its context between sessions, claimed broken code works, and ran three independent tasks one after another

those are three different bugs in three different places, and not one of them lives in your prompt

there are only three layers under an agent. the harness is the environment it works in, the loop is what happens after each model call, and the graph is who runs when. every failure you have had belongs to exactly one of them, and the symptom tells you which.

doesn't matter if you're building agents for code, research or ops, you need this triage

read the symptom, fix the layer:

> state lost between sessions - harness. add .claude/memory/progress.json and stop treating the chat thread as a database

> agent says the patch works, tests say otherwise - loop. terminate on a pytest exit code, never on the model announcing it is done

> three independent tasks ran sequentially - graph. fan them out with asyncio.gather and join them at one sync gate

> the same file read five times and billed five times - harness. sha-256 the file, cache the content, serve the hash on every repeat read

> retries that never stop - loop. an attempt cap plus an escalation rule, or it burns budget looping on confidence

> nothing catches the edge cases - graph. route the finished patch to an adversarial node that writes tests designed to break it before the PR opens

> string parsing, dedup and file filtering done by the model - none of that is reasoning. a 20-line python script does it for zero tokens

the whole harness is 7 files: CLAUDE.md for the rules, settings.json for timeouts and budget caps, two hooks for caching and audit, three memory files for progress, tool cache and a rollback log.

the article closes with a benchmark table - 6.7x faster, 80% cheaper, 98.6% pass rate. it gives no task set, no sample size and no method, so read it as the author's own figures rather than a measurement. the architecture stands without them.

the terminal in the clip is a separate build of mine and everything on it is a visual, not a live feed.

the full code for the hashing cache, the evidence loop, the fan-out and the red-team node is in the 4-minute article below.

[marfin @marfinxx]

https://t.co/aOcpuc8n3v

[vivi @xbtvivi]

in the early 80s Compaq spent about $1 million rebuilding IBM's BIOS without ever reading its code. that is the only reason PC clones exist

the method was called clean room. it is 40 years old, it is still legal, and it now runs from your terminal

the tool is REA, MIT licensed, built by @morluto. first release july 12, and it passed 100,000 github stars about three months later. it plugs your coding agent into Ghidra, IDA, JADX and the rest so the agent decides what to look at next instead of you running each one by hand.

the whole point is one word: evidence. ask a model how an app works and it recalls patterns from a million other programs and sounds confident. with REA every claim has to point at a function it found, a string it pulled or a call it watched happen, and anything unproven gets labelled a guess.

doesn't matter if you're a developer, a security researcher or just tired of apps being black boxes, you need this

the setup:

> install with npx rea-agents@latest setup, pick your agent, read the change plan before you approve it, restart, then confirm with doctor --client claude_code --json

> start on the official Notes example, not on Photoshop. one Export button traced from renderer.js through preload.js to main.js, and you watch exactly where notes.csv gets written

> pick one feature, not one app. "take apart Notion" is not a task. "find out how Notion handles pasted content" is, and that one is already a published case study

> put the evidence rule in the prompt itself: trace the path, show proof for every conclusion, separate confirmed facts from hypotheses, stop if the scope grows

> write SPEC.md before a single line of code - behavior, inputs, outputs, edge cases, nothing copied. this is the clean-room step and it is the part that keeps you on the right side of it

> build from the spec, then test against the original. the DX-Ball sound function passes 3,205 test cases and reproduces all 63 compiled bytes. a Touhou rebuild hit 696 exact matches out of 697 functions in about 12 days

> know the wall. it reads what is on your disk and never the server, stripped C++ stays brutal, and it hands you no IDA or Hopper license

study the original, ship your own. Sega v. Accolade and Sony v. Connectix both went that way. copying the code or unlocking paid features is the other road, and geohot settled at the end of it.

the dashboard in the clip is my own concept of that loop, with the before and after numbers illustrative rather than measured. the engines named on it are real and the firmware target is not.

the full 10 steps, the prompt pack and the honest list of what the tool cannot do are in the 12-minute course below.

