Загрузка видео...

Не удалось загрузить видео

Возникла проблема при загрузке этого видео. Это может быть связано с временными проблемами сети или видео может быть недоступно.

На главную

You're still typing every turn The guy who built the tool wrote the loop that types them for him "I don't prompt Claude anymore I write loops that prompt Claude" A loop takes four shapes turn-based, you close it each round goal-based - all tests pass, stop after 5... show more

Skaly_Bull

1,162 subscribers

12,525 просмотров • 9 дней назад •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

Комментарии: 0

Нет доступных комментариев

Здесь появятся комментарии из оригинального поста

Похожие видео

there are four types of agent loops. most people only know one. loop engineering is a choice between four structures, each handing off one more job than the last. every one answers two questions: what starts a run, and what ends it. hand-run, you answer both yourself, every time. 1) turn-based → you prompt, it acts, you review, you prompt again. both jobs stay with you. use when requirements are still forming. 2) goal-based → "/goal hit Lighthouse 90, stop after 5 tries." an evaluator checks, a no sends it back. use when the outcome is measurable but the path isn't. 3) time-based → a clock fires, it runs "check the PR, fix CI," then waits. /loop local, /schedule survives a closed laptop. use for recurring work. 4) proactive → no human present. it watches a channel, spawns triage, fix, and a reviewer, closes the task itself. use for standing duties you can't predict. not which one is most advanced. whether your task is exploratory, measurable, recurring, or standing. the more you hand off, the less you babysit. full breakdown in the article below.

there are four types of agent loops. most people only know one. loop engineering is a choice between four structures, each handing off one more job than the last. every one answers two questions: what starts a run, and what ends it. hand-run, you answer both yourself, every time. 1) turn-based → you prompt, it acts, you review, you prompt again. both jobs stay with you. use when requirements are still forming. 2) goal-based → "/goal hit Lighthouse 90, stop after 5 tries." an evaluator checks, a no sends it back. use when the outcome is measurable but the path isn't. 3) time-based → a clock fires, it runs "check the PR, fix CI," then waits. /loop local, /schedule survives a closed laptop. use for recurring work. 4) proactive → no human present. it watches a channel, spawns triage, fix, and a reviewer, closes the task itself. use for standing duties you can't predict. not which one is most advanced. whether your task is exploratory, measurable, recurring, or standing. the more you hand off, the less you babysit. full breakdown in the article below.

Hanako

476,080 просмотров • 1 месяц назад

SOMEONE TURNED 33 PILES OF DEAD BOOKMARKS INTO A GRAVITY MAP CLAUDE REBUILDS ITSELF EVERY NIGHT - AND IT RUNS ON THE 80% OF CLAUDE NOBODY TOUCHES most people drive Claude Code like a chatbot with file access - type a prompt, watch it edit, move on. that's maybe 20% of the tool this is the opposite. she's not typing at Claude. she's running it - loops on a mac mini overnight, claude linking every node while she sleeps the gravity map in the video is just the 80% maxed out: 1 system that organizes itself, not a human babysitting a chat box the other 80% is a steering layer Anthropic shipped quietly on june 18 - 7 ways to instruct the model, and a stack of commands almost nobody opens /context to see your bloat. /clear between tasks. path-scoped rules, subagents, hooks - conventions that load themselves the exact second they matter i stopped typing at Claude months ago - now i configure it once and it shows up already running the work, 10x cleaner a prompt helps for 1 message. the steering layer pays you back every session, for life the people who learn it stop being users and become operators - everyone else is still arguing about which model is smartest the article below is the full map - all 4 layers, every file and command, start to finish

SOMEONE TURNED 33 PILES OF DEAD BOOKMARKS INTO A GRAVITY MAP CLAUDE REBUILDS ITSELF EVERY NIGHT - AND IT RUNS ON THE 80% OF CLAUDE NOBODY TOUCHES most people drive Claude Code like a chatbot with file access - type a prompt, watch it edit, move on. that's maybe 20% of the tool this is the opposite. she's not typing at Claude. she's running it - loops on a mac mini overnight, claude linking every node while she sleeps the gravity map in the video is just the 80% maxed out: 1 system that organizes itself, not a human babysitting a chat box the other 80% is a steering layer Anthropic shipped quietly on june 18 - 7 ways to instruct the model, and a stack of commands almost nobody opens /context to see your bloat. /clear between tasks. path-scoped rules, subagents, hooks - conventions that load themselves the exact second they matter i stopped typing at Claude months ago - now i configure it once and it shows up already running the work, 10x cleaner a prompt helps for 1 message. the steering layer pays you back every session, for life the people who learn it stop being users and become operators - everyone else is still arguing about which model is smartest the article below is the full map - all 4 layers, every file and command, start to finish

KingWilliam

12,305 просмотров • 1 месяц назад

In our Anthropic Claude Design study, 5 designers approved a design system before they typed their first prompt. >Brand palette >type system >components the whole thing all set up. Only 1 of them named any of it in their opening prompt. That designer was the only one to finish production-ready. The other 4 assumed Claude would carry the system over. It didn't. TLDR: Claude doesn't reliably carry the design system you just approved. If you don't name it in the prompt, it doesn’t exist. It's never been a better time to be a designer, but you must learn the art of the prompt.

