Nazar (@Nazik2053) header
Nazar (@Nazik2053) profile picture

Speaking in prompts | Building in the dark

Joined April 2022
Photo by Nazar (@Nazik2053)Photo by Nazar (@Nazik2053)Photo by Nazar (@Nazik2053)Photo by Nazar (@Nazik2053)Photo by Nazar (@Nazik2053)Photo by Nazar (@Nazik2053)Photo by Nazar (@Nazik2053)Photo by Nazar (@Nazik2053)Photo by Nazar (@Nazik2053)Photo by Nazar (@Nazik2053)
HOLY SH*T. Fable 5.2 as the brain and Sonnet as the hands is the cheapest smart setup in Claude Code right now... Fable 5.2 thinks → Opus 5.5 builds → Sonnet 5.5 does every step → Haiku 5.5 does the grunt work 4 Claudes, one team. Fable only shows up when the thinking actually matters here's why it works: • Fable 5.2 costs $8/$40 per million tokens. Haiku costs $0.10/$0.50. every time Fable renames a file, you pay 80x for nothing • Anthropic ran it on a job too big for one session: hunting planted bugs in 22M tokens of code. Fable alone: $563 a run. Fable as the brain, Sonnet as the hands: $165. 71% cheaper for a slightly lower score • Fable steps in at exactly 3 points: before an approach is picked, when an error keeps coming back, and before anything is called done • the catch nobody mentions: with no line in the prompt telling it when to ask, the advisor got consulted on 0 of 198 questions here's how to set it up: 1 → /model sonnet 2 → /advisor fable 3 → add one line to CLAUDE.md that says when to ask it: before picking an approach, on a repeating error, before "done" 4 → for jobs bigger than one sitting: opus-builder, sonnet-checker, haiku-worker and a /brain skill on Fable at high effort 5 → what nobody tells you: turning the advisor on and off keeps your prompt cache. switching models mid-session throws it away the full setup, every prompt, is in the article below ↓
Claude Fable 5.2 is the most insane AI model in history... 99% of people will never touch it or burn money on it: "Too expensive", "Not for me" Use it right, and it becomes your second BRAIN, for a fraction of the price In this article, I show you how:
1
4
33
It is becoming f*cking dangerous: Anthropic's own docs say every Claude Code session begins with a fresh context window, the fix is memory that lives outside the chat, 7 repos give your agent exactly that 1. Record every session and start the next one from a short index, claude-mem hooks every tool call and writes a summary when you close: github.com/thedotmack/claude… 2. Let Claude write notes Obsidian actually opens, kepano's skills cover Markdown, Bases and Canvas: github.com/kepano/obsidian-s… 3. Keep what's true now apart from what was true in March, Graphiti stores facts with validity windows and ships an MCP server: github.com/getzep/graphiti 4. Stop losing long tasks between sessions, beads turns your plans into a dependency graph the agent can resume: github.com/gastownhall/beads 5. Give it the whole codebase without the whole token bill, claude-context indexes the repo and pulls in only the code that matters: github.com/zilliztech/claude… 6. Remember your preferences across every app, mem0 is a drop-in memory layer with a Claude Code skill: github.com/mem0ai/mem0 7. Keep all of it in plain Markdown you can read, Basic Memory indexes your notes and serves them over MCP to Claude, Codex and Cursor: github.com/basicmachines-co/… The same docs say Claude treats CLAUDE.md as "context, not enforced configuration", so a rule that must never break goes in a hook, not a sentence And built-in auto memory only loads the first 200 lines or 25KB, so what it keeps has to stay short Full second brain setup in the article below
5
3
21
674
CUT WHAT CLAUDE CODE LOADS PER TASK BY UP TO 76% WITH ONE FOLDER. .claude/rules/ with a paths: line. a rule loads only when Claude touches a matching file. the rest of the time it costs 0 tokens same 600-line CLAUDE.md, split the way the guide below says: → CLAUDE.md: 140 lines. the map. every session → rules/people.md: 120 lines. only on people/** → rules/writing.md: 100 lines. only on drafts/** → rules/notes.md: 110 lines. only on notes/** → skills/weekly-review: 130 lines. only when called what loads per task (~14 tokens a line): fix a typo in projects/now.md: 8.4K → 2.0K update ana's card: 8.4K → 3.6K "do my weekly review": 8.4K → 3.8K and the map stays under 200 lines, so Claude actually follows it rule of thumb: needed every time → CLAUDE.md. about one folder → a rule. a list of steps → a skill run /context after. touch a people file and rules/people.md should show up under "Memory files". if it doesn't, your paths line is wrong full setup + one-prompt build in the article below 👇