[vivi @xbtvivi]

in the early 80s Compaq spent about $1 million rebuilding IBM's BIOS without ever reading its code. that is the only reason PC clones exist

the method was called clean room. it is 40 years old, it is still legal, and it now runs from your terminal

the tool is REA, MIT licensed, built by @morluto. first release july 12, and it passed 100,000 github stars about three months later. it plugs your coding agent into Ghidra, IDA, JADX and the rest so the agent decides what to look at next instead of you running each one by hand.

the whole point is one word: evidence. ask a model how an app works and it recalls patterns from a million other programs and sounds confident. with REA every claim has to point at a function it found, a string it pulled or a call it watched happen, and anything unproven gets labelled a guess.

doesn't matter if you're a developer, a security researcher or just tired of apps being black boxes, you need this

the setup:

> install with npx rea-agents@latest setup, pick your agent, read the change plan before you approve it, restart, then confirm with doctor --client claude_code --json

> start on the official Notes example, not on Photoshop. one Export button traced from renderer.js through preload.js to main.js, and you watch exactly where notes.csv gets written

> pick one feature, not one app. "take apart Notion" is not a task. "find out how Notion handles pasted content" is, and that one is already a published case study

> put the evidence rule in the prompt itself: trace the path, show proof for every conclusion, separate confirmed facts from hypotheses, stop if the scope grows

> write SPEC.md before a single line of code - behavior, inputs, outputs, edge cases, nothing copied. this is the clean-room step and it is the part that keeps you on the right side of it

> build from the spec, then test against the original. the DX-Ball sound function passes 3,205 test cases and reproduces all 63 compiled bytes. a Touhou rebuild hit 696 exact matches out of 697 functions in about 12 days

> know the wall. it reads what is on your disk and never the server, stripped C++ stays brutal, and it hands you no IDA or Hopper license

study the original, ship your own. Sega v. Accolade and Sony v. Connectix both went that way. copying the code or unlocking paid features is the other road, and geohot settled at the end of it.

the dashboard in the clip is my own concept of that loop, with the before and after numbers illustrative rather than measured. the engines named on it are real and the firmware target is not.

the full 10 steps, the prompt pack and the honest list of what the tool cannot do are in the 12-minute course below.

[codila @0xCodila]

REA is most dangerous thing on the internet right now...

It lets your AI agent reverse engineer any app, down to the binary

If you use it correctly, you can rebuild any feature you pay for, for FREE

In this article, I show you how: https://t.co/3P1FNXiXzl

[vivi @xbtvivi]

i stopped rewriting prompts. one markdown file replaced 20 lines of them and it survives every model switch

same model, two setups. one claims success on broken work. the other stops at the failing test and ships only after 3 of 3 checks pass

the problem was never the wording. Karpathy's framing is the one that fixed it for me: the model is the CPU, the context window is the RAM. dump twenty pages into chat and you flood the RAM, and the rules you care about are the first thing it drops.

a harness is everything outside the weights - what the model is allowed to read, which tools it can trigger, what survives a context reset and when it has to stop.

doesn't matter if you're shipping code, research or content, you need this

the setup:

> write HARNESS.md in the project root - boundaries, map, verification order, escalation. in the clip it replaces 20 prompt lines and gets read on every single run

> turn the request into a bounded contract - deliverable, constraints, non-goals, acceptance criteria. without one the model decides when it is finished, and it always decides early

> give it a map, not an encyclopedia. 1 file read, 19 skipped, context still under half full

> externalize memory into CLAUDE.md or .cursorrules so the work survives chat wipes, rate limits and the next model release

> build sensors before autonomy - npm run lint, npm test, a schema check. asking "are you sure?" is just the model grading its own guess

> cap the attempts at 4 and escalate at 2. two failures on the same error means stop, print the log, wait for a human - not try a fifth variation

> dual-encode every rule. tell it never to delete tests, then block rm -rf at the permission level so the operating system refuses it instead of the prompt asking nicely

the two bars in the clip are labelled illustrative for a reason. that is my own dashboard of the idea, not a benchmark, and the only honest number on it is 3 attempts out of 4 allowed.

models update every few months and your prompts drift with them. your map, your tests and your permission boundaries do not. the three layers and the full HARNESS.md template are in the 5-minute article below.