ben

163,778 просмотров • 2 месяцев назад

ANTHROPIC JUST DROPPED THE OFFICIAL GUIDE TO PROMPTING FABLE 5. This is the most important prompting framework I've seen. Bookmark this before you forget. Most people treat Fable 5 like a chatbot. That's the mistake. > don't over-engineer prompts — it degrades output. > use /loop for autonomous multi-step work. > give it the goal, not step-by-step commands. > add a memory file. it learns from past runs. > spin up 50+ subagents for complex tasks. Fable 5 isn't an assistant. it's a consultant that leads the work. Read it before you write another prompt. Claude → Fable 5 → Autonomous Work → Real Output → Money

ANTHROPIC JUST DROPPED THE OFFICIAL GUIDE TO PROMPTING FABLE 5. This is the most important prompting framework I've seen. Bookmark this before you forget. Most people treat Fable 5 like a chatbot. That's the mistake. > don't over-engineer prompts — it degrades output. > use /loop for autonomous multi-step work. > give it the goal, not step-by-step commands. > add a memory file. it learns from past runs. > spin up 50+ subagents for complex tasks. Fable 5 isn't an assistant. it's a consultant that leads the work. Read it before you write another prompt. Claude → Fable 5 → Autonomous Work → Real Output → Money

Kirill

414,789 просмотров • 1 месяц назад

an agent is four parts in a loop. you own one. the other three break it. that's why the demo works and prod doesn't. you can't debug what you can't see. 1) the prompt → what you tell the model each turn. you own this one. good. 2) the context window → what it sees right now. the framework fills it with junk, and you never notice until it rots. 3) the tools → what it can do. you own the list, not when or why it fires them. 4) the control flow → what happens next, when to stop. the framework owns this. it's what breaks at 80%. own all four and your agent stops being a magic trick that works on stage and dies on call. this isn't my idea. it's the 12-factor agents guide (24k stars) github: the whole thing every serious builder ends up rewriting their stack around. full breakdown in the article below.

an agent is four parts in a loop. you own one. the other three break it. that's why the demo works and prod doesn't. you can't debug what you can't see. 1) the prompt → what you tell the model each turn. you own this one. good. 2) the context window → what it sees right now. the framework fills it with junk, and you never notice until it rots. 3) the tools → what it can do. you own the list, not when or why it fires them. 4) the control flow → what happens next, when to stop. the framework owns this. it's what breaks at 80%. own all four and your agent stops being a magic trick that works on stage and dies on call. this isn't my idea. it's the 12-factor agents guide (24k stars) github: the whole thing every serious builder ends up rewriting their stack around. full breakdown in the article below.

Hanako

38,184 просмотров • 1 месяц назад

🚨Anthropic just gave Claude Code eyes and hands. read that again. it can now open apps on mac, click around macOS, find bugs visually, screenshot them, fix the code, rebuild, and verify the fix. one prompt. absolute zero human input. full autonomy. seems like Claude Code is becoming a super app. i’m so here for it.

🚨Anthropic just gave Claude Code eyes and hands. read that again. it can now open apps on mac, click around macOS, find bugs visually, screenshot them, fix the code, rebuild, and verify the fix. one prompt. absolute zero human input. full autonomy. seems like Claude Code is becoming a super app. i’m so here for it.

