Aayush Manchanda · Library

Cursor's prompt for improving your agent harness

Thu, 24 Sep 2026 00:00:00 GMT

Cursor shares a prompt shaped by lessons from improving its agent harness.

Source: x.com

Make money, make no mistakes.

Thu, 24 Sep 2026 00:00:00 GMT

BugBasher used public restaurant inspection data for cold outreach and earned $75 from pest-control referrals in four weeks.

The idea: to build an agentic exterminator service. New York City publishes the result of every restaurant health inspection, including pest violations.

Four weeks and a lot of phone calls later, it had made $75.

Source: x.com

Eight mistakes people make coding with AI

Thu, 24 Sep 2026 00:00:00 GMT

Eight recurring AI coding mistakes include over-orchestration, preserving stale context, tool sprawl, code-quality anxiety, and unnecessary branch or worktree habits.

Source: x.com

How I make launch videos with AI

Thu, 24 Sep 2026 00:00:00 GMT

Jake Moran turns product context into story angles, approved still-frame storyboards, then reusable HTML, CSS, and JavaScript launch-video scenes with HyperFrames.

Then I ask the agent to build one still frame per scene and lay them out in a grid - a contact sheet, basically a page of thumbnails, one for each row of the table.

The rule I push hardest: everything on screen is real UI. Anyone who uses the product knows what it looks like, and an invented button breaks it for all of them.

A HyperFrames video is a folder of HTML files, so once a scene works, you can reuse it in the next video.

Source: x.com

How I built a $300K ARR consumer app in 8 months

Thu, 24 Sep 2026 00:00:00 GMT

Akshaya details how he validated, built, and marketed a healthcare app to $300K ARR in eight months with a small AI-assisted team.

Now that AI has made it so easy to build an app, you should be spending as much time figuring the right things to build as you are actually building it.

Before you even start building anything, start a TikTok & Instagram account and start posting founder-POV content talking about your app.

Make a Posthog dashboard to track your entire funnel: app install → go through onboarding questions → paywall conversion rates → app usage → retention.

Source: x.com

Inside Cursor

Thu, 24 Sep 2026 00:00:00 GMT

Brie Wolfson reports on Cursor’s in-person culture, talent-first recruiting, deep-work norms, dogfooding, constructive friction, and focus on raising expert developers’ ceiling.

Michael often says he wants Cursor to be a “haven for self-motivated individual contributors.” And so far, it is.

Because Cursor team members are Cursor’s ideal users, everyone is looking for product-market fit internally on the features and infrastructure they believe in.

At Cursor, critics are also problem-solvers. The “friction” here works because everyone genuinely wants the best for the product and each other.

Source: colossus.com

Writing Rust code that's faster than state-of-the-art…

Thu, 24 Sep 2026 00:00:00 GMT

Max Woolf shows agents repeatedly optimizing Rust benchmarks by 2× to 20× with baselines, explicit targets, correctness gates, and anti-cheating constraints.

After months of testing and experimenting since the release of Opus 4.5 made agentic coding more viable, I can confidently confirm that modern agentic LLMs can indeed write Rust code that is significantly faster than current state-of-the-art approaches if given appropriate guardrails and constraints.

What I needed was a clear target goal that can be pass/failed, so I refined the prompt:

It must be reiterated that agents can and will cheat if they can.

Source: minimaxir.com

An All-Night Client Design

Thu, 24 Sep 2026 00:00:00 GMT

Harshit shares a client design made overnight that the client loved.

Source: x.com

AI Will Expand the Creative Canvas

Thu, 24 Sep 2026 00:00:00 GMT

Jeffrey Katzenberg argues AI will expand creative work, but fair adoption requires artists' consent, credit, compensation, and leadership.

Tools are never the point. The instruments change with every generation. What endures is taste and imagination.

The right path forward includes them by design, with credit, with consent, and with compensation. Build this with the storytellers. Not on top of them.

The tools-versus-no-tools argument is a trap. First, we must all agree that there should be terms.

Source: x.com

How Volatile Are AI Responses? (70K Answer, 28 Day Study)

Thu, 24 Sep 2026 00:00:00 GMT

Across 70,000 responses, leading brands stayed visible most days while exact brand and citation sets changed constantly, making single-check AI visibility misleading.

On ChatGPT, the leading brand appeared on 82% of days.