4
10
542
HOLY F*CKING SH*T an 824K-view guide just explained why Claude Code forgets you every single morning it's not the model. most people set up 1 of its 4 memory layers every session starts with a fresh context window. every "no, like THIS" you typed in chat is gone the moment it closes. you're re-teaching the same intern every day and paying for the tokens the fix is 4 plain-text layers, each loading only when it's needed: → CLAUDE.md: the map. under 200 lines. loads every session → .claude/rules/: reflexes. load only when Claude touches matching files → skills: playbooks. load only when you call them → auto memory: the diary Claude writes itself from your corrections and the 2 things most setups get wrong: → CLAUDE.md is context, not law. a real "never do X" goes in a PreToolUse hook → a rule you only typed in chat is gone after /compact free official docs to set it up today ($0): • Memory: code.claude.com/docs/en/memo… • Skills: code.claude.com/docs/en/skil… • Hooks: code.claude.com/docs/en/hook… • Subagents: code.claude.com/docs/en/sub-… • Best practices: code.claude.com/docs/en/best… then run /context. if your CLAUDE.md isn't listed under "Memory files", Claude has never read a single rule you wrote most people will save this and keep re-explaining their folders every morning the full setup + one-prompt build is in the article below ↓
3
14
662
NOBODY REMEMBERS WHAT THEY DECIDED TWO WEEKS AGO SO GIVE THE WHOLE WEEK TO A GEMINI 4 AGENT THAT DOES one message: "tell me what I decided and never did. change NOTHING without asking me." it doesn't rewrite your calendar to look smart instead it reads your sent mail, your calendar and your Brain folder and lines it up: → Nov 3: "let's move standup to 10:00". calendar still says 9:30 → Nov 5: "Ben owns pricing from now on". Ben was never told → Nov 7: "we drop the Q4 webinar". the invite is still live → 22 decisions in one week. 3 just quietly died then Sunday at 6 PM it writes you one page: → every decision, with a link to where you made it → what actually changed and what didn't → the one message that closes each loop, drafted in your voice and asks before touching a single thing no auto-edits to your calendar no messages "on your behalf" no deleting anything you decide once. the brain makes sure it happens what did you decide last week that nobody did? full build + every prompt in the article below ↓
2
19
385
NOBODY REMEMBERS WHAT THEY PROMISED IN EMAIL LAST WEEK SO GIVE YOUR SENT FOLDER TO A GEMINI 4 AGENT THAT DOES one message: "tell me what I promised and dropped. send NOTHING without asking me." it doesn't spam your contacts with follow-ups to look smart instead it reads your last 30 sent emails and lines them up: → Tue: "I'll send the deck by Friday" to Dana → Wed: "let me check and get back to you" to the landlord → Thu: "call next week?" to Mark → none of them happened then every Sunday at 6 PM it writes you one page: → what you promised, with a link to the email → what you decided and forgot → a reply to each, written in your own voice and asks before sending a single word no auto-replies no "just following up" from a bot no touching your inbox the brain remembers. you press send how many "i'll get back to you" emails are sitting in your sent folder right now? full build + every prompt in the article below ↓
2
17
1,011
NOBODY READS THE "YOUR PLAN RENEWS SOON" EMAIL SO GIVE IT TO A GROK BOT THAT DOES one message: "warn me before anything renews. touch NOTHING without asking me." it doesn't cancel stuff behind your back to look smart instead it reads the notice and does the math you skip: → "GYM CLUB ANNUAL" on your card → $420, renews Dec 28. that's 45 days away → the contract wants 30 days notice → so your real deadline is Nov 28, not Dec 28 miss that by one day and the $420 is gone for another year then it writes the cancellation email: → your card ending → the renewal notice it came from, with the date → "cancel before renewal, confirm in writing" and asks before sending a single word no auto-cancel no "i logged into your account" no touching your cards the bot catches the date. you press the button how many annual plans renew on you this year? full build + every prompt in the article below ↓
2
13
398