sui ☄️

27,357 просмотров • 4 месяцев назад

20 days ago, I connected Claude Code to my newly created instagram handle.. I gained 4.3M views and 6500+ followers in less than a month [ i post Ai generated animated stories ] Full workflow: i let claude study my account before i write another reel.. This is the cleanest content workflow i've built on claude. give it your IG first. 4 prompts handle the rest.. niche research, the reel script, the hook, and the daily automation.. the whole loop is basically, give claude your IG → find what's working → write retention-optimized scripts → engineer the hook → automate the daily output.. ▫️ Setup: give claude your instagram open claude code. claude code has a built-in web tool that browses any public URL. or install any agentic browser like Browser Harness or Firecrawl or Comet browser paste this with your handle filled in: "Browse and pull the last 30 reels and posts. Analyze my recurring topics, top-performing hooks, formats, and engagement patterns. Then map out my actual audience and what they consistently respond to." claude reads your profile, pulls every reel down, and now has the context to personalize every prompt below to YOUR account, not a generic niche. if you're on claude desktop, the same works with firecrawl MCP connected. ▫️ Prompt 1 find what actually goes viral in your niche: "Analyze the highest-performing Instagram Reels, TikToks, and Reddit posts in the [niche] niche from the last 30 days. Identify repeating hooks, visual styles, emotional triggers, and content formats that consistently generate high engagement. Then summarize the 5 strongest content angles optimized for AI-generated content and short-form videos." run this after the setup. you get 5 angles backed by what's already working in your niche, cross-checked against what's already working on YOUR account. ▫️ Prompt 2 write a high-retention reel script "Write a short-form Instagram Reel script about [topic] with an aggressive hook in the first 2 seconds. Create immediate curiosity, tension, or controversy to stop scrolling, then deliver a fast and satisfying payoff. Keep it under 30 seconds and optimize the structure for watch time, replays, comments, and shares. Finish with a subtle CTA." the line that matters: "optimize the structure for watch time, replays, comments, and shares." claude writes for the metrics, not just the word count. ▫️ Prompt 3 engineer better hooks "Study the top-performing Reels in [niche] and break down the hook structure, pacing, and emotional triggers used in the first 3 seconds. Then generate 5 new hook variations that are even more curiosity-driven, emotionally charged, and optimized to stop scrolling instantly. Focus on triggers like surprise, fear, ego, urgency, or desire." most reels die in the first 2 seconds. this prompt has claude reverse-engineer what already works, then give you 5 sharper versions to swap in. ▫️ Prompt 4 automate the whole workflow "Build a complete AI-powered content workflow for Instagram in the [niche] niche. The system should identify trending topics daily, generate high-retention scripts, create matching AI visuals, turn them into short-form videos, and generate optimized captions and hashtags. Structure everything as a repeatable workflow designed for consistent daily posting and growth." once the niche and script structure are validated, this turns it into a daily loop. one prompt that handles topic → script → visual → video → caption. these 4 prompts are the building blocks. the setup is what makes them yours. your real value is in the [niche] you plug in. content workflow built in one weekend, daily posting on autopilot from monday.

Axel Bitblaze 🪓

199,569 просмотров • 2 месяцев назад

FABLE 5 + HIGGSFIELD TURN A $35,000 ANIMATED SITE INTO A ONE-SESSION, $12 BUILD. HERE'S EXACTLY HOW. a studio runs this across four people and three weeks. you run it across one chat window and one afternoon. THE BUILD, STAGE BY STAGE: STAGE 1 - THE CONCEPT Claude reads your brief and scripts the scroll before a line of code exists - what the visitor feels at second 3, 15, 40. prompt: "read this brief. script the scroll beat by beat, then scaffold the project with GSAP ScrollTrigger + Lenis." STAGE 2 - THE VISUALS (Higgsfield) every hero shot, transition, and ambient loop comes out of 30+ generative models - matched to the story, not pulled from a stock library. prompt: "generate the hero sting and one b-roll clip per section. 3-5s, high-res, cinematic." STAGE 3 - THE MOTION (Claude Code) Claude writes the ScrollTrigger timelines and Lenis smooth-scroll, extracts frames, optimizes every asset. zero hand-coded keyframes. prompt: "wire the scroll: pin the hero, scrub the video, reveal each section on scroll. keep it 60fps on mobile." STAGE 4 - THE POLISH six cinematic effects baked in, no config: film grain, particles, vignette, glass cards, color tints, scroll pacing. prompt: "bake in the cinematic layer, then QA load speed, mobile breakpoints, and whether the scroll actually lands - rewrite what doesn't." CONNECT HIGGSFIELD (MCP): add it as a custom connector in Claude Code: mcp_servers: higgsfield: url: " one OAuth flow. Claude generates and pulls clips directly - no exporting by hand. THE MATH: → what a studio charges: $6,000-$35,000+ → what it costs you: a Claude sub + a few dollars of Higgsfield credits → what it takes: 4 people + 3 weeks → 1 operator + 1 session the pipeline was the moat. it just became four prompts. Follow me, comment "MATH" and I'll send you the full step-by-step Playbook. full breakdown in the article 👇

FABLE 5 + HIGGSFIELD TURN A $35,000 ANIMATED SITE INTO A ONE-SESSION, $12 BUILD. HERE'S EXACTLY HOW. a studio runs this across four people and three weeks. you run it across one chat window and one afternoon. THE BUILD, STAGE BY STAGE: STAGE 1 - THE CONCEPT Claude reads your brief and scripts the scroll before a line of code exists - what the visitor feels at second 3, 15, 40. prompt: "read this brief. script the scroll beat by beat, then scaffold the project with GSAP ScrollTrigger + Lenis." STAGE 2 - THE VISUALS (Higgsfield) every hero shot, transition, and ambient loop comes out of 30+ generative models - matched to the story, not pulled from a stock library. prompt: "generate the hero sting and one b-roll clip per section. 3-5s, high-res, cinematic." STAGE 3 - THE MOTION (Claude Code) Claude writes the ScrollTrigger timelines and Lenis smooth-scroll, extracts frames, optimizes every asset. zero hand-coded keyframes. prompt: "wire the scroll: pin the hero, scrub the video, reveal each section on scroll. keep it 60fps on mobile." STAGE 4 - THE POLISH six cinematic effects baked in, no config: film grain, particles, vignette, glass cards, color tints, scroll pacing. prompt: "bake in the cinematic layer, then QA load speed, mobile breakpoints, and whether the scroll actually lands - rewrite what doesn't." CONNECT HIGGSFIELD (MCP): add it as a custom connector in Claude Code: mcp_servers: higgsfield: url: " one OAuth flow. Claude generates and pulls clips directly - no exporting by hand. THE MATH: → what a studio charges: $6,000-$35,000+ → what it costs you: a Claude sub + a few dollars of Higgsfield credits → what it takes: 4 people + 3 weeks → 1 operator + 1 session the pipeline was the moat. it just became four prompts. Follow me, comment "MATH" and I'll send you the full step-by-step Playbook. full breakdown in the article 👇

