Aaron

24 • 55+ projects shipped • Building AI tools that work while I sleep

[Ronin @DeRonin_]

https://t.co/J0ULhSHLLH

[Aaron @CodeByAaron]

Everyone asks "how do I validate my idea?"

Wrong question.

What actually matters in week 1:

→ Talk to 10 people about the problem (not the solution)
→ Find 3 people who've tried to solve it themselves
→ Ask what they used, why it failed, what they wished existed

By the end you'll either know if it's real or know to stop.

Are you validating your idea or just building it?

[Aaron @CodeByAaron]

This is actually useful.

Most AI "guides" are just marketing.

Anthropic just dropped 33 pages of the real thing.

Building with Claude just got easier.

[0xMarioNawfal @RoundtableSpace]

Anthropic dropped a 33 pages cheat sheet for building Claude skills

https://t.co/jEuH95NGn3

[Aaron @CodeByAaron]

@PrajwalTomar_ Been doing this with a SOUL.md, USER.md and AGENTS.md setup. Agent reads them before every session. Took an afternoon to write.

Saved me weeks of re-explaining context. Most underrated hour you'll spend on any agent setup.

[Aaron @CodeByAaron]

@EXM7777 The system prompt portability thing is the hidden cost nobody talks about. What works on Claude needs a rewrite for anything else.

The knowledge base part is right though, that's the one thing that transfers cleanly.

[Aaron @CodeByAaron]

Qwen3.5:35b running locally on dual RTX 3090 Ti cards.

23GB model. MoE architecture. Only 3B parameters active per inference. Near-zero cost per token.

Took three hours to get it running. Two full downloads. An Ollama version mismatch. A model tag that didn't exist in the old library.

Now it runs fast enough to power all 5 agents as a fallback when the API drops.

Local inference isn't the future. It's already the boring default.

[Aaron @CodeByAaron]

Most people treat Claude like a ChatBot.

Ask it a question. Get an answer. Close the tab.

That's leaving 90% of it on the table.

The shift that changes everything:

→ Give it tools, not just prompts.
→ Give it memory files, not just context windows.
→ Give it cron triggers, not just manual runs.
→ Give it isolated sessions, not shared state.
Now it's not a ChatBot. It's infrastructure.

[Aaron @CodeByAaron]

@meta_alchemist The hardest part of building agents isn't the intelligence. It's making the feedback loop tight enough that the improvement is actually measurable.

How are you evaluating Spark's progress on the goal? Curious what your ground truth looks like.

[Aaron @CodeByAaron]

@gregisenberg The answer isn't quitting your job tomorrow. It's using whatever hours you have now to get so good at these tools that when the transition happens, it feels like a choice.

[Aaron @CodeByAaron]

@EXM7777 Ran a production agent for months and the issue was always stale context, not missing context. Fresh compact summaries beat a full conversation history every time. More window just means more rot.

[Aaron @CodeByAaron]

@PrajwalTomar_ Gave mine write access to the wrong directory once. Didn't delete anything but it reorganized 200 files at 3am. Now every agent gets scoped permissions before it touches anything in production.

[Aaron @CodeByAaron]

If your AI agent only works when you're watching, it's not an agent. It's an assistant.

Real autonomy needs 3 things:

→ Reliable cron triggers. Not manual runs you forget to press.

→ Isolated context per session. State bleed is silent and it will bite you.

→ File-based memory. In-context memory dies with the session.

Everything else is UI.

[Aaron @CodeByAaron]

@gregisenberg Spent months building UI wrappers around APIs thinking that was the product.

The moment I ripped the UI out and exposed the API directly, the thing actually got used.

The interface that survives is the one with the lowest friction, usually no interface at all.

[Aaron @CodeByAaron]

My agent knows what I worked on 3 weeks ago, which rules I added after things broke, and the history of every decision in the codebase.

None of that is intelligence. It's context.

Context is what separates genuinely useful AI from impressive demos.

[Aaron @CodeByAaron]

@levelsio That's the thing about X. You build something and post it, and the person who inspired it might turn up in your replies with actual advice. No other platform does that.

[Aaron @CodeByAaron]

@aaditsh Context is the prompt. Most prompt engineering is just compensating for missing context.

Once the model has the right inputs loaded, the actual wording becomes almost irrelevant.

[Aaron @CodeByAaron]

@claudeai Spent months bodging together Telegram pings so I could check long sessions from my phone.

Glad it's native now. The real unlock is being able to kick something off and actually walk away.

[Aaron @CodeByAaron]

First time my agent committed code to the repo while I was at work, I spent 15 minutes checking every line when I got home.

Now I check the diffs. Ten seconds.

Trust is a system you build over time. Same with code review. Same with most things.

[Aaron @CodeByAaron]

People romanticise "building in public" like it's a content strategy.

It's not. It's just being honest about the fact that most of what you ship doesn't work. And being okay with that in front of strangers.

[Aaron @CodeByAaron]

@EXM7777 The number is probably wrong AND the real number is still surprisingly large.

Both things are true.

Most of those "AI users" set their phone autocomplete on day one and haven't thought about it since.