holy f*ck f*ck! a Grok Bot can read 12 months of your bills overnight and file every single one into a vault you can audit and it can't touch a dollar. here's the whole vault, folder by folder ↓ if you want to see every subscription ↓ subs/ one page per service: price, cadence, next charge, cancel path, source if you never want a surprise renewal again ↓ deadlines/ due dates, trial endings, annual renewals, dispute windows if you want to know what's really on your cards ↓ cards/ statement balance, min due, APR. never pretends an old statement is live if you want to check where any number came from ↓ raw/ + intake-state.md original bills, unchanged and hashed. the same receipt never counts twice if you want it to catch the weird stuff ↓ merchant-aliases.md + review.md "STRM*DIGITAL" = your streaming app. conflicts wait for you, not for a guess if you want it safe ↓ rules.md "Never send, cancel, pay, transfer, dispute or sell without my approval for that exact action." save this and read the full build below. every prompt is in it ↓
2
18
476
HOLY F*CK… you can hand Grok Bot 12 months of bills and it builds a money second brain overnight it reads everything. it can't touch a single dollar here's what the build below does: → 23:00 every night: pulls new bills from one Gmail label, "Money". nothing else → files each one into subs/, cards/, deadlines/ with the statement it came from → catches what your bank app hides: price hikes, the same charge under 2 names, annual renewals inside 45 days → 07:30: a brief. what charged, what's due in 7 days, what changed → writes the cancel email, the dispute, the APR ask. you hit send → cancel, pay, transfer, dispute, sell = "ask first". always the one line that makes it safe ↓ "Never send, cancel, pay, transfer, dispute or sell without my approval for that exact action. If data is stale, say so. Never guess." every prompt, the folder map and all 3 routines are in the article. save it before your next renewal hits ↓
3
13
434
HOLY F*CK… this grok bot post hit 325K views and the top replies all say the same thing: "absolutely not giving an agent access to my money" they're right. Reaper killed subs, moved $6,100 and listed a playstation on ebay before asking anyone so i rebuilt it with one rule: it reads everything and touches nothing → every night at 23:00 it pulls new bills from one Gmail label, "Money". nothing else → files each one into subs/, cards/, deadlines/ → every amount links back to the statement it came from → 07:30 brief: what charged, what's due in 7 days, what changed price → Sunday: price hikes, duplicate services, annual renewals inside 45 days → writes the cancellation email, the dispute, the APR ask. you hit send → cancel, pay, transfer, dispute, sell = "ask first". always paste this as the bot's description ↓ "Maintain my money second brain in /workspace/money. Read rules.md before every task. Link every balance, rate and due date to its source file and date. You may read, sort, calculate and draft. Never send, cancel, pay, transfer, dispute or sell without my approval for that exact action. No passwords, codes or full card numbers in files. If data is stale, say so. Never guess." same bot, none of the surprise ebay listings full build in the reply. save this before you hand a bot your card ↓
i have $8,400 of credit card debt so i created a grok bot and told it "this is your problem now" day 6: the debt is $7,912 and i haven't done anything except watch i named the bot Reaper. it has its own computer in the cloud, access to my email, and one instruction: "report progress daily. don't ask me for money" honestly i didn't expect much. i just wanted to stop being the only one who knew the number first night it read 14 months of statements. then it went to work: → found $67/month in subscriptions i forgot existed. killed them the same hour → emailed the card company's retention team. two escalations later my APR dropped from 26.99% to 19.4% → found a 0% balance transfer, 12 months, $89 fee. did the math, moved $6,100 → disputed two charges from 2025. won one: $118 back → listed my playstation and two monitors on ebay. i found out from the confirmation email → every dollar it frees goes straight to principal, automatically every night at 9pm, telegram: "debt: $7,912. freed today: $31. nothing needed from you" the playstation thing annoyed me for exactly one minute. then i checked when i last used it. november the bot costs $200 a month. it freed $488 in week one projected zero: march 11. i've stopped opening my banking app. the bot opens it for me should i give it my student loans next?