The individual pages cited change constantly. Only about a quarter of the pages ChatGPT cited for a prompt were cited again the next day.

Source: detailed.com

Ship Your First App With Claude Code

Thu, 24 Sep 2026 00:00:00 GMT

JJ Englert lays out a beginner workflow for shipping an app with Claude Code, GitHub, Vercel, upfront planning, user mapping, and design.

Source: x.com

claude.dev — Technical Writing for People Building with…

Thu, 24 Sep 2026 00:00:00 GMT

Anthropic's developer blog collects practical engineering notes, playbooks, and videos on building with Claude, including context engineering, skills, caching, and tool design.

Source: claude.dev

Thinking in Systems, Shipping in Loops

Thu, 24 Sep 2026 00:00:00 GMT

Tomasz Tunguz argues software engineers now design agent systems, constraints, and verification loops rather than hand-writing every line of code.

This is the new software engineering : designing the systems that let AI write code correctly at scale.

Three properties define great system design : resilience, self-organization & hierarchy.

Self-organization is the loop learning, so a class of failure becomes a new skill the agent applies next time.

Source: x.com

Build and Automate With Claude

Thu, 24 Sep 2026 00:00:00 GMT

A one-hour course on using Claude to build and automate projects.

Source: x.com

Recreating 1906 Market Street in Blender

Wed, 23 Sep 2026 00:00:00 GMT

Alex Albert shares a Claude Tag prompt rebuilding 1906 pre-earthquake Market Street in Blender from Sanborn maps and period film, generating every asset in Python.

Record every building with footprint, height, facade material, occupant, and the source for each fact with a confidence level.

No downloaded meshes, textures or HDRIs.

Source: x.com

Brooke Joseph on joining OpenAI through an acquired startup

Wed, 23 Sep 2026 00:00:00 GMT

Julia Fedorin interviews Brooke Joseph, who left Waterloo without graduating, interned at 12-person Sky, joined OpenAI when it acquired Sky, and researches computer-use agents.

Source: x.com

Polar for GTM turns an ICP into saved outbound workflows

Wed, 23 Sep 2026 00:00:00 GMT

Polar for GTM takes an ICP, finds accounts and decision makers, enriches leads, updates sheets, preps follow-ups, and saves the workflow to rerun.

Source: x.com

Build a full-stack app with coding agents

Tue, 22 Sep 2026 00:00:00 GMT

Kun Chen builds and ships a full-stack app from scratch with a FirstMate-led team of coding agents, covering parallel research, UX prototyping, and end-to-end testing.

1:01 FirstMate: he talks to one agent, which hands work to crewmate agents, so he isn't juggling tabs and context.

30:58 The model's VPS estimate of about $30 a month was wrong. He pushed it toward a roughly $5 Hetzner box.

1:00:58 Secrets live in Atomic Vault. He approves every command that touches a production key instead of using .env files.

1:22:44 GPT models, especially the 5.6, has this tendency to over-engineer.

Source: youtube.com

How I shipped 2,500 PRs last month

Tue, 22 Sep 2026 00:00:00 GMT

Lauren (poteto) shares a talk, originally meant for Cursor Compile London, on shipping 2,500 PRs to production last month, releasing it free on X instead.

Source: x.com

Automating repetitive work at OpenAI with Codex

Tue, 22 Sep 2026 00:00:00 GMT

Jeremy Lewi describes using Codex and Runme notebooks to plan, execute and document repetitive OpenAI engineering work like Kubernetes cluster setup and model evaluations.

Codex reads the notebook and updates it as the work progresses.

The result is a notebook that documents the steps required to complete the task, along with the dead ends.

Runme brings the intent, actions, decisions, and outcomes into the same artifact.

I still decide when a plan is ready and when a consequential choice needs human judgment.

Source: developers.openai.com

How to Pivot Your SaaS Company in the Age of AI

Tue, 22 Sep 2026 00:00:00 GMT

Jon Yongfook argues cheap AI-built features killed SaaS moats, then lists seven ways founders can still compete: operations, distribution, agents, underserved markets, hardware, pricing, consolidation.

Competing on features is over, because anyone can build any feature in a weekend now.

The hard part is the deliverability relationships, the compliance, the sheer boredom of keeping the thing up.

If two products have identical features, and they will, the one with the audience wins.

Own the market, then sell them the things that aren't commodities yet.