ZEUS⚡️

47,174 просмотров • 1 месяц назад

in a similar vein to Kieran Klaassen & Every 📧 's "compound engineering", I've used a prompt every day for the past few months called "self-improving Cline" (linked below). at the end of the task, it evaluates any rule I had turned on and proposes updates based on any friction points during the task. now, all my clinerules are extremely refined and improving every time I use them.

in a similar vein to Kieran Klaassen & Every 📧 's "compound engineering", I've used a prompt every day for the past few months called "self-improving Cline" (linked below). at the end of the task, it evaluates any rule I had turned on and proposes updates based on any friction points during the task. now, all my clinerules are extremely refined and improving every time I use them.

Nick

42,766 просмотров • 1 год назад

Seedance 2.0 + Claude Code is f*cking insane 🤯 I built a Claude skill that creates UGC ads on demand. One product + one prompt = the AI creator, the script, the scene-by-scene shot list, and the finished video. All inside Claude Code. Perfect for DTC brands and agencies who can't afford to keep paying $500-$1,500 per UGC video and waiting 2 weeks for revisions. This skill eliminates the entire loop: → Tell Claude the product, ad angle, and length → Skill writes the GPT Image 2.0 prompt to generate the AI creator from scratch → Skill writes every scene prompt, dialogue line, and delivery direction → Pipes it into Seedance 2.0 with character + product + voice locked → Speed up + caption in CapCut → Ship the ad in 20 minutes No more paying $11 per video on Arcads. No more 2-week revision cycles. No more PR boxes to creators who ghost you. What you get: → Perfect character consistency across every scene → Voice consistency that holds clip-to-clip → Real product fidelity using your actual product photo as a reference → Multi-scene day-in-life, testimonial, and action-shot formats out of the box Built 100% with a Claude skill + Seedance 2.0. I recorded a full step-by-step tutorial showing the exact workflow so you can build these AI UGC ads yourself. Want the full breakdown? > Like this post > Comment "UGC" And I'll send it over (must be following so I can DM)

Seedance 2.0 + Claude Code is f*cking insane 🤯 I built a Claude skill that creates UGC ads on demand. One product + one prompt = the AI creator, the script, the scene-by-scene shot list, and the finished video. All inside Claude Code. Perfect for DTC brands and agencies who can't afford to keep paying $500-$1,500 per UGC video and waiting 2 weeks for revisions. This skill eliminates the entire loop: → Tell Claude the product, ad angle, and length → Skill writes the GPT Image 2.0 prompt to generate the AI creator from scratch → Skill writes every scene prompt, dialogue line, and delivery direction → Pipes it into Seedance 2.0 with character + product + voice locked → Speed up + caption in CapCut → Ship the ad in 20 minutes No more paying $11 per video on Arcads. No more 2-week revision cycles. No more PR boxes to creators who ghost you. What you get: → Perfect character consistency across every scene → Voice consistency that holds clip-to-clip → Real product fidelity using your actual product photo as a reference → Multi-scene day-in-life, testimonial, and action-shot formats out of the box Built 100% with a Claude skill + Seedance 2.0. I recorded a full step-by-step tutorial showing the exact workflow so you can build these AI UGC ads yourself. Want the full breakdown? > Like this post > Comment "UGC" And I'll send it over (must be following so I can DM)

Mike Futia

40,039 просмотров • 3 месяцев назад

SOMEONE BUILT A CLAUDE CODE PLUGIN THAT REFUSES TO WORK UNTIL YOU DO PUSH-UPS OR SQUATS its called workout gate and it is exactly as cursed as it sounds when the hook fires, it blocks your prompt, opens your webcam, and makes you exercise to keep coding > it counts your reps live through the webcam with mediapipe, no honor system rounding down > only releases your prompt once you actually finish the reps > quit halfway and the leftover reps get saved as debt for your next session > chill, demo and hardcore presets, plus prompt based, time based or random triggers > tracks stats, streaks and personal records right in the claude code statusline we automated the coding, now we're automating the part where we actually take care of ourselves

SOMEONE BUILT A CLAUDE CODE PLUGIN THAT REFUSES TO WORK UNTIL YOU DO PUSH-UPS OR SQUATS its called workout gate and it is exactly as cursed as it sounds when the hook fires, it blocks your prompt, opens your webcam, and makes you exercise to keep coding > it counts your reps live through the webcam with mediapipe, no honor system rounding down > only releases your prompt once you actually finish the reps > quit halfway and the leftover reps get saved as debt for your next session > chill, demo and hardcore presets, plus prompt based, time based or random triggers > tracks stats, streaks and personal records right in the claude code statusline we automated the coding, now we're automating the part where we actually take care of ourselves

