ex: @claudeai | independent r̗̝esearcher on frontier ai | dm open
Pacific Heights, SF
Joined July 2014
- Tweets25.2K
- Following442
- Followers13.1K
- Likes44.7K
plusxbt retweeted
I don't understand why everyone isn't doing this yet: Anthropic's own Claude Code guide lists 7 habits you can stop
1. Stop typing think hard, it's plain text, effort is the control: set /effort or type ultrathink for one turn
2. Delete IMPORTANT and ALL CAPS, shouted rules get over-applied, turn hard rules into a hook or deny rule
3. Cut architecture and file trees from CLAUDE.md, Claude reads the repo, keep only what isn't in there
4. Drop keep going and don't be lazy, it's more likely to cause over-work, give Claude a check to pass, /goal for unattended runs
5. Quit spawning subagents just to save context, most models have a 1M window, keep them for parallel work
6. Stop handing a task to a subagent for a better answer, a fresh reviewer helps because it hasn't seen your reasoning
7. Stop rationing context, it compacts on its own, still /clear between tasks for better, cheaper answers
Opus 5.5 runs at medium effort by default, /effort status shows yours
In the terminal ultrathink turns rainbow when it can raise that turn to xhigh, v2.1.311+
Full guide in the article below
plusxbt retweeted
It's pretty wild that people still run Claude Code bare when these 7 repos have over 1,000,000 GitHub stars combined
1. Load 4 coding rules drawn from Karpathy's post (230K): /plugin marketplace add forrestchang/andrej-karpathy-skills
2. Make Claude write the least code that works (186K): /plugin marketplace add DietrichGebert/ponytail
3. Add 7 design skills with 192 palettes (146K): /plugin marketplace add nextlevelbuilder/ui-ux-pro-max-skill
4. Run Garry Tan's own setup, /office-hours to /ship (141K): paste the install line from github.com/garrytan/gstack, needs Bun
5. Cut Claude's replies to caveman length, code stays exact (118K): /plugin marketplace add JuliusBrussee/caveman
6. Add Addy Osmani's 25 skills, spec to ship (114K): /plugin marketplace add addyosmani/agent-skills
7. Shrink the shell output Claude reads (86K): brew install rtk, then rtk init -g
After each marketplace add, run /plugin and install it from the Discover tab
gstack adds a Stop hook and Codex reviews of your diffs, rtk patches settings.json: read both READMEs first
Full guide in the article below
plusxbt retweeted
Everyone is obsessed with making AI smarter. Almost nobody is talking about the fact that their AI keeps forgetting everything they teach it.
Claude Code can become something much more interesting than a coding assistant.
A second brain that remembers your projects, decisions, people and notes across sessions.
Here's how to build it:
1. Create a folder called SECOND BRAIN
2. Open Claude Code inside it
3. Add four folders: MEMORY/, notes/, people/, projects/
4. Create a CLAUDE.md that tells Claude how to organize and retrieve important information
5. Use @-imports to load your most important memory files at startup
6. Tell Claude to save key decisions as Markdown files, one topic per file
7. Create a custom /brain skill to search your notes and bring relevant context back
Now watch the video.
Every branch represents stored knowledge. The glowing line visualizes a memory being retrieved.
Something you decided weeks ago can become useful context in a completely new session.
No vector database. No expensive memory app. Just Markdown files and Claude Code.
The crazy part? Your knowledge doesn't have to disappear when the conversation ends.
Full SECOND BRAIN setup in the article below.