Source: bannerbear.com

Find your writing voice by copying others' style

Tue, 22 Sep 2026 00:00:00 GMT

Jason Cohen argues that copying other writers' style, not their words, is a practical way to discover your own voice and strengths as a writer.

High School English III doesn't teach you how to write prose in the real world any more than a college CS degree teaches you how to write code in the real world.

Rather, copying is one of the best ways of learning, growing, evaluating, and exploring. It's a valid tool so long as we regard it as a means to an end.

The goal is to learn and grow as a writer, and this was fun to do.

Copying ironically helps you discover yourself: The way to discovering who you are, what you value, what you're good at, what you believe, is to try things on, like clothes, to see what fits.

Source: longform.asmartbear.com

Short design engineering tips on type and motion

Fri, 18 Sep 2026 00:00:00 GMT

Gustavo Fior's Craft collects short essays and one-line tips on typography, color, layout, motion, sound and data for design engineers, most still marked coming soon.

How fast it feels is a design choice.

Great work starts with great inputs.

Tighten big text, loosen small text.

Ease-out for anything the user triggers.

Source: craft.gustavofior.com

Historical awe for the everyday things at home

Tue, 15 Sep 2026 00:00:00 GMT

Ordinary Abundance pairs a modern apartment's daily objects, music, light, photography, medicine, plumbing, with historical quotes describing each as a miracle when new.

If we could have devised an arrangement for providing everybody with music in their homes, perfect in quality, unlimited in quantity, suited to every mood, and beginning and ceasing at will, we should have considered the limit of human felicity already attained.

The strange, weird light, exceeded in power only by the sun, yet mild as moonlight, rendered the Court House square as light as midday.

A museum without walls has been opened to us, and it will carry infinitely farther that limited revelation of the world of art which the real museums offer us within their walls.

You have erased from the calendar of human afflictions one of its greatest.

Source: ordinaryabundance.com

Simon Willison's LLM cliché highlighter hits 38 patterns

Mon, 14 Sep 2026 00:00:00 GMT

Simon Willison expands his LLM cliché highlighter tool to 38 detected patterns for spotting AI-generated writing tells.

Source: x.com

Cut your AGENTS.md down

Mon, 14 Sep 2026 00:00:00 GMT

PostHog's Jina Yoon argues AGENTS.md needs subtraction not addition, and gives 3 methods: doctor checks, testable evals, and direct agent feedback.

Boris Cherny, the creator of Claude Code, recommended deleting your CLAUDE.md every six months to stay on the bleeding edge.

Our agents therefore had wrong instructions for 21 hours. During that time, one engineer’s PR got stuck for 10 hours; another lost 45 minutes investigating the cause before updating the instructions.

it suggested turning off 3 unused plugins and 3 skills across our website repo and estimated saving an average of 6K tokens per session.

For each line, if you can’t name the failure it prevents, delete it.

Source: x.com

An illustrator's code for playful homepage effects

Mon, 14 Sep 2026 00:00:00 GMT

Illustrator Noah Farrar breaks down the code behind his homepage's interactive moments: dither reveals, spring-physics heart trails, and ball physics, with exact numbers.

Dots use spring physics (stiffness 0.02, damping 0.9) and fade in with a trailing effect over your last 12 cursor positions, creating a comet effect.

Ten SVG circles with real physics. Gravity at 0.015px/frame², 0.5% drag per frame, 50% vertical restitution on bounce.

Source: noahfarrar.com

Use your code-based design system with Magic Patterns

Mon, 14 Sep 2026 00:00:00 GMT

A Lenny's Workshop session covers connecting Magic Patterns to Claude Code and Cursor over MCP so agents produce on-brand, production-synced components.

We’ll cover pulling generated UI straight into your codebase, reading and writing your design system over MCP so agents produce on-brand components, and keeping prototypes and production code in sync without copy-paste.

Source: lennysworkshops.com

Get less generic design out of AI

Mon, 14 Sep 2026 00:00:00 GMT

Anshu Chimala, ex-Apple AI design lead, outlines a Discover-Define-Deliver process pushing AI past generic outputs using seed strings, ambitious prompts, and subagent design critics.

Large language models are next-token predictors: at each step, they look at a sequence of text and predict what comes next based on millions of examples.