Om Patel

671,324 просмотров • 2 месяцев назад

Claude Fable 5 is insane for voice-of-customer research 🤯 I just built a Claude Code skill that catches your customers quoting your own ads back to you. It reads your reviews, cross-references every recurring phrase against your website + ad copy, and sorts your "voice of customer" into three piles: Planted, category-standard, organic gold. All inside Claude Code. Perfect for DTC brands and creative strategists who brief ads off review mining. If you're pulling ad copy from your reviews, some of that language is real customer voice, some of it is your own tagline, and every time you re-use it, you're marketing to yourself a little harder. This skill breaks the loop: → Drop in any review export (Judge .me, Okendo, Amazon, Shopify) → It scrapes your site + ad copy automatically → Every recurring phrase gets 3 forensic tests (overlap, independence, category) → Verdicts come with receipts: counts, sources, confidence levels → Dark-mode dashboard + 5 ready-to-test hooks from the gold pile No API keys. No pip installs. No copy-paste prompt rituals. What you get: → The "planted" list — phrases you taught your customers (stop briefing off these) → The organic gold list — language customers use that your ads never have → 5 hooks built from real customer phrasing → A dashboard your whole team can read Runs 100% in Claude Code Want full playbook for free? > Like this post > Comment "Claude" And I'll send it over (must be following so I can DM)

Claude Fable 5 is insane for voice-of-customer research 🤯 I just built a Claude Code skill that catches your customers quoting your own ads back to you. It reads your reviews, cross-references every recurring phrase against your website + ad copy, and sorts your "voice of customer" into three piles: Planted, category-standard, organic gold. All inside Claude Code. Perfect for DTC brands and creative strategists who brief ads off review mining. If you're pulling ad copy from your reviews, some of that language is real customer voice, some of it is your own tagline, and every time you re-use it, you're marketing to yourself a little harder. This skill breaks the loop: → Drop in any review export (Judge .me, Okendo, Amazon, Shopify) → It scrapes your site + ad copy automatically → Every recurring phrase gets 3 forensic tests (overlap, independence, category) → Verdicts come with receipts: counts, sources, confidence levels → Dark-mode dashboard + 5 ready-to-test hooks from the gold pile No API keys. No pip installs. No copy-paste prompt rituals. What you get: → The "planted" list — phrases you taught your customers (stop briefing off these) → The organic gold list — language customers use that your ads never have → 5 hooks built from real customer phrasing → A dashboard your whole team can read Runs 100% in Claude Code Want full playbook for free? > Like this post > Comment "Claude" And I'll send it over (must be following so I can DM)

Mike Futia

10,164 просмотров • 1 месяц назад

This is genuinely insane 🤯 I sent the same prompt to 2 Claude Opus 5 agents 1 agent used no MCPs 1 agent used 2 MCPS the goal was to clone a very famous game, that has over 4 billion downloads and the results are scary.. The 1st agents game looks pretty basic. but has all the mechanics in place and even has power ups. The 2nd agents game looks polished especially for 1 single prompt. the 2nd agent used higgsfield mcp to create images for all of the assets then sent those images to Meshy AI with a prompt to create them it also generated all the UI elements. the results speak for themselves. you need to start MCP maxxing

This is genuinely insane 🤯 I sent the same prompt to 2 Claude Opus 5 agents 1 agent used no MCPs 1 agent used 2 MCPS the goal was to clone a very famous game, that has over 4 billion downloads and the results are scary.. The 1st agents game looks pretty basic. but has all the mechanics in place and even has power ups. The 2nd agents game looks polished especially for 1 single prompt. the 2nd agent used higgsfield mcp to create images for all of the assets then sent those images to Meshy AI with a prompt to create them it also generated all the UI elements. the results speak for themselves. you need to start MCP maxxing

Ernesto Lopez

68,858 просмотров • 22 дней назад

For Anthropic’s Claude Design, your opening prompt sets the ceiling of the project. If your prompt is a comprehensive brief, it will land somewhere near production-ready Seems obvious, but here is something interesting we observed: one designer wrote: >16 prompts >climbed in specificity over the session >and still finished worse than they started A single "reconsider" prompt at iteration 2 tanked output quality by -0.7 and the session never recovered. 2/3 of every prompt typed across all 5 sessions was refinement or correction. TLDR: Your first move is the whole game. and prompting is an art. What a time to be alive.

