
Building with AI agents • learning in public • sharing what works | @nabulines 👁️⃤
Ai World
Joined December 2021
- Tweets31.7K
- Following937
- Followers894
- Likes32.4K
Xman retweeted
$100 to one person
who likes this post ↓
Keep your Bitcoin. Get your liquidity.
Crypto Loan is now fully rolled out in Ledger Wallet™: take a stablecoin loan backed by your BTC, right inside Ledger Wallet™. Powered by @Morpho via @yield_xyz.
Xman retweeted
Video editing always felt like a skill you either have or you don't
Most people have the idea. They get stuck on cuts, captions, music and switching between tools
Cardboard is now a plugin inside ChatGPT and Claude → drop your clips, describe the video and it builds the edit
Then you keep chatting until it's right
🤝 Paid partnership
Introducing Cardboard for ChatGPT and Claude.
You never know when your best ideas will find you.
So we made sure that @usecardboard is there, whenever they do.
At your desk, or on the go, wherever you are :)
AI built the city 👾
AI writes the laws
AI is the police
If you lived there, what’s the first thing you’d get arrested for? 👀
🤖 Made with AI
Xman retweeted
Kimi finished in 2 min
Claude took 40.
I'd still pick Claude !
Same prompt:
Easy tasks → almost identical
Harder one → Claude won
Kimi wrote the code and stopped.
Claude wrote it, then tested it itself.
Speed or verification, which matters more to you?
Kimi finished in 2 min
Claude took 40.
I'd still pick Claude !
Same prompt:
Easy tasks → almost identical
Harder one → Claude won
Kimi wrote the code and stopped.
Claude wrote it, then tested it itself.
Speed or verification, which matters more to you?
Prompt 👇🏻
Build an interactive 3D "bioluminescent abyss" experience in ONE self-contained HTML file. This is a showcase art piece: it must look like a premium, award-winning WebGL site, not like a game prototype.
ART DIRECTION (the most important part, follow strictly):
- Organic, smooth, elegant forms only. NO boxy, voxel, blocky or Minecraft-like look. Avoid BoxGeometry for anything visible. Build shapes from high-segment spheres, LatheGeometry, TubeGeometry along curves, displaced and noise-deformed geometry, and custom shaders.
- Soft glowing materials: custom ShaderMaterials with fresnel rim light, translucent layered membranes, emissive gradients and subtle iridescence. No flat default materials.
- A curated limited palette: deep navy and teal darkness with cyan, soft violet and warm amber glows.
- Cinematic post-processing: strong but tasteful bloom, depth of field, gentle chromatic aberration, vignette, film grain and color grading (one custom final shader pass).
- Atmosphere: volumetric-looking light shafts from above, depth fog, floating particles (marine snow) with depth.
- Everything moves slowly and gracefully: breathing, pulsing, swaying, drifting.
- UI like an editorial magazine: a large elegant serif headline (your own short poetic title, system serif fonts only), tiny uppercase monospace labels with letter spacing, thin lines, a semi-transparent frosted panel.
Tech: Three.js 0.160.0 via an import map from cdn.jsdelivr.net/npm/three@0… and addons from cdn.jsdelivr.net/npm/three@0… (OrbitControls, PointerLockControls, EffectComposer, RenderPass, UnrealBloomPass, BokehPass, ShaderPass, OutputPass, and other official addons as needed). No other libraries, no image, model, font or audio files: everything is generated in code.
The world (procedural, generated from a numeric seed shown on screen):
1. A seafloor of soft dunes with glowing sea plants, anemones and curved branching coral that sway in a current.
2. A swarm of translucent jellyfish with pulsing bells and long flowing tentacles (animated curves), drifting upward and around.
3. A giant gentle creature (an ancient whale-like silhouette or a huge manta) slowly circling far away, partly lost in the fog.
4. Schools of small glowing fish moving with simple flocking behaviour, reacting to the camera.
5. A central ancient structure: an elegant ring or arch covered in bioluminescent growth, with a slow pulse of light running through it.
Interaction:
6. Depth slider: from twilight zone (more blue light from above) to the deep abyss (almost black, only living light), smoothly changing light, fog and palette.
7. Sonar pulse: click anywhere or press a button to send a glowing spherical wave; plants, jellyfish and fish light up as it passes and the fish scatter.
8. Regenerate world (new seed, smooth fade), Camera drift on/off (slow cinematic tour), Ambient sound on/off (deep evolving underwater drone, soft whale-like calls and bubbles, synthesized with Web Audio, starting only after a click).
