Francesco

i build open-source and make it profitable while sharing what i learn

- run X on autopilot
- agent skills for 10x devs

[Francesco @prafaolo]

I want to grow and market on X, but I'd rather focus on building the business

I want to post more, but it takes up a lot of time. I feel like I should use AI to write faster, but nobody enjoys reading AI slop

I solved this and open-sourced it. Here's the sauce 👇

[Francesco @prafaolo]

Following X hook best practices can make your post easier to scroll past

You can use the psychology without sounding like a parrot

1. Pick one to three levers that fit your point

Payoff: Give the reader something concrete to gain
Curiosity: Open a question they want answered
Surprise: Break an expectation
Identity: Make the right reader feel seen
Proof: Use a number, name or timeframe
Exclusivity: Offer a sense of insider access
Loss aversion: Show what a mistake could cost
Vulnerability: Share a real doubt or failure

2. Match the combination to what you want the post to achieve

For bookmarks: curiosity + payoff + proof
For replies: surprise + identity
For shares or DMs: identity + payoff
For trust and follows: proof + vulnerability, with curiosity if it fits
For clicks, sign-ups or DMs: payoff + proof + loss aversion

Treat these as starting points. If you don't have real value to provide, share a proof or mistake you made to build trust

3. Write the hook in your own words

Best practices are useful, but familiar phrasing like "hot take" can make you sound like everyone else... and people just scroll past you

This post's hook uses surprise: following familiar advice can make people scroll past. The next line offers a payoff and opens a question: how do you use the psychology without sounding copied?

Save this for the next time you're writing a hook

[Francesco @prafaolo]

@joncphillips @peerpush_com done. well deserved :)

[Francesco @prafaolo]

@stedA_on Give it voice interaction and you got Jarvis at your service haha

[Francesco @prafaolo]

@joncphillips Thanks man for taking the time to do all of this. It's rare to find open source tools worth installing and using daily

[Francesco @prafaolo]

@JohnAthayde Actually, the agent can look at the container logs (which it does) and driving the UI is the most realistic form of testing after automated e2e tests

[Francesco @prafaolo]

Here's how I let my agents review 1000+ PRs a month unattended:

At this volume of PRs, the bottleneck is how much can you trust your validation layer, not coding

Unit tests are only the tip of the iceberg. A validation flow you can trust involves multiple code reviews and feature testing in the app

I get asked about my flow quite often, so here it is:

1. Prerequisite: I use git worktrees and Vercel's portless library to run 5-10 agents working in parallel without codebase or locahost port conflicts

2. After every set of changes, Opus reviews the changed code and cleans up AI-slop in comments - AI tends to write unnecessarily long comments that make the codebase seem AI-generated, which I don't want

3. The CI pipeline runs a code review Sonnet agent as a GitHub action to review every commit on the PR. I gave it my Claude subscription token so that I don't have to pay for API tokens - you can do the same using Codex

4. Opus drives a hands-on test on my app in localhost to actually test the new feature and all edge-cases - it clicks through the app and does exactly what a human tester would... all unattended

To do this on every project without prompting the flow every time, I created a set of skills that contain everything

If you also want to reach your 10x dev potential, the skills are open-source -> pumpkit-io/sdlc on GitHub

[Francesco @prafaolo]

I shipped 71 PRs in the last three days, just on side projects

The unlock was letting agents work in parallel without merge conflicts, port conflicts or shared accounts

Codex turns ideas into tickets, then each agent works unattended on a ticket in a separate git worktree

Vercel's portless library gives each worktree its own app instance on its own .localhost URL, so agents can test in the browser without fighting over ports

Vercel's Emulate gives each worktree its own local, stateful mocks for Stripe, Google OAuth, Gmail, Resend, and GitHub, so there are no shared sandbox accounts or per-instance OAuth and webhook setup

Without isolated worktrees, app URLs and integrations, parallel agents just step on each other. With them, I can run 10+ instances at once and let the work continue unattended for entire days

The agentic harness that packs all of this is open-source on GitHub

[Francesco @prafaolo]

@joncphillips For products that integrate LLMs especially, shipping an MCP means users can use your product through their Claude Code/Codex subscription, delegating part of the tokens cost to their subscription and thus saving you money

[Francesco @prafaolo]

@joncphillips Well deserved man. Thanks for taking the time to make Runlight open🤘

[Francesco @prafaolo]

Link to my agent skills: https://t.co/vhVXR4FhGx

[Francesco @prafaolo]

@stedA_on I also don't like that tests aren't public. I looked at the repo but found nothing on this. I'm willing to try it anyway of out necessity to spend fewer tokens somehow

[Francesco @prafaolo]

everyone moving to t3code now

[Aaron Delasy @aarondelasy]

idk how theo and his team pulled this off but my tests show that t3code is now saving around ~40% of usage on avg, but how is this even possible???

anyway, sorry team but everyone have to install t3code now 😂

[Francesco @prafaolo]

@joncphillips very likely i'll reach out when i'm done

[Francesco @prafaolo]

@joncphillips sick! thanks a lot for taking the time for all the details

i'm asking claude to setup runlight on my app, i'll throw this at it for context

i think i'll have to stick to posthog in parallel for some other things, like user session recordings and LLM traces, but that's fine

[Francesco @prafaolo]

@stedA_on Makes sense

[Francesco @prafaolo]

5 websites for travelers to lean back in your seat and relax ✈️

1. https://t.co/2Kcgzf1Ey3 - see where your plane really is and spot delays before airlines communicate them

2. https://t.co/EuvPQSbH13 - see which window seats are actually aligned with a window in your plane

3. https://t.co/WQTZKizoV7 - find the cheapest flights across all airlines. Set the destination to "Everywhere" and the date to "Cheapest month" to see where you can fly for the least money

4. https://t.co/pjZtJcdA9w - disposable email for airport and cafe Wi-Fi logins and one-off sign-ups. Don't use it for bookings though

5. https://t.co/QmXmVIGWKU - line up worldwide time zones at a glance

Travel gets easier when you know where to look for what

[Francesco @prafaolo]

@stedA_on I should setup an automation as well. For now I just scroll through X to find out

[Francesco @prafaolo]

🚨 X just changed its open-source ranking weights and now pays for people who stay and like, not for taps and clicks

- Likes: 0.66 to 0.708, the highest ever, up from 0.5 in early October
- Time spent on a post: 0.006 to 0.0075, nearly double since October
- Tapping into a post: 0.3 to 0.274 and falling
- Opening a video: 0.02 to 0, only people who keep watching count now
- "Not interested" got harsher: -53.07
- AI replies under 250K+ accounts can get an "llm slop" label and skip notifications unless they follow you
- The first 2 hours and first 200 views are still when early likes decide the extra push

Write stuff people actually read to the end and like, and write your own replies

[Francesco @prafaolo]

@joncphillips At this point wait until Mon. The days with most active X users are Mon, Tue, Wed

[Francesco @prafaolo]

@stedA_on You mean automating the discovery of good websites or their fruition by putting an automation layer in front of them?