Fernando

Building AI agents that survive outside the demo: harness, evals, security, MCP. Engineer @iobuilders. Repo Rules Keeper. Futer, football matchmaking in BCN.

[Fernando @ferlogs]

AGENTS.md doesn't make your coding agent smarter. it makes it obedient.

an eth zurich paper tested 4 coding agents: context files didn't raise task success and added 20%+ cost. but the agents followed the instructions closely.

so that file should hold real rules, not a summary of your repo.

i've been working this way for over a year, across multiple microservices in production, and still improving it. repo rules keeper is that method as a template:
- reads the code, commits and review comments
- keeps only rules with evidence, deduped against the readme and docs
- hard rules fail CI, soft rules warn
- weekly, new review comments become proposed rules. a human merges them

tested on a quiet next.js 16 + supabase + cloudflare repo: 11 hard rules, 12 soft, 1 real bug caught.

template in the reply.

[Fernando @ferlogs]

@jasonzhou1993 thanks jason. the historical rank step is what changed the story, without it i'd have posted the 35x as real growth

[Fernando @ferlogs]

@sharenow_today appreciate it. the club pages pattern was the one i didn't see coming, it's basically SEO that grows every time a club signs up

[Fernando @ferlogs]

i'm building futer, an app for people who organize amateur football games. so i ran /rea-growth on playtomic, the biggest court booking app in spain

88 of their top 100 google keywords in spain are club names. people search for the place they play and land on playtomic's club page

0 active meta ads, 0 paid search. and the 35x traffic jump in 2025 was just a domain move from .io to .com

stealing the club pages idea for futer. whole teardown cost me $0.39

[Jason Zhou @jasonzhou1993]

Introducing /REA-growth skill

Reverse-engineer any company's growth

- Name any company, get its growth blueprint
- Every move on its timeline: SEO, AEO, ads, social
- $0.58 for a full Lovable teardown
- Fully OpenSource

Try it at https://t.co/m9kFVsgRCV
Partnered with @sharenow_today

Repo link below 👇

[Fernando @ferlogs]

@jasonzhou1993 same, the .com chart alone looks like a 35x rocket. add the .io traffic back and it's +29% in 17 months. the real engine was the club pages

[Fernando @ferlogs]

if you want to run it on your own competitor, it's open source: https://t.co/7dHHVNl0hL by @treg_ai

[Fernando @ferlogs]

the file, if you want to check: https://t.co/dVvNoBrESF

[Fernando @ferlogs]

i read the x algorithm repo instead of another thread about it

the weights in home-mixer/params/param.rs:

copy link share 20
reply 5
quote 5
share via dm 5
follow 4
like 0.66
video view 0
not interested -52
mute -59

these multiply the predicted chance of each action. raw counts never enter the formula, so one report doesn't cancel 234 likes

a post people send to a friend beats a post people like

which one surprised you?

[Fernando @ferlogs]

what got my attention in agentweb is the permissions model

your agent calls a declared action instead of clicking through a site. payments carry an amount limit, and sensitive steps wait for a human

then a receipt comes back. no guessing from a screenshot whether the button got clicked

that's how agents get let into real accounts

[Fernando @ferlogs]

@chiefkittenme 29, engineer and founder in barcelona, building futer and agent tooling after work. what are you building as a solo founder?

[Fernando @ferlogs]

@rwspatin @X old school software architecture made the list, that's the one i'd defend hardest right now. ddd and hexagonal are what keep agent-written code changeable a year later

[Fernando @ferlogs]

@jimisdima @X good luck with the launch, what kind of creative work is it for? engineer in barcelona here, building futer and agent tooling after work

[Fernando @ferlogs]

@dev_coma engineer by day in barcelona too, building futer at night, football matchmaking for amateur games here. what are you shipping right now?

[Fernando @ferlogs]

@waulrz clean code and system design on a connect list is rare. backend engineer in barcelona here, ddd and hexagonal at the day job, now building rules so coding agents stop breaking all that

[Fernando @ferlogs]

@hirenthakore @waulrz terminal screenshots and regret is painfully accurate. what does shipreceipt do?

[Fernando @ferlogs]

@rvgdev @mkappworks a crew of agents each owning a project is a neat split. how do you stop two of them touching the same files, the inbox or permissions on the vm?

[Fernando @ferlogs]

@SamiDaarx honest results is the rare part. the test i'd love to see is the same AGENTS.md across claude code and codex, same task, which one actually follows it

[Fernando @ferlogs]

@Priteesshhh @X founder of futer here, football matchmaking in barcelona, and building agent tooling after work. what's the ai engineering side at lakefusion like, more agents or more data?

[Fernando @ferlogs]

@potwires stack is typescript and react, plus groovy on micronaut at the day job. goal is coding agents that follow repo rules ci can actually check

[Fernando @ferlogs]

@kasike boring is the right kind. the hours saved only stick if the owner can see what the agent did and say no before it sends anything

[Fernando @ferlogs]

@mkappworks ai harnesses made your list, which is my corner. i'm building repo rules keeper, rules for coding agents that ci actually enforces, and curious what benchlabs does