For Anthropic’s Claude Design, your opening prompt sets the ceiling of the project. If your prompt is a comprehensive brief, it will land somewhere near production-ready Seems obvious, but here is something interesting we observed: one designer wrote: >16 prompts >climbed in specificity over the session >and still finished worse than they started A single "reconsider" prompt at iteration 2 tanked output quality by -0.7 and the session never recovered. 2/3 of every prompt typed across all 5 sessions was refinement or correction. TLDR: Your first move is the whole game. and prompting is an art. What a time to be alive.

ben

229,081 просмотров • 3 месяцев назад

Don't train the model, evolve the harness. I read a brilliant blog post from Hugging Face where they took a frozen open model scoring 0% on a hard legal agent benchmark, left its weights alone, and let an automated loop rewrite only the code around it. That code layer is the harness, the runtime wrapper that feeds the model context, runs its tool calls, and decides when a run ends. By the time the loop finished, the system had essentially matched Sonnet 4.6 on the benchmark's headline metric, at roughly 7x lower cost per task. Zero weights changed. The gain existed because of where the model was failing. The judge only grades files saved in the right place under the exact requested filename, and the model kept doing the legal analysis correctly, then saving it under the wrong name, dropping it in a scratch folder, or never writing it at all. So the 0% was never measuring legal reasoning. It was measuring the harness. Hand-tuning that layer is slow and model-specific, so they automated it. A Claude proposer adds exactly one mechanism per iteration, and an outer loop keeps it only if it clearly beats the current best, so accepted mechanisms compound. What the loop discovered says a lot about where agents actually fail. → The biggest single gain was file handling, not intelligence. An automatic step that lands the deliverable exactly where the judge expects it beat every prompt change, with zero extra model tokens. → Code fixes transferred across models, prompt playbooks did not. The same harness lifted a smaller model from the same family by 14 points, but the tuned prompts hurt a different model family on tasks it could already finish. → The harness mattered more than anything else. Same model, same judge, same tasks, and five different harnesses scored anywhere between 3.5% and 80.1%. The gains do eventually flatten, and the remaining misses look like real capability gaps. At some point the wrapper runs out of tricks and the model has to carry the work. But the lesson holds. A benchmark score measures the model and its harness together, and until the harness is fixed, it's impossible to know which one failed. I highly recommend reading this: I also wrote a deep dive on agent harness engineering a while back, covering the orchestration loop, tools, memory, context management, and everything that turns a stateless LLM into a capable agent. The article is quoted below.

Don't train the model, evolve the harness. I read a brilliant blog post from Hugging Face where they took a frozen open model scoring 0% on a hard legal agent benchmark, left its weights alone, and let an automated loop rewrite only the code around it. That code layer is the harness, the runtime wrapper that feeds the model context, runs its tool calls, and decides when a run ends. By the time the loop finished, the system had essentially matched Sonnet 4.6 on the benchmark's headline metric, at roughly 7x lower cost per task. Zero weights changed. The gain existed because of where the model was failing. The judge only grades files saved in the right place under the exact requested filename, and the model kept doing the legal analysis correctly, then saving it under the wrong name, dropping it in a scratch folder, or never writing it at all. So the 0% was never measuring legal reasoning. It was measuring the harness. Hand-tuning that layer is slow and model-specific, so they automated it. A Claude proposer adds exactly one mechanism per iteration, and an outer loop keeps it only if it clearly beats the current best, so accepted mechanisms compound. What the loop discovered says a lot about where agents actually fail. → The biggest single gain was file handling, not intelligence. An automatic step that lands the deliverable exactly where the judge expects it beat every prompt change, with zero extra model tokens. → Code fixes transferred across models, prompt playbooks did not. The same harness lifted a smaller model from the same family by 14 points, but the tuned prompts hurt a different model family on tasks it could already finish. → The harness mattered more than anything else. Same model, same judge, same tasks, and five different harnesses scored anywhere between 3.5% and 80.1%. The gains do eventually flatten, and the remaining misses look like real capability gaps. At some point the wrapper runs out of tricks and the model has to carry the work. But the lesson holds. A benchmark score measures the model and its harness together, and until the harness is fixed, it's impossible to know which one failed. I highly recommend reading this: I also wrote a deep dive on agent harness engineering a while back, covering the orchestration loop, tools, memory, context management, and everything that turns a stateless LLM into a capable agent. The article is quoted below.

Akshay 🚀

244,567 просмотров • 1 месяц назад

I STOPPED REVIEWING MY OWN AGENT, SOMETHING ELSE DOES IT NOW I used to read every diff it produced and approve most of them, because an agent grading itself always says the work is good. -> Now a second model with different instructions tries to break the work first, and I only read what survived. Here is what is actually in the folder that took over the night shift: • the brief > CONTRACT.md -- what it may touch, and what it may never touch. > VISION.md -- the destination, so turn 47 still knows why it started. • the gate > judge/ -- a different model, never the one that wrote the code. > break-it.md -- it opens the page, clicks, screenshots, reports back. > -- no opinion, just zero or non-zero. > shift.yml -- 03:30 every night, laptop closed. • the memory > receipts/ -- one folder per night, dated and graded. > STATE.md -- where it stopped and what it escalated. > lessons.log -- the flaky test, written down once instead of rediscovered weekly. • the brakes > caps.json -- turn limit, retry limit, spend limit. > -- written on day one, used never. The generator decides what your loop can produce -> The judge decides what it refuses to produce. One of those is the part everyone builds -> The other is why most loops quietly fail. Bookmark it & Read Full breakdown below ↓