[vivi @xbtvivi]

i stopped rewriting prompts. one markdown file replaced 20 lines of them and it survives every model switch

same model, two setups. one claims success on broken work. the other stops at the failing test and ships only after 3 of 3 checks pass

the problem was never the wording. Karpathy's framing is the one that fixed it for me: the model is the CPU, the context window is the RAM. dump twenty pages into chat and you flood the RAM, and the rules you care about are the first thing it drops.

a harness is everything outside the weights - what the model is allowed to read, which tools it can trigger, what survives a context reset and when it has to stop.

doesn't matter if you're shipping code, research or content, you need this

the setup:

> write HARNESS.md in the project root - boundaries, map, verification order, escalation. in the clip it replaces 20 prompt lines and gets read on every single run

> turn the request into a bounded contract - deliverable, constraints, non-goals, acceptance criteria. without one the model decides when it is finished, and it always decides early

> give it a map, not an encyclopedia. 1 file read, 19 skipped, context still under half full

> externalize memory into CLAUDE.md or .cursorrules so the work survives chat wipes, rate limits and the next model release

> build sensors before autonomy - npm run lint, npm test, a schema check. asking "are you sure?" is just the model grading its own guess

> cap the attempts at 4 and escalate at 2. two failures on the same error means stop, print the log, wait for a human - not try a fifth variation

> dual-encode every rule. tell it never to delete tests, then block rm -rf at the permission level so the operating system refuses it instead of the prompt asking nicely

the two bars in the clip are labelled illustrative for a reason. that is my own dashboard of the idea, not a benchmark, and the only honest number on it is 3 attempts out of 4 allowed.

models update every few months and your prompts drift with them. your map, your tests and your permission boundaries do not. the three layers and the full HARNESS.md template are in the 5-minute article below.

[mirku @mirku21]

https://t.co/PaAzxlyrsH

[vivi @xbtvivi]

everyone starts at the animation. that is why most launch videos look cheap

the order is design, assets, motion, depth, export. five stages, 20 github repos, and the animation is stage three

Opus 5.5 runs the whole thing from a terminal. it shipped on september 22, the announcement hit 28 million views, and it costs 40% less to run than Opus 5.

the composition lives in code, so nothing in it is a decision you make with a mouse. every landing, every hold, every cursor press is a frame number you can change and re-render.

doesn't matter if you're a developer, a designer or someone who just wants one thing to sell, you need this

the build loop:

> 01 design the scene - storyboard in Excalidraw, lay it out in Penpot, lock the headline box first. if your longest headline wraps onto 3 lines, stop and fix the copy, not the motion

> 02 prepare the assets - clean vectors with SVGO, resize screens with sharp, self-host type with Fontsource and icons with Lucide. check dimensions and font loading before a single frame is built

> 03 direct the motion - Motion and GSAP for timelines, Anime.js for SVG, Theatre.js to tune by eye. write timing in seconds first, then convert. a 0.6 second entrance is 18 frames at 30 fps

> 04 build the depth - Three.js, React Three Fiber, Drei, Postprocessing. keep the camera still on the first pass and change one parameter at a time

> 05 render and ship - Remotion composes, FFmpeg encodes, Mediabunny handles browser pipes. npx remotion render Launch out/launch.mp4 --codec=h264

> before polishing anything, render 3 stills at frames 60, 240 and 540 and fix message, crops and readability first. movement cannot rescue a shot nobody can read

the board in the clip is my own workshop map of that stack, not a product screen. the character walking it is mine too.

five stages, one evening, and the next revision costs you a number instead of an invoice. every prompt and a $150 starter package, which the author calls an illustrative price, are in the 8-minute guide below.