Great design, on the other hand, starts with feeling and aims to create an emotional response. It bends the rules and delights users with memorable, unexpected choices.

One technique for this is String Seed of Thought, published by Sakana AI. We make the AI generate a random string and use it as design inspiration.

AI loves to add more, but it rarely takes away. One of the biggest signs that a design is AI-generated is that it overexplains everything or contains elements that don’t serve any practical purpose.

Source: lennysnewsletter.com

Claude Code builds a Blender restaurant scene unattended

Mon, 14 Sep 2026 00:00:00 GMT

space-z.ai's brief tasks Claude Code with building a photorealistic restaurant and bar unattended for 12+ hours, judged by a 100-point rubric and six fresh-context critics.

You are the lead environment artist, Blender technical artist, hospitality interior architect and restaurant operations designer, lighting artist, and final quality owner for this project.

The user may be unavailable. Make conservative, reversible assumptions, record them, and continue.

Perform at least four complete cycles of: build or change → render the fixed judgeset → inspect actual pixels and contact sheet → diagnose → fix → rerender identical cameras

Each important asset must work at three reading distances: a recognizable macro silhouette, functional medium-scale construction, and close-range joins, controls, hardware, fasteners, seams, gaskets, labels, edge treatment, and contact with adjacent surfaces.

Source: restaurant-bar.space-z.ai

How a Professional Writer Writes With AI

Mon, 14 Sep 2026 00:00:00 GMT

Every's Natalia Quintero interviews Katie Parrott on turning a $20 ChatGPT habit into compound writing, a context-file system now packaged as a plugin.

8:06 Turning AI into a content agency of one.

10:59 Building the context files and style guides that make AI output useful.

32:27 What compounding means, and how it became the Compound Writing plugin.

38:00 Borrowing Vonnegut, Hitchcock and Sorkin as AI editing skills.

Source: youtube.com

Kenny Damian's six systems for scaling client delivery

Mon, 14 Sep 2026 00:00:00 GMT

Kenny Damian details 6 systems that keep AI delivery costs flat as Frontal scales from 2 to 275 client accounts.

You price a result instead of hours, you say software does the work, the client signs.

At Frontal we have done this work across 275 companies and the same thing is true every time: companies buy the tools and cannot make them run.

I open-sourced the 15 skills we run across 70+ clients. 5 pillars, roughly 4,200 lines of methodology.

Across the agency that store now holds 11,000 campaigns, 132,000 snapshots and 30+ clients.

Source: x.com

Building a $1/month scraper to replace ScrapingBee

Mon, 14 Sep 2026 00:00:00 GMT

Pieter Levels replaces his $249/month Scrapingbee plan with a self-built Playwright scraper on his VPS, reaching about 90% success for roughly $1 a month.

plain curl got 7% blocked but OpenSERP's headless browser got 100% blocked

built my own with Playwright and a real Google Chrome on my VPS, headful and persistent context

Now I'm at about 90% success rate, which is similar or better than Scrapingbee

Source: x.com

Inside OpenAI’s Forward Deployed Engineer Role

Mon, 14 Sep 2026 00:00:00 GMT

Colin Jarvis, OpenAI's Global Head of Forward Deployed Engineering, describes an FDE org that grew from 2 to 140 people alongside a 200-person deployment company.

A forward deployed engineer is focused on outcomes. Which means that their goal should be to work on problems where solving them has a much bigger impact than the work itself.

They started with just 2 people and have now grown to around 140 people in the FDE org.

An FDE might spend the full day talking with the payments team to understand what specific things are important to them so that the migration would be successful.

They generally tell new joiners to expect around 50% of their time on site.

Source: newsletter.eng-leadership.com

Stripe's Kai and the company brain playbook

Mon, 14 Sep 2026 00:00:00 GMT

Sharadh Krishnamurthy explains how Stripe's Kai, an internal AI agent used by 10,000+ employees weekly, handles governance, data access, and 2,000 packaged skills.

Sharadh Krishnamurthy is an engineering manager at Stripe, where he helped build Kai, the company’s internal AI agent used by more than 10,000 employees every week.

Why Stripe built Kai from scratch instead of buying, and what tipped the decision

Why “projects” at Stripe are a governance mechanism, not just a folder

What Sharadh learned the hard way when agents nearly took down production systems

Source: lennysnewsletter.com

Design details from Honk's live messaging app

Sat, 05 Sep 2026 00:00:00 GMT

Benji catalogues Honk's real-time messaging interactions: live typing bubbles, presence animations, emoji collisions, custom fonts and sounds, arguing playfulness wasn't enough for product-market fit.