I STOPPED REVIEWING MY OWN AGENT, SOMETHING ELSE DOES IT NOW I used to read every diff it produced and approve most of them, because an agent grading itself always says the work is good. -> Now a second model with different instructions tries to break the work first, and I only read what survived. Here is what is actually in the folder that took over the night shift: • the brief > CONTRACT.md -- what it may touch, and what it may never touch. > VISION.md -- the destination, so turn 47 still knows why it started. • the gate > judge/ -- a different model, never the one that wrote the code. > break-it.md -- it opens the page, clicks, screenshots, reports back. > -- no opinion, just zero or non-zero. > shift.yml -- 03:30 every night, laptop closed. • the memory > receipts/ -- one folder per night, dated and graded. > STATE.md -- where it stopped and what it escalated. > lessons.log -- the flaky test, written down once instead of rediscovered weekly. • the brakes > caps.json -- turn limit, retry limit, spend limit. > -- written on day one, used never. The generator decides what your loop can produce -> The judge decides what it refuses to produce. One of those is the part everyone builds -> The other is why most loops quietly fail. Bookmark it & Read Full breakdown below ↓

slash1s

37,556 просмотров • 19 дней назад

19-year-old from china makes $9,000/month designing product sites and ships each one in an afternoon. here's his exact setup the whole thing runs on two tools that each do one job: > brief written by hand: 5 min > Moonchild builds the design system, then every screen from it: 20 min > MCP hands the design to Claude as real structure, not a screenshot: instant > Claude Code reads those exact tokens and builds the live app: 20 min > second Claude session reviews the build for drift: 10 min total: about an hour. screen five still matches screen one. no agency, no dev, no design team the trick is MCP. the design tool passes Claude the actual colors, components and layout, so it builds from the source instead of guessing from a picture. full pipeline, every prompt, in the article above.

19-year-old from china makes $9,000/month designing product sites and ships each one in an afternoon. here's his exact setup the whole thing runs on two tools that each do one job: > brief written by hand: 5 min > Moonchild builds the design system, then every screen from it: 20 min > MCP hands the design to Claude as real structure, not a screenshot: instant > Claude Code reads those exact tokens and builds the live app: 20 min > second Claude session reviews the build for drift: 10 min total: about an hour. screen five still matches screen one. no agency, no dev, no design team the trick is MCP. the design tool passes Claude the actual colors, components and layout, so it builds from the source instead of guessing from a picture. full pipeline, every prompt, in the article above.

Ridark

19,477 просмотров • 2 месяцев назад

Claude, Obsidian and n8n form a second brain that clears $5,200 months from brand UGC deals. Most creators still open a blank CapCut file and hope the brand brief sticks. That loop dies at $1,800. The stack replaces hope with a closed system. Obsidian holds every brand voice, every past script that hit 400k views, every rejection note. Claude reads the vault through a local project and rewrites the next five UGC packages in 3 minutes. n8n watches the inbox, creates the note, triggers Claude, formats the brief, and drops the finished package into the brand portal while the creator sleeps. No more research tabs. No more “what angle worked last time.” Month 1: build the vault and wire the three tools. Month 2: first automated batch ships. Two brands pay $1,100 total. Month 3: the same loop runs on four brands. Average deal $480. Twelve packages leave. Bank shows $5,200. The second brain does not store notes. It compounds every win into the next model. Script that scored 18% engagement gets tagged. Claude pulls only those tags next run. The model improves without extra prompt engineering. UGC brands pay for consistency, not inspiration. This system delivers the same quality every week at the same speed. Set the vault once. Let the three tools run the rest. If this was useful - follow.

Rugikk

10,318 просмотров • 1 месяц назад

Making OpenCode as lean as Pi agent? Just trimmed 25k out of OpenCode's system prompt (from 30k to 4-5k tokens) How? Just disable skills and get rid of massive skill definition bloat. Who needs skills anyway? Just kidding, this is the not the way. It makes the agent lame and defeats the point of using one. But it sets a precedent: Find a way to use skills without their definitions pre-loaded into the system prompt every single turn. Another interesting stuff: Upon testing this temporary "no skill setup" with two of hottest OpenCode Zen free models, Mimo V2.5 vs DeepSeek V4 Flash: One thinks more and talks less One thinks less and talks more Check the video to see which is which If you made it here, I'm finding a way to leanest OpenCode setup that I can get I simply don't believe that OpenCode can't be as lean as Pi Upon tinkering, I made a plugin that temporarily extracts the system prompt while I test, and noticed the hundreds of definitions in it from my .agents/skills directory which is shared across all my coding agents (Cursor, Antigravity, Claude, etc.) Of course disabling skills is not the answer, but it just proved that there is a way to strip the system prompt of these massive skill defs Aside from the system prompt hierarchy that injects confusion imo if you have a conflicting and redundant AGENTS.md which I discovered upon digging into OpenCode's source code Apparently it has prompt.ts/system.ts/instruction.ts/llm.ts and loads base .txt prompts based on model family (claude/gpt-o/gpt-5/codex/gemini/others) that all work together to make OpenCode aware of who it was and how it should use tools and become a "coding agent" Gotta find the most minimal mix that fits right into my workflow Make OpenCode as lean as Pi? We'll see. All in