Camera and viewing (very important):
9. Orbit mode with smooth damping and limits (never below the seafloor).
10. A "Views" menu with at least 5 presets the camera glides to: wide panorama, close-up of a jellyfish (follows it), through the central arch, looking up at the light shafts from the seafloor, and following the giant creature.
11. Free-swim mode: WASD + mouse look with pointer lock, gentle inertia, Esc to return.
12. Photo mode: hides UI, click any object to set focus, adjust blur, download a PNG.
Performance: quality preset (Low / Medium / High) and a small FPS counter. Smooth on a normal laptop at Medium, no console errors, works when opened directly from a file, resizes correctly, touch works on mobile.
Correctness matters more than length: every listed feature must actually work. Output only the complete file in a single code block, with no explanation.
3
73
One of my early AI projects wasn’t one giant prompt
RSS → dedup/store → Gemini filter → analysis → report → Telegram
Some stages were deterministic, others used the model
That split shaped how I think about AI systems: the model is a component, not the whole system
12
17
208
Xman retweeted
Hazels x Nabulines 👁⃤
50 WL spots up for grabs.
How to enter:
❤️ Like
🔁 Repost
👥 Follow @nabulines & @0xhazels
Drop your EVM address below.
50 winners.
Good luck.
Failure isn’t one state
I stopped treating every agent error as “retry”
Depending on what failed, the next move may be RETRY, FALLBACK, SKIP, RECONCILE, or STOP
Detection tells me something broke.
Recovery decides what can safely happen next
9
15
178
I stopped treating an agent restart as “run it again”
In my workflow, STOP, RESUME, and RESET became different operations
Resume now means:
inspect preserved state → validate against current reality → continue from the safest known point
Persistence alone isn’t recovery
6
11
153
A simple agent doesn’t have to mean an unstructured one
A small core can still have clear scope, task context, state, verification, and recovery
Professional architecture isn’t about how many components you have, It’s about whether each responsibility is clear
16
18
180
When I was building with free API limits, I couldn’t afford to keep throwing the whole project at an agent every time
So I started narrowing the task first:
what does it need to know, what can stay out, and what is actually relevant right now?
At first, that was just a workaround for limited resources but later, even with better tools and fewer constraints, I kept doing it
Because the useful part wasn’t saving tokens
The constraint had forced me to notice how much vague tasks and unnecessary context were costing the workflow
What started as a limitation became a habit:
give the agent what the current task needs, not everything you happen to have
11
14
134
Prompts tell coding agents what to achieve
Boundaries define what they’re allowed to do getting there
In my workflows, READ / / MODIFY / APPROVAL-GATED are different authority levels
More capable agents make that distinction more important, not less
18
28
355
Chat with a coding agent is a good workspace
It shouldn’t be your project’s source of truth ⚠️
Decisions, state, constraints, and next steps should survive the session.
I moved that project truth into persistent files the next agent can read
New session, Same project truth
When an agent says “done,” I don’t treat it as a state change
No Evidence = No Project Fact ⚠️
The claim has to earn its status with matching proof
What matters isn’t how certain the agent sounds. It’s what the evidence actually lets you claim
Replying to @Xman_base_eth
agent done signals belief only
CODE TEST GIT E2E layers prove the work is complete
A coding agent saying “done” tells me what the agent believes happened.
It doesn’t tell me what has actually been verified
I ran into this in one of my projects, so I started separating different kinds of evidence:
“Implementation exists” → CODE
“Automated check passes” → TEST
“Change is recorded” → GIT
“Feature works end to end in the tested workflow” → E2E
Those claims are related, but they’re not interchangeable.
Code existing doesn’t prove the test passes.
A passing test doesn’t prove the feature works end to end in the tested workflow, that changed how I define completion when building with coding agents:
verification isn’t something that happens after the work, It’s part of what makes the work complete
A coding agent can have access to your entire repo
That doesn’t mean the entire repo belongs in every task
I learned this first because context was expensive.
Now I care about it for a different reason: system design.
For each task, I try to separate:
• what the agent can access
• what the agent actually needs
• what should stay out of its working context
Project knowledge can be large.
Runtime context should be deliberate.
The goal isn’t minimum context.
It’s minimum sufficient context.
19
26
239
A bad assumption can spread like a silent virus 🦠
With OpenClaw 2.0, context can persist across people, agents, and handoffs
That’s powerful but it creates a subtle failure mode.
One wrong assumption enters the shared context, the next person builds on it. Another agent inherits the result.