[vivi @xbtvivi]

everyone starts at the animation. that is why most launch videos look cheap

the order is design, assets, motion, depth, export. five stages, 20 github repos, and the animation is stage three

Opus 5.5 runs the whole thing from a terminal. it shipped on september 22, the announcement hit 28 million views, and it costs 40% less to run than Opus 5.

the composition lives in code, so nothing in it is a decision you make with a mouse. every landing, every hold, every cursor press is a frame number you can change and re-render.

doesn't matter if you're a developer, a designer or someone who just wants one thing to sell, you need this

the build loop:

> 01 design the scene - storyboard in Excalidraw, lay it out in Penpot, lock the headline box first. if your longest headline wraps onto 3 lines, stop and fix the copy, not the motion

> 02 prepare the assets - clean vectors with SVGO, resize screens with sharp, self-host type with Fontsource and icons with Lucide. check dimensions and font loading before a single frame is built

> 03 direct the motion - Motion and GSAP for timelines, Anime.js for SVG, Theatre.js to tune by eye. write timing in seconds first, then convert. a 0.6 second entrance is 18 frames at 30 fps

> 04 build the depth - Three.js, React Three Fiber, Drei, Postprocessing. keep the camera still on the first pass and change one parameter at a time

> 05 render and ship - Remotion composes, FFmpeg encodes, Mediabunny handles browser pipes. npx remotion render Launch out/launch.mp4 --codec=h264

> before polishing anything, render 3 stills at frames 60, 240 and 540 and fix message, crops and readability first. movement cannot rescue a shot nobody can read

the board in the clip is my own workshop map of that stack, not a product screen. the character walking it is mine too.

five stages, one evening, and the next revision costs you a number instead of an invoice. every prompt and a $150 starter package, which the author calls an illustrative price, are in the 8-minute guide below.

[beamnxw ./ @beamnxw]

Opus 5.5 can run a whole motion design studio

Sell a 20-second launch clip built from real product screens

Copy the motion engineering prompts and delivery process

Build your offer before the next product launch ↓ https://t.co/KjiDrYRDjQ

[vivi @xbtvivi]

AN ON-CHAIN SCOUT THAT WATCHES WALLET CLUSTERS FORM BEFORE A TOKEN EVEN TRENDS - PAPER MODE, 19 SECONDS, 295 WALLETS TRACKED

noctis runs as one agent, patrolling, not trading - 720 scans logged by second 19 of the trace alone

→ wallet graph builds live - 18 linked nodes at second 7, 71 by second 19, clustering around the same mint before the chart even moves
→ entry density heatmap stacks by the minute - peak cluster count jumps from 7 to 31 across the run, that's the actual signal, not the price
→ simulated pnl (paper, 1 sol base) climbs $80.95 → $686.34 → $1,535.74 → $2,508.96 as confidence rises 22% → 84%
→ signal feed tags every token watch, hold, skip, or buy? - never a flat buy, always a question mark until the cluster confirms
→ exit curve tracks 8+ tokens at once, surf score drifting from 2.68x down to 1.52x the longer the trace runs - the opportunity decays while it's still watching

the coverage calendar has real gaps - wed and sat rows are visibly thinner than mon and tue, so even an always-on scout isn't actually always on, and that gap sits right there on its own dashboard, not hidden

most scouts show you one alert and a price. this one shows the wallet graph underneath the alert, so you can tell a real cluster from one buyer pretending to be a trend

everything here is paper, 1 sol base, labeled research only twice on the same screen - built that way on purpose so nobody mistakes a simulated trace for a live position

what's the actual gap in your own tracking - a couple of days you don't check, or a signal you're not watching live at all?