Making OpenCode as lean as Pi agent? Just trimmed 25k out of OpenCode's system prompt (from 30k to 4-5k tokens) How? Just disable skills and get rid of massive skill definition bloat. Who needs skills anyway? Just kidding, this is the not the way. It makes the agent lame and defeats the point of using one. But it sets a precedent: Find a way to use skills without their definitions pre-loaded into the system prompt every single turn. Another interesting stuff: Upon testing this temporary "no skill setup" with two of hottest OpenCode Zen free models, Mimo V2.5 vs DeepSeek V4 Flash: One thinks more and talks less One thinks less and talks more Check the video to see which is which If you made it here, I'm finding a way to leanest OpenCode setup that I can get I simply don't believe that OpenCode can't be as lean as Pi Upon tinkering, I made a plugin that temporarily extracts the system prompt while I test, and noticed the hundreds of definitions in it from my .agents/skills directory which is shared across all my coding agents (Cursor, Antigravity, Claude, etc.) Of course disabling skills is not the answer, but it just proved that there is a way to strip the system prompt of these massive skill defs Aside from the system prompt hierarchy that injects confusion imo if you have a conflicting and redundant AGENTS.md which I discovered upon digging into OpenCode's source code Apparently it has prompt.ts/system.ts/instruction.ts/llm.ts and loads base .txt prompts based on model family (claude/gpt-o/gpt-5/codex/gemini/others) that all work together to make OpenCode aware of who it was and how it should use tools and become a "coding agent" Gotta find the most minimal mix that fits right into my workflow Make OpenCode as lean as Pi? We'll see. All in

raymel 👋

37,478 просмотров • 2 месяцев назад

1 Neural Network + Obsidian + Karpathy’s 1-file method = the most unhinged second brain build of 2026. It remembers everything you’ve ever done, and it costs $0 on top of what you already pay. The base is Karpathy’s append and review: 1 giant note, new thoughts stack on top, old ones sink, every few days you reread and pull the survivors back up. No folders, no tags, no plugins the rereading IS the system, because review is what turns storage into thinking. The flaw: past 10,000 lines, no human rereads anything. That’s where the neural network takes over. You keep the note in Obsidian 1 vault, everything dumps to the top: ideas, links, meeting fragments, half-thoughts. You never organize, you only dump. It all lives as plain markdown on your own disk, and that detail is the whole trick. Because now you point Claude Code at the vault folder, and it reads every line you’ve ever written. “What did I think about pricing in March.” “Find the 3 ideas I keep circling.” “What did I drop that deserves a second look.” It answers from YOUR notes, with quotes, in 15 seconds. Then once a week, 1 prompt closes the loop: read the last 7 days, surface the 5 entries worth pulling back up, flag anything that contradicts what I wrote a month ago. The model does the sinking and surfacing Karpathy did by hand, and the note stays alive instead of turning into a graveyard. Week 1 feels like nothing. Week 4 you hit the first “I already solved this in January.” Month 3 you consult your past self more than Google. Most second brains die in 11 days under 40 plugins and 200 folders. This one is 1 file and a loop, and it compounds because dumping takes 0 discipline. Notion stores what you thought. This thing argues back.

1 Neural Network + Obsidian + Karpathy’s 1-file method = the most unhinged second brain build of 2026. It remembers everything you’ve ever done, and it costs $0 on top of what you already pay. The base is Karpathy’s append and review: 1 giant note, new thoughts stack on top, old ones sink, every few days you reread and pull the survivors back up. No folders, no tags, no plugins the rereading IS the system, because review is what turns storage into thinking. The flaw: past 10,000 lines, no human rereads anything. That’s where the neural network takes over. You keep the note in Obsidian 1 vault, everything dumps to the top: ideas, links, meeting fragments, half-thoughts. You never organize, you only dump. It all lives as plain markdown on your own disk, and that detail is the whole trick. Because now you point Claude Code at the vault folder, and it reads every line you’ve ever written. “What did I think about pricing in March.” “Find the 3 ideas I keep circling.” “What did I drop that deserves a second look.” It answers from YOUR notes, with quotes, in 15 seconds. Then once a week, 1 prompt closes the loop: read the last 7 days, surface the 5 entries worth pulling back up, flag anything that contradicts what I wrote a month ago. The model does the sinking and surfacing Karpathy did by hand, and the note stays alive instead of turning into a graveyard. Week 1 feels like nothing. Week 4 you hit the first “I already solved this in January.” Month 3 you consult your past self more than Google. Most second brains die in 11 days under 40 plugins and 200 folders. This one is 1 file and a loop, and it compounds because dumping takes 0 discipline. Notion stores what you thought. This thing argues back.

West Lord

24,679 просмотров • 1 месяц назад