Honk took a completely different approach. It had two chat bubbles, no send button, no chat history, and no drafts.

As you typed, your chat bubble dynamically resized; when you paused, it shrank back down.

Users could send (or more commonly spam) as many emojis as they wanted, in any order, even if this made the conversation briefly unreadable for both people.

Honk facilitated over 2.6 billion characters typed, 170 million Honks sent, and hundreds of millions of images, audio, and videos exchanged.

Source: benji.org

How We Built Interfere’s New Website

Sat, 05 Sep 2026 00:00:00 GMT

Interfere's founders detail a week-long sprint on visual identity, code-built homepage graphics, staggered hero animations, and a stack migration from Next.js to TanStack Start.

They gave us the two signature details the site is built on: the Interfere gradient and a pixelated final word in the headline that echoes the logo.

Every asset on the website is built in code, with a few exceptions where it wasn’t possible.

We originally built the website with Next.js. Since then, we’ve migrated to TanStack Start and Vite for both the marketing pages and the product.

It’s significantly more work this way, but we weren’t willing to compromise or take shortcuts.

Source: interfere.com

You Cannot Mandate an AI Transformation

Fri, 04 Sep 2026 00:00:00 GMT

Julie Zhuo outlines a ten-step path for AI adoption, from automating one hated task through building trust with evals to redesigning what organizations reward.

Build personal belief. Turn belief into shared wins. Turn wins into reliable systems. Then redesign the organization around what has become possible.

A leader cannot mandate enthusiasm. A leader can arrange for relief.

You earn trust in an agent through evaluations, usually shortened to evals. An eval is a repeatable test of whether the AI did the job right.

A win is a result, not a good performance of transformation.

Source: lg.substack.com

Six free courses on packaging and selling ideas

Fri, 04 Sep 2026 00:00:00 GMT

Jack Butcher's Visualize Value offers six free courses on visualizing ideas, productizing services, apprenticing without permission, and using AI as leverage rather than replacement.

You don't need to be a designer.

Service is the leverage AI can't commoditize.

Source: visualizevalue.com

Building Shared Memory for AI Agents in Notion

Wed, 02 Sep 2026 00:00:00 GMT

Notion's Hesham Salman describes Lore, an open-source shared memory system storing agent knowledge as Notion pages, which recovered about 46% of failures models couldn't solve.

Lore is the tool we built for that gap: shared, persistent memory for agents, backed by Notion.

The first bad version of a memory system is a junk drawer with an API.

When we filter the data set only to the hardest tasks (those that the no-memory agent never succeeded in doing), the seeded memory agent recovered about 46% of failures!

A key takeaway is not that memory magically fixes agents.

Source: notion.com

How our agents build on-brand pages with design.md

Tue, 01 Sep 2026 00:00:00 GMT

Vercel describes building design.md, a public brand file for coding agents, cutting a known layout failure from 91 to 39 occurrences across matched test pages.

The naive approach we tried first was to simply port product-design into a public prompt, collapsing the skill's reference files into one file any agent could read from a URL.

The pages generated with design.md had 39 of those failures. The pages generated without it had 91, which works out to 57% fewer in this test.

We created a stylesheet because agents kept inventing their own typography, spacing, and layout, so we took those decisions away from the model entirely.

Source: vercel.com

The Modern App Starter Checklist

Mon, 31 Aug 2026 00:00:00 GMT

burcs.dev lays out a 14-phase Cloudflare app checklist covering domain setup, data stores, staging, auth, SEO, agent surfaces, secrets, and monetization.

Allowlist, don't denylist. Public things live in public/. Nothing is public by accident.

The single most common architecture mistake is one store for everything.

KV is eventually consistent - reads elsewhere can lag a write by up to a minute.

Rule one: don't invent crypto, and don't put tokens in localStorage.

Source: burcs.dev

The Claude Code Guide for Startups

Mon, 31 Aug 2026 00:00:00 GMT

Anthropic profiles more than a dozen startups, citing ClickHouse's 30% more shipped features and Artemis Security's 6,000-plus weekly PRs, to outline five Claude Code principles.