[vivi @xbtvivi]

AN ON-CHAIN SCOUT THAT WATCHES WALLET CLUSTERS FORM BEFORE A TOKEN EVEN TRENDS - PAPER MODE, 19 SECONDS, 295 WALLETS TRACKED

noctis runs as one agent, patrolling, not trading - 720 scans logged by second 19 of the trace alone

→ wallet graph builds live - 18 linked nodes at second 7, 71 by second 19, clustering around the same mint before the chart even moves
→ entry density heatmap stacks by the minute - peak cluster count jumps from 7 to 31 across the run, that's the actual signal, not the price
→ simulated pnl (paper, 1 sol base) climbs $80.95 → $686.34 → $1,535.74 → $2,508.96 as confidence rises 22% → 84%
→ signal feed tags every token watch, hold, skip, or buy? - never a flat buy, always a question mark until the cluster confirms
→ exit curve tracks 8+ tokens at once, surf score drifting from 2.68x down to 1.52x the longer the trace runs - the opportunity decays while it's still watching

the coverage calendar has real gaps - wed and sat rows are visibly thinner than mon and tue, so even an always-on scout isn't actually always on, and that gap sits right there on its own dashboard, not hidden

most scouts show you one alert and a price. this one shows the wallet graph underneath the alert, so you can tell a real cluster from one buyer pretending to be a trend

everything here is paper, 1 sol base, labeled research only twice on the same screen - built that way on purpose so nobody mistakes a simulated trace for a live position

what's the actual gap in your own tracking - a couple of days you don't check, or a signal you're not watching live at all?

[Lyrics @lyricsxbt]

https://t.co/8Asg9jdLCw

[vivi @xbtvivi]

i genuinely don't understand why everyone isn't using this yet

Andrej Karpathy, a co-founder of OpenAI, posted a simple idea that hit 22 million views: stop using AI to write code, use it to build a second brain.

you give Grok Bot one folder on its own cloud computer, feed it your emails, links and photos from your phone, and it reads them, links them and files them into a living wiki of everything you know. it compounds like interest, the more you feed it the sharper it gets.

doesn't matter if you're a developer, a trader or a student, you need this

the full setup:

> create a bot in Grok Bot, name it Brain and add Gmail and Google Calendar in Connect apps. it runs on its own cloud computer, so it keeps working while your laptop is shut

> paste the prompt from chapter 07 of the article below into its chat. one message builds the folder, the routines and the rules

> it reads 20 emails you already sent and writes voice.md, 10 rules of how you write, each with a real line of yours. no more "hope this finds you well" drafts

> label any email Brain and every night at 11 PM it saves the original into raw/ and files it on the page of that person or project

> at 7:30 AM you get one page with what came in, what it drafted and what waits for your yes, ranked by the 5 projects you actually work on

> correct it once and the fix goes into rules.md, which it rereads before every task. every Sunday it tells you what you decided and forgot

keep it to a few runs a day. one hourly routine can eat a week of your plan.

five minutes to set up, and you never start from a blank chat again.

the terminal in the clip is my own visual concept and the market data in it is simulated. same engine, louder feed: one agent ingesting all day on a computer that is not my laptop.

every prompt is in the 7-step course below. or paste the whole article into Brain's chat and it builds the brain for you.

[vivi @xbtvivi]

@Atlas_fomo 100%

this is the future of personal AI

[vivi @xbtvivi]

i genuinely don't understand why everyone isn't using this yet

Andrej Karpathy, a co-founder of OpenAI, posted a simple idea that hit 22 million views: stop using AI to write code, use it to build a second brain.

you give Grok Bot one folder on its own cloud computer, feed it your emails, links and photos from your phone, and it reads them, links them and files them into a living wiki of everything you know. it compounds like interest, the more you feed it the sharper it gets.

doesn't matter if you're a developer, a trader or a student, you need this

the full setup:

> create a bot in Grok Bot, name it Brain and add Gmail and Google Calendar in Connect apps. it runs on its own cloud computer, so it keeps working while your laptop is shut

> paste the prompt from chapter 07 of the article below into its chat. one message builds the folder, the routines and the rules

> it reads 20 emails you already sent and writes voice.md, 10 rules of how you write, each with a real line of yours. no more "hope this finds you well" drafts

> label any email Brain and every night at 11 PM it saves the original into raw/ and files it on the page of that person or project

> at 7:30 AM you get one page with what came in, what it drafted and what waits for your yes, ranked by the 5 projects you actually work on

> correct it once and the fix goes into rules.md, which it rereads before every task. every Sunday it tells you what you decided and forgot

keep it to a few runs a day. one hourly routine can eat a week of your plan.

five minutes to set up, and you never start from a blank chat again.

the terminal in the clip is my own visual concept and the market data in it is simulated. same engine, louder feed: one agent ingesting all day on a computer that is not my laptop.

every prompt is in the 7-step course below. or paste the whole article into Brain's chat and it builds the brain for you.

[Argona @Argona0x]

https://t.co/XkfsgqCUug

[vivi @xbtvivi]

told jev's night desk: 0 rule breaks or it's off the desk. ten trades later, still 0

$250 to start, $391 by the time i stopped the replay - paper account, not real money, keeping that part honest

zero code between the watcher and the desk - two models splitting the job, nobody doing the other one's part

→ haiku 5.5 scores every signal on spread, news, session, volatility, rate - score under 0.55, skip it and log why
→ opus 5.5 only sees what haiku already passed - decides size, stop, target, risk, nothing else
→ 10 trades closed, 67% win rate, 1.9r average win against 1r average loss
→ max drawdown capped at 5.0%, never broke past it across the whole run
→ cpi news hit in 4 minutes - the desk tightened size on its own before the release, nobody told it to
→ 12 agents on dispatch - read, score, decide, size, mirror, log, risk, news, sync, payout, guard, review - most sit idle, only fire when their job comes up

early in the run it skipped more than it traded - signal #214, range fade, logged as skip, spread too wide at 1.9. that's in the same feed as the wins, not hidden somewhere else

most trading bots show you the equity curve and bury the skips. this one logs "rules broken: 0" next to every single entry, skips included

it's a $250 paper account - the point isn't the return, it's that every miss gets logged the same way as every win

which agent in your stack logs its skips as carefully as its wins?

[tekario @tekari0]

https://t.co/Vv7hiBCdFc

[vivi @xbtvivi]

I LOCKED 2 STEPS AND TOLD IT: TOUCH THEM WITHOUT ASKING AND I DELETE THE WHOLE SETUP

the 2 are tag release and post changelog. the other 4 run while i sleep.

1,000 coding tasks cost $800 if opus 5.5 does all of them. about $480 if sonnet 5.5 takes the first 80%.

laptop stays shut.

→ watch the release once - pull main, run checks, bump version, build image, tag, changelog
→ capture it as a named routine - release_patch v1, not a prompt i retype every morning
→ replay it with fresh context per run instead of one chat that grows all week
→ send the first pass to sonnet 5.5 at $2 per 1M read, keep opus 5.5 for the messy 20%
→ let github carry the state - issue, branch, commits, test output, pull request
→ hold tag release and post changelog behind my approval, let the other 4 steps run
→ schedule the whole loop daily at 09:00

remember that part about approval: a held run does not wait politely in the background. it clears 4 steps, stops, and sits there until i look at it. the gate is doing exactly what i told it to and it still reads as broken the first time you see it.

the panel in this clip is the internal view i use to watch it think. the numbers on it are illustrative, not my live bill.

and the honest part from the guide: there is no connect button. openai and anthropic ship nothing that joins a dot to claude code. you build the github bridge yourself, and if your tests are weak you have only automated a bad merge faster.

everyone shows you the chat window. nobody shows you the two steps that are not allowed to run without a human.