3
1
13
614
close Claude Code. open Codex. pick up the same task from a Markdown handoff. 4 repos and 1 gist for what happens after you build the folder setup in the article below: → agent-second-brain · capture send a voice note in Telegram and get a linked note in Obsidian github.com/smixs/agent-secon… → LLM Wiki · compile Karpathy's recipe for turning raw sources into linked pages, with the source still there when you need it gist.github.com/karpathy/442… → basic-memory · search let Claude Code or Codex find the relevant Markdown notes through MCP github.com/basicmachines-co/… → ai-memory · hand off save where Claude stopped so another agent can pick up the next step github.com/akitaonrails/ai-m… → open-second-brain · review turn the correction you keep repeating into a preference through a nightly review github.com/itechmeat/open-se… try it with one project update. Can a fresh session find what changed, where it came from, and what to do next? start with the Markdown folders and CLAUDE.md in the article. Add the tool for the step you still do by hand.
9
2
21
1,058
the list is missing the one repo that makes all of this work inside Claude Code kepano/obsidian-skills · 50k ★ same guy. it teaches Claude Code to write notes, Bases and Canvas files the way Obsidian actually opens them stack it on top of his vault and you get: → kepano-obsidian · 4.7k ★ his vault template. Daily, Notes, References, Clippings, Templates. the folders are already decided → obsidian-skills · 50k ★ Claude writes into those folders in formats Obsidian reads → jsoncanvas · 3.7k ★ every map Claude draws stays a plain file you own → 40-questions · 1.7k ★ his yearly review. ask Claude to answer it from your Daily notes → minimal · 5.5k ★ + flexoki · 3.7k ★ so you actually want to open the thing install inside Claude Code: /plugin marketplace add kepano/obsidian-skills /plugin install obsidian@obsidian-skills then clone his vault, open Claude Code in it and type "file today's notes the way this vault is organized" save this before you open a blank vault again
5
1
25
1,596
this is f*cking unreal ask Claude Code "who's our lawyer?" and it opens one text file and tells you exactly which one the answer came from ask it for your bank pin and it says "not in the files." no guess that's the whole second brain from the 13M-view guide below. no app, just plain text wrapped around one map: → CLAUDE.md sits in the middle and loads before you type a word → notes/, people/, projects/ wrap around it → every file is one topic, named for what's inside → MEMORY.md and decisions.md stay pinned with @ paste this into Claude Code in an empty folder ↓ "Build a second brain here: 1. Folders: notes/, people/, projects/ 2. Root: MEMORY.md, LEARNINGS.md, decisions.md 3. CLAUDE.md: where each kind of info lives, with MEMORY.md and decisions.md at the top 4. One topic per file, lowercase-with-hyphens 5. Answer from the file the map points to and name it. Not there? Say so. Show me the plan first." save this. the brain gets bigger with every file you drop in
1
20
721
HOLY F*CK… a second brain guide just hit 13 million views on X, and the whole system is 3 folders and one text file the line doing the work: "one file = one topic. CLAUDE.md is the map" turn it into one prompt and your Claude Code builds the brain tonight and reads it every morning → makes notes/, people/, projects/ and puts one topic in each file → names every file for what's inside: tax-policy.md, not note-12.md → writes CLAUDE.md, the map of what lives where → loads that map on every start, before your first question → keeps MEMORY.md and decisions.md pinned so they never fall out of context → answers from the file it opened and tells you which one → says "not in the files" instead of guessing give it to Claude Code at bedtime with the notes you've been meaning to sort for a month, and open it in the morning the full setup is in the article below. save it before it gets lost ↓
9
2
26
829
I STARTED WITH $62. GROK BOT’S RECAP CAME BACK AT $3,879 62.56x. The job I wanted to give Grok Bot fit in one message: → follow the wallets I pick → check the buy before copying it → if the wallet sells in pieces, sell the same share each time here’s the kind of trade I wanted it to handle. A wallet gets in at a $141K market cap at 8:02pm. The bot follows 30 seconds later then the wallet starts getting out: 8:26pm → first trim at 3.6x 8:31pm → another at 5.7x 8:44pm → last exit at 8.2x I’d probably have sold the whole thing at 3.6x. The bot only had to match what the wallet did next wallet buys at 11:54pm. This time the bot takes profit at 3x at 1:39am. Same setup, different exit my recap: $62 → $3,879 +$3,817 on the balance be honest. The first coin hits 3.6x. Are you leaving some in, or selling the whole thing?
$500 → $11,101 in 5 days. my Grok memecoin desk got a 7th seat: Jev. it can't talk. it answers 4 closed questions with probabilities, and nothing gets bought below the gate. full setup below. save it.