Agentic coding lowers the barrier to entry, so the person who understands the problem can ship the first version of the fix.

"Not only were engineers shipping much more, but non-technical people (like me) were also suddenly shipping UI changes and other product improvements."

You can't automate a process unless you have a reliable means of monitoring and verifying the outcome.

Source: claude.com

Teach coding agents your design taste with rules

Sun, 30 Aug 2026 00:00:00 GMT

Emil Kowalski packages design taste into skill files with strict rules for easing, duration, and typography so agents produce better animations.

The correct animation below feels right, because it animates from a higher initial scale value. It makes the movement feel more gentle, natural, and elegant.

So take a step back, ask yourself why you made a certain decision, articulate clearly why something has to be done this way, set the rules, be strict.

Source: emilkowal.ski

Why accountability wins in AI services

Sat, 29 Aug 2026 00:00:00 GMT

Yaman argues AI-native services firms win on accountability, not automation, citing Sequoia's autopilot framing and a $1 billion opportunity cost from a bad FDA filing.

No one would ever trust anything that has nothing to lose, and AI agents have nothing to lose.

The pilot isn't there for the 10,000 normal flights; he's there for the abnormal one.

Sell the capabilities of an agent behind the accountability of a credible human

Source: x.com

Tell your agent to mine your conversations

Fri, 28 Aug 2026 00:00:00 GMT

Aaron Francis suggests telling your agent to mine your saved conversations and report how you work, your strengths, weaknesses, and how to use AI better.

Source: x.com

Orwell's rules for writing with LLMs

Fri, 28 Aug 2026 00:00:00 GMT

Joe Wyer shares Orwell's six writing rules as a fix for LLM prose, including avoiding cliche metaphors, cutting unnecessary words, and preferring active voice.

They write with so much damn passive voice, incoherent figures of speech, and unnecessary -ly words.

If it is possible to cut a word out, always cut it out.

Never use the passive where you can use the active.

Source: x.com

Learn faster than the default pace allows

Fri, 28 Aug 2026 00:00:00 GMT

Derek Sivers recounts how Kimo Williams taught him a full semester of music theory in one three-hour lesson, compressing years of coursework into five sessions.

I graduated from Berklee and taught there, too. I'll bet I can teach you two years of theory and arranging in only a few lessons. I suspect you can graduate in two years if you understand there's no speed limit. Come by my studio at 9:00 tomorrow for your first lesson, if you're interested. No charge.

In our three-hour lesson that morning, he taught me a full semester of Berklee's harmony courses. In our next four lessons, he taught me the next four semesters of harmony and arranging classes.

If you're more driven than most people, you can do way more than anyone expects. And this principle applies to all of life, not just school.

Source: sive.rs

The Anatomy of an Agent Harness

Thu, 27 Aug 2026 00:00:00 GMT

Vivek Trivedy of LangChain defines agent as model plus harness, then walks through core components: filesystems, bash execution, sandboxes, memory, and context-rot mitigation.

Parked. I already run this ladder: harness first, loop second, graph only when branching earns it.

If you are building your first harness, this is the vocabulary sheet. For me it names things I already do, which is worth one read and no more.

A harness is every piece of code, configuration, and execution logic that isn't the model itself. A raw model is not an agent. But it becomes one when a harness gives it things like state, tool execution, feedback loops, and enforceable constraints.

The model contains the intelligence and the harness is the system that makes that intelligence useful.

Compaction addresses what to do when the context window is close to filling up. Without compaction, what happens when a conversation exceeds the context window? One option is that the API errors, that's not good.

Opus 4.6 in Claude Code scores far below Opus 4.6 in other harnesses.

Source: langchain.com

See how Gumroad's AI operations agent runs daily

Thu, 27 Aug 2026 00:00:00 GMT

Gumclaw, Gumroad's AI operations agent, runs a cron-triggered loop: check policy, verify facts against live sources, do the work, then report to founder Sahil.

Read now: this is the setup my own agents already half-run, written down end to end.

Two rules are worth stealing straight into my stack: verify against a live source before every public claim, and turn every real mistake into a dated policy rule.

Finance ops: maintain a daily finance-event ledger from processor reports and prepare month-end journal entries for review.

Public presence: reply on X autonomously within policy; standalone tweets are drafted first and approved by Sahil.

Source: gumclaw.github.io