$0 extra infra. github already stores the task, the diff, the proof and the history. the only thing you add is which model goes first.

which one would you refuse to hand over - the tag, or the changelog?

full breakdown is in the article below

[vivi @xbtvivi]

I LOCKED 2 STEPS AND TOLD IT: TOUCH THEM WITHOUT ASKING AND I DELETE THE WHOLE SETUP

the 2 are tag release and post changelog. the other 4 run while i sleep.

1,000 coding tasks cost $800 if opus 5.5 does all of them. about $480 if sonnet 5.5 takes the first 80%.

laptop stays shut.

→ watch the release once - pull main, run checks, bump version, build image, tag, changelog
→ capture it as a named routine - release_patch v1, not a prompt i retype every morning
→ replay it with fresh context per run instead of one chat that grows all week
→ send the first pass to sonnet 5.5 at $2 per 1M read, keep opus 5.5 for the messy 20%
→ let github carry the state - issue, branch, commits, test output, pull request
→ hold tag release and post changelog behind my approval, let the other 4 steps run
→ schedule the whole loop daily at 09:00

remember that part about approval: a held run does not wait politely in the background. it clears 4 steps, stops, and sits there until i look at it. the gate is doing exactly what i told it to and it still reads as broken the first time you see it.

the panel in this clip is the internal view i use to watch it think. the numbers on it are illustrative, not my live bill.

and the honest part from the guide: there is no connect button. openai and anthropic ship nothing that joins a dot to claude code. you build the github bridge yourself, and if your tests are weak you have only automated a bad merge faster.

everyone shows you the chat window. nobody shows you the two steps that are not allowed to run without a human.

$0 extra infra. github already stores the task, the diff, the proof and the history. the only thing you add is which model goes first.

which one would you refuse to hand over - the tag, or the changelog?

full breakdown is in the article below

[Sprytix @Sprytixl]

https://t.co/aRj8erlMO9

[SITA @S1TA10]

Clavicular's friend went out with him to the club and spent the whole night watching him pull girls while sitting on his phone 👀💀

"being Clavicular's wingman is just being a spectator with a front row seat "

[vivi @xbtvivi]

@semichenkko the plan was tnt. arthur just followed it like always

[vivi @xbtvivi]

this is f*cking absurd.

opus 5.5 dropped minecraft into red dead redemption 2 and valentine did not survive it.

not a texture pack. rockstar's physics is still doing the work.

same mud, same saloon, same horses. minecraft just runs on top of it.

-> light tnt with flint and steel in the middle of main street
-> the blast throws out pixel smoke and real rockstar ragdolls
-> draw a minecraft bow and arthur morgan goes straight to wanted
-> swing a netherite sword at a rider and the horse goes down with him
-> tie a lead to a horse and fling it over the rooftops of valentine

read the binary -> find the player -> graft one rule set onto another -> test it in the running game

half the town is gone. the wanted notice says animal cruelty.

rockstar built a whole justice system.
it has no line item for tnt.

[vivi @xbtvivi]

I took a minecraft sword into dragonsreach and it is f*cking absurd

not a texture pack. both games are keeping score.

same throne room, same jarl, same guards. minecraft just runs on top of it.

-> throw a creeper and a block of tnt at jarl balgruuf mid-fight
-> swing an iron sword and skyrim says one-handed increased to 25
-> block an arrow with a minecraft shield. "advancement made: not today, thank you"
-> wall off the throne room with oak planks while the guards shoot into the wood
-> pour lava by the doors. the guards walk straight in and catch fire

read the binary -> find the player -> wire two progression systems together -> test it in the running game

1000 bounty added to whiterun. the law of skyrim has no word for creeper.

skyrim levels you up. minecraft hands you the achievement.
neither one knows the other is watching.

[vivi @xbtvivi]

@sorennxbt rockstar spent years on horse physics. one minecraft lead and it's a kite now