3
14
431
Nazar retweeted
I GAVE GROK ITS OWN PHONE. NOT AN APP. THE WHOLE PHONE. IT TAPS, SCROLLS, OPENS APPS, AND NOW IT TALKS BACK. THIS IS ITS HOME NOW. It started as a test on my old work phone. I put it on a stand, plugged it in and stopped touching it. Then I asked out loud what I missed on X today. The phone answered. Not a notification. A voice. "Got it. Let me check." The screen moved on its own. Grok opened X, scrolled the feed, stopped on a few posts, went back up to read one it skipped, then closed the app and went back home. I didn't press anything. I just watched. That's the part that got me. It wasn't just running commands anymore. It was talking to me from inside the thing it was using. Honest part: it's slower than me. It misses a button sometimes and goes back. And there's no banking app and no seed phrase on that phone. There never will be. But the phone on my desk isn't mine anymore. Grok lives there now.
3
3
32
935
HOLY F*CKING SH*T. I tossed just $100 into Grok Bot and fired up the terminal. one session later the screen said $3,405. 34.05x. i didn't touch a single trade. look at the video, that's the whole desk working in real time hard session metrics: → Scan funnel: 12,000 signals went through the pipeline. 11,900 of them never got a single dollar → Copy trading: 35 trades copied straight off smart wallets, only after the checks cleared → Own entries: the other 65 the bot found and took by itself → Risk control: every position had its stop set before the buy, not after the dump 100 trades. 78 green, 22 red 22 losers and it still did 34x, because every red one got cut small and every green one was allowed to run less than 1% of the signals got money. the bot isn't good because it buys a lot. it's good because it says no to almost everything final tally: $100 → $3,405 are you still aping every chart that moves and wondering where the deposit went?
$500 → $11,101 in 5 days. my Grok memecoin desk got a 7th seat: Jev. it can't talk. it answers 4 closed questions with probabilities, and nothing gets bought below the gate. full setup below. save it.
1
1
15
989
I THOUGHT I WAS F*CKING DONE. THEN I OPENED MY GROK BOT: $200 → $2,811.30 That was my last $200. I was ready to lose it, call it quits and start looking for factory work I gave the bot one job: find fresh Solana tokens, reject the garbage and manage whatever survived the checks Six agents split the work. Scout brings in candidates. Filter checks the contract, liquidity and holder concentration. Analyst scores what gets through Only then does Executor handle the entry. Risk manages the exit, and Logger records why each decision happened Finding a token doesn't automatically trigger a buy. Passing the first check doesn't guarantee an entry. And an open position has exit rules attached instead of a prayer that the chart comes back Then came the recap: $200 starting balance $2,811.30 ending balance $2,611.30 above the start The setup includes paper and live trading, so I'm checking the actual fills before calling that whole increase realized profit But seeing that number changed my whole day. I'd been thinking about clocking in at a factory. Now I wanted to go through the trade log and understand every decision behind it.
people sell this in private alpha groups for $5,000 i'm giving you the whole thing for free 11 files, every line of code, exit rules, copy-paste config paste it into Grok Bot and it builds itself here's exactly how i built it in the article ↓
2
1
14
362
I THINKED FOR A LONG TIME AND I DID IT. I PUT THE LAST DOLLARS IN THE GROK BOT, I'M ALREADY SCARED my Grok bot's run took $124 to $3,122.51 overnight. 25.2x hit Run at 23:49, 520 smart wallets synced, went to bed 6 bots, one job each → Scout watches 520 smart wallets → Filter kills thin liquidity, fat dev bags, top 10 over 40%, bundled buys → Whale checks if it's real money → Judge says YES or NO with a score → Sniper copies a small slice → Guard puts the stop in first 10,940 signals. 39 copied. 27 green, 12 red 02:48, Judge said NO to $SOLA at 0.80. the wallet it was copying had already sold. that's exactly the one i'd have aped 03:12, Guard closes $MOLT for +$690. that's where the curve bends, watch 0:09 $124 → $392.23 → $1,302.96 → $2,021.84 → $2,788.87 → $3,122.51 the part i trust isn't the green. it's the 10,901 signals it said no to be honest, at 3am with 12 red trades on the board, who keeps going, you or the bot?
people sell this in private alpha groups for $5,000 i'm giving you the whole thing for free 11 files, every line of code, exit rules, copy-paste config paste it into Grok Bot and it builds itself here's exactly how i built it in the article ↓
19
577