Aseprite en exemple. Tu peux l'avoir gratuitement en le compilant toi-même avec CMake, Ninja, Skia. C'est drm-free Pourtant, énormément de gens préfèrent payer 20 € pour avoir un exécutable officiel prêt à l'emploi. Ils n'achètent pas quelque chose d'impossible à faire eux-mêmes, ils achètent du temps et de la simplicité.
Over the past few months, I've mostly been building things to remove my own pain points.
Not because I had a roadmap.
Not because I was chasing ideas.
Just because I needed better tools and better workflows.
But I did have a goal I wanted to achieve.
Along the way, I built:
- A local AI agent isolation and hardening protocol with a defense-in-depth architecture, built around Hermes Agent, Claude Code, OpenCode, and OpenClaw.
- A secure MCP-based environment allowing AI agents running in a VM to interact with tools like Unreal, Blender, and Aseprite hosted on my main machine.
- A spritesheet generation tool originally built for Unity that evolved into a standalone Windows application, now becoming a SaaS to test the market.
- Reverse-engineering the AKI Engine to understand its internals and build new things on top of it.
And yet, the further I go, the more one question keeps coming back:
Does any of this actually have value for other people?
Does it solve problems beyond my own?
Maybe the hardest part isn't building systems.
Maybe it's figuring out whether the solutions we create for ourselves can become useful to others.
I wasn't joking :
>be me
>using OpenAI since late 2022
>predicted the LLM boom on French JVC Finance
>switch entirely to Claude in 2023
>never really go back to OpenAI
>even after Codex drops in 2025
>try Codex and GPT 5.4 on a free account
>immediately realize how powerful it is
>still stay with Claude
>workflow is good
>building is going well
>Opus 4.8 breaks my generator's core rules
>Opus 5 fails to fix it
>Fable 5 is the only model that figures it out
>but safeguards keep triggering before it can properly finish the job
>Anthropic turns user experience into a psychological experiment
July 27, 2026
>switch back entirely to OpenAI
>run Codex with GPT-5.6 Sol
>medium reasoning
>it fixes the problem
30 regression tests pass
zero failures
Fable 5 is back.
But that won't stop me from canceling my Anthropic subscription and switching to OpenAI until I finish what I'm working on. I really dislike the way you treat your customers.
@ThomasItsMe Oui et non, il faut vraiment faire attention a ce qu'elles disent, quand j'ai fait mon mcp pour aseprite, ChatGPT m'a dit que ca n'existait sous la forme .aseprite-extension, j'ai vérifié par moi même et cela existait, mon seul levier, j'ai fait mieux que le premier.
@LHelvetique Vrai. Perso, ma spécialisation, c'est l'infra et les systèmes, même si j'ai eu la chance d'avoir appris à coder avant l'arrivée de chatgpt, fin 2022 grâce à des potes dev, aujourd'hui je refuse que l'on me qualifie de dev.
@NotWill42 I didn't know about that fishing method. But grab the Scatter Sphere Launcher instead; it's a lot more fun to use. It'll chew through your Pal Sphere supply, but it's worth it.