plusxbt retweeted
I still don't get why everyone isn't fixing this: Claude Code silently deletes your chats after 30 days, one line changes it
1. Open ~/.claude/settings.json and add "cleanupPeriodDays": 3650, 0 fails validation
2. Keep secrets out, transcripts sit in ~/.claude/projects as plain text, check the size: du -sh ~/.claude/projects
3. Name work while you still remember it: /rename auth-fix, later in the same repo claude --resume auth-fix
4. Query an old chat without touching it: claude -p --resume <session-id> --fork-session "list the decisions we made" >> decisions.md
5. Export the open chat: /export auth-fix.txt
6. Ask Claude to add a SessionEnd hook that copies transcript_path into your second brain folder, every chat lands there when it exits normally
7. Run /insights: an HTML report on what you work on and where you hit friction, it spends your tokens
Plans, checkpoints, subagent logs and /insights reports go in the same sweep, no message shown
Transcripts of sessions started or last continued in the desktop app or Cowork are exempt by default
Full guide in the article below
plusxbt retweeted
It is becoming f*cking dangerous: one command in Claude Code turns your codebase into a map, 141K stars in 7 months
1. Install with uv, the package has two y's: uv tool install --python 3.13 graphifyy
2. Add the skill to Claude Code: graphify install
3. List tests, docs and images in .graphifyignore: test factories took 6 of the top 10 spots in a repo Graphify reviewed, each image gets its own subagent
4. In a repo you trust, start Claude Code at the root, type /graphify . (tree-sitter parses code with no LLM, Claude still spends tokens running it)
5. Open graphify-out/graph.html and read GRAPH_REPORT.md: most connected nodes first, then surprising connections
6. Ask with words from your code: graphify query "auth token", or trace a path: graphify path "A" "B" --undirected
7. Remind Claude to check the graph: graphify claude install, rebuild the code part on every commit: graphify hook install
Every edge is tagged EXTRACTED, INFERRED or AMBIGUOUS: check the source before you trust the last two
Official repo: github.com/Graphify-Labs/gra…
Full guide in the article below
plusxbt retweeted
It is hard to believe, but 66,000 GitHub stars went to a way of running Claude Code without paying Anthropic
1. Open github.com/Alishahryar1/free…, on Mac or Linux run the curl installer from the README
2. Run fcc-server on Linux or open the Free Claude Code app on Mac and Windows, it opens the Admin UI
3. Turn on Proxy Authentication with your own token, by default the proxy is open to your whole network
4. Create an API key at build.nvidia.com, paste it under Providers, NVIDIA NIM, Add key, Save, then click Apply
5. Start the agent with fcc-claude
6. Switch in the native /model picker among providers you set up, 61 in the catalog, free and paid, DeepSeek, Kimi, Groq, Ollama
7. Add Fallback Models so a dead provider rolls to the next one, a failed request can consume usage from more than one
Free routes mean other models, not Claude, Anthropic does not support this, and your code goes to the provider unless it is local
The README claims 1.3B+ free tokens a month, availability and limits are set by each provider and can change
Full guide in the article below
plusxbt retweeted
Whoever built this will put video editors out of work: raw footage in a folder, Claude Code hands back final.mp4
1. Paste the setup prompt from github.com/browser-use/video… into Claude Code, it installs itself
2. Export your ElevenLabs key first, the tool is free and open source, you pay for transcription and Claude usage
3. Drop your raw takes into one folder, cd there and run claude, transcription starts right away so test on one short clip
4. Type: edit these into a launch video
5. Read the plan it proposes and say OK, it does not touch the cut before that
6. Wait while it removes umms and dead space, ask for color grade, subtitles and overlays too, each overlay gets its own subagent
7. Approve the preview to get edit/final.mp4, every cut is checked first, up to 3 fix passes
30K+ GitHub stars, MIT license, works for talking heads, montages, tutorials, travel, interviews
Claude reads a ~12KB transcript plus PNG filmstrips, frame by frame would be about 45M tokens
Full guide in the article below
plusxbt retweeted
I still don't get why anyone builds a second brain for Claude Code from scratch when these 7 repos have over 75,000 stars combined
1. Teach Claude Obsidian's formats and tools with 6 skills from its CEO: github.com/kepano/obsidian-s…
2. Run wiki-ingest on text in inbox/, review the plan, get linked notes: github.com/AgriciDaniel/clau…
3. Save decisions and tasks with obsidian-save, load your current state next session with obsidian-world: github.com/eugeniughelbur/ob…
4. Run an unofficial Karpathy-style wiki with ingest, query and lint: github.com/Astro-Han/karpath…
5. Copy a starter vault plus 18 skills and 72 commands: github.com/undefined-ui/seco…
6. Turn your ChatGPT and Claude exports into linked notes, run only step 1 of its README: github.com/charlie947/ai-sec…
7. Split auto-loaded memory from wiki pages read on demand, ignore its tip to keep secrets in memory: github.com/MehmetGoekce/llm-…
Pick one, they overlap: all seven work with plain Markdown you can open in any editor
Full guide in the article below
plusxbt retweeted
It is becoming f*cking dangerous: one line in Claude Code builds a turbofan with 3,597 parts you can cut open and break
1. Install the open source skill (Node 22.20+): npx skills add AirsupHQ/machine-skill -g -a claude-code
2. Add Playwright in its folder: cd ~/.claude/skills/machine && npm i && npx playwright install chromium
3. Open a new Claude Code session in an empty folder and type: /machine a turbofan jet engine, the kind that pushes a widebody off the runway
4. Let Claude write down real dimensions, parts list and physics, then build it
5. Open the HTML file, drag the throttle, try explode, cutaway, x-ray
6. Hit the break button: it throws a fan blade and the solver computes the load the case has to hold
7. Film it from that folder with your file's name (needs ffmpeg): node ~/.claude/skills/machine/scripts/record.mjs turbofan.html out/turbofan.mp4
The makers' estimate: about an hour per machine
You get one HTML file: the makers' demo counts 3,597 parts, nearly all blades, yours will differ, and the physics is a small model, not CFD
Full guide in the article below
plusxbt retweeted
It is becoming f*cking dangerous: Claude Code turns into a web crawler for almost any website, and you write zero lines of code
1. Install Microsoft's Playwright CLI (needs Node.js 20+): npm install -g @playwright/cli@latest
2. Run playwright-cli install --skills in your project folder, it adds a browser skill to .claude/skills
3. Open Claude Code in that folder, paste the start URLs and add: use playwright-cli
4. Name the fields you need: titles, prices, links, dates
5. Tell it to follow internal links, skip duplicates and respect robots.txt
6. Open a second terminal tab and run playwright-cli show to watch the browser live
7. Ask Claude to save the results to data.json, then turn the run into a script and test it once so you can rerun it
Even Microsoft's Playwright MCP README says coding agents increasingly favor CLI with skills over MCP because it is more token-efficient
Full guide in the article below
plusxbt retweeted
This is getting out of hand: Karpathy shared one text file, and Claude Code turns your PDFs into a wiki it maintains for you
1. Make a folder and open Claude Code inside it
2. Create raw/ for your sources and wiki/ for the pages
3. Paste Karpathy's LLM Wiki idea file and ask Claude to write CLAUDE.md for your topic
4. Put a PDF, transcript or article in raw/
5. Tell Claude to ingest it: a summary page, index.md and related pages updated, a new entry in log.md
6. Ask anything: it answers from wiki/ with citations, good answers can be saved as new pages
7. Ask for a lint pass now and then: contradictions, stale claims, orphan pages
One source might touch 10-15 pages, so your knowledge is compiled once and kept current, not re-derived on every question
Full guide in the article below
plusxbt retweeted
You are making a painful mistake if you still use Claude Code's default answers: one Markdown file rewrites how it talks to you
1. Type /output-style to see the four built-in styles besides Default
2. Run /output-style concise: every answer leads with the result, no preamble
3. Switch to explanatory: short ★ Insight blocks show why Claude wrote that code
4. Pick learning: Claude leaves TODO(human) spots for you to write yourself
5. Create your own: save diagrams-first.md in ~/.claude/output-styles with your rules
6. Add keep-coding-instructions: true so Claude still codes the same way
7. Restart Claude Code, run /output-style diagrams-first and ask it to explain your app
It sets Claude's role, tone and format for every response in the session, so you stop repeating the same request in each prompt
Full guide in the article below
plusxbt retweeted
It is becoming f*cking dangerous: the creator of Claude Code says one habit makes the final result 2-3x better
1. Give Claude a way to check its own work: tests, a bash command, a browser
2. Say "Prove to me this works" and have Claude diff the behavior of main vs your feature branch
3. Spin up a second Claude to review the plan as a staff engineer
4. End every correction with "Update your CLAUDE.md so you don't make that mistake again"
5. Check CLAUDE.md into git so the whole team shares the lessons
6. Set an agent Stop hook on long tasks, it runs the verification for you
Boris calls it probably the most important thing for great results in Claude Code
His own setup: Claude opens a browser, tests the UI and keeps iterating until the code works
Full guide in the article below
plusxbt retweeted
I don't understand why everyone isn't using this yet: Claude Code can audit your CLAUDE.md and catch rules meant for older models
1. Update Claude Code to v2.1.283 or newer: claude update
2. Open your project and run /doctor prompt-audit
3. It reads CLAUDE.md, AGENTS.md, rules, skills and commands in .claude/ and ~/.claude/
4. It flags instructions aimed at older models
5. It finds references to files and commands that don't exist
6. It catches files that contradict each other
7. You get a report with proposed edits, nothing changes until you ask Claude to apply them
One folder only: /doctor prompt-audit .claude/skills/deploy
Anthropic's docs: keep each CLAUDE.md under 200 lines, and if two rules contradict each other, Claude may pick one arbitrarily
Full CLAUDE.md and folder setup in the article below
plusxbt retweeted
It is becoming f*cking dangerous: one prompt turns Claude Code into a FULL crypto trading agent that switches tactics when the market flips
Zero code. Copy, paste, fill in exchange, pairs, timeframe, capital:
Time: 1 min to paste
Action: drop into Claude Code
One-strategy bots trade a flat day and a dump the same way. This one does not
Regime Detector: labels each bar SIDEWAYS, CHOPPY or DUMP via ADX+ATR, swaps playbook
Kill Switch: equity falls 15% below peak, positions close, halt survives restarts
Backtest Gate: 3+ years with fees, slippage and unseen data, blocks any strategy that lost or fell >15%
Nightly Review: writes a tweaks report overnight, nothing changes without your OK
Paper Mode: sim orders on live prices, no real money until you set LIVE=true
Old way: code and test every rule by hand
Now: one prompt, Claude Code builds the bot, tests risk, shows diff before running
Passed backtest = survived the past, not a profit guarantee. Stay on paper until you trust it
Full prompt in the reply below, bookmark this
plusxbt retweeted
THIS IS AN ABSOLUTE GOLDMINE
Anthropic documented one of the strongest tricks in Claude Code
The idea: Sonnet 5.5, Haiku 5.5 and Opus 5.5 in one setup, and each model only gets the work it's BEST at
Opus stays on the hard calls, and it typically costs LESS than running Opus the whole time
One command starts it:
claude --model sonnet --advisor opus
Who does what:
> Sonnet 5.5 drives the main session and takes the routine work: writing and fixing code
> Haiku 5.5 runs as read-only subagents (search, grep, docs), each in its own context window, sending back only a summary
> Opus 5.5 is the advisor, called at decision points, and it sees the FULL conversation including every tool call and result
> Opus isn't called every turn, so routine turns run at Sonnet prices, not Opus prices
Claude decides when to call Opus, and it TENDS to do it at three moments:
> Before it commits to an approach
> When an error keeps coming back
> Before it calls a task done
Sonnet codes. Haiku scouts. Opus advises
The math: per token Opus costs 2x Sonnet and 40x Haiku, so every routine job kept off Opus runs at half the price or less
The full tree:
Sonnet 5.5: main session, writes and edits code
Search: read-only Haiku 5.5 subagent that maps the repo
Grep: read-only Haiku 5.5 subagent that finds call sites
Docs: read-only Haiku 5.5 subagent that reads API docs
Opus 5.5: advisor that sees the whole conversation and every tool result when called
Paste the tree plus this prompt into Claude Code
Set up my Claude Code project to match this tree:
1. Make Sonnet 5.5 the main model and Opus 5.5 the advisor (advisorModel in settings)
2. Add three read-only subagents in .claude/agents/ (model: haiku, tools: Read, Grep, Glob): search (repo map), grep (call sites), docs (API docs)
3. Add one rule to CLAUDE.md: consult the advisor before committing to an approach, when an error repeats, and before calling a task done
Before touching anything, show me the full diff and wait for my OK
Bookmark this setup, you'll want it every session
Step 3 lives in CLAUDE.md, the full CLAUDE.md guide is in the article below
plusxbt retweeted
I don't understand why everyone isn't doing this yet
Anthropic documented a Claude Code feature where the STRONGEST model only kicks in at key moments instead of running the whole time
The setup: Agent Teams in Claude Code
Rather than one session doing everything, the work splits across a team:
Lead (Opus 5.5, high effort): the main session, creates the tasks and runs the team
Teammates (Sonnet 5.5, medium effort): each owns its own files, UX or backend
Advisor (Fable 5.1): consulted before picking an approach, on a repeated error, before done
Shared task list: teammates claim tasks, file locks stop two of them grabbing the same one
Start the lead from the terminal:
claude --advisor fable
Then tell it: Spawn three teammates, one on UX, one on the backend, one playing devil's advocate. Use Sonnet for each teammate
Paste this to set up your workspace:
In settings.json, add CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 to the env block, set effortLevel to high
Define roles in .claude/agents/ with model sonnet and effort medium, plus a devil's advocate role
Give each teammate its own set of files to avoid overwrites
Add a TaskCompleted hook that blocks a task from being marked done if checks fail
Show me every config change before applying it
The CLAUDE.md map your agents read, full guide in the article below
> code.claude.com/docs/en/agen…
plusxbt retweeted
Build an absolute monster. Merge six Claude Code pieces into one system, it works the NIGHT SHIFT while you sleep
Opus 5.5 + subagents + routines + web search + Playwright + CLAUDE.md
Six pieces. One system. Nobody at the keyboard
Opus 5.5 -> the main model that plans every task
Subagents -> independent workers in .claude/agents/, each with its own role and context, returning only a summary
Routines -> scheduled cloud agents on a cron, the system runs at night while your laptop is closed
Web search -> scans public pages and forums for real complaints
Playwright -> opens a real browser, clicks through every screen, catches bugs
CLAUDE.md + memory files -> the map the agent reads on launch, every step and cost logged
Give it one goal before bed: find a real gap people complain about and have a working prototype by 7 AM
One hard rule: nothing spent, nothing published, nobody contacted without your yes
Start the routine at 11:40 PM and go to sleep. Here is how one night can look:
12:10 AM: web search pulls ~1,900 public posts with the same complaint, gig drivers tracking income across 3 apps
1:00 AM: it maps the paid tools people already use and the 2 complaints they keep repeating
2:30 AM: Opus 5.5 plans 9 screens and one database without asking a question
4:10 AM: subagents write the code and a landing page draft
5:45 AM: Playwright clicks through every screen and logs bugs
7:20 AM: the memory log holds every step
Morning: a working prototype and a landing draft waiting for your yes, $0 spent beyond tokens, nothing published
A subagent with no payment or email tool physically CANNOT buy or send anything. The rules in CLAUDE.md apply to every session
The biggest mistake: treating AI like a chat for answers. Give yours a GOAL, a DEADLINE and HARD RULES
You write zero code. Playwright does the clicking. You're asleep the whole time
Everyone is waiting for AGI. The night shift already exists
Bookmark this. Full breakdown of the CLAUDE.md map in the article below
plusxbt retweeted
It is becoming f*cking dangerous: when Anthropic shipped CLAUDE.md, they didn't even realize they were opening a Pandora's box
1. Make an empty folder on your computer, call it SECOND BRAIN
2. Open Claude Code inside it
3. Add subfolders: MEMORY/, notes/, people/, projects/
4. Drop a CLAUDE.md in the root with one rule: save every key fact to MEMORY, one topic per file
5. Add @-links in CLAUDE.md so key files load into context every session
6. Run claude mcp add to connect Playwright MCP, now Claude controls a real browser
7. Build a skill in .claude/skills/ so one command runs the entire routine
Folder setup: about 5 minutes
Your Claude becomes a second brain: it keeps every key fact between sessions, loads your main files by itself and drives a real browser
Full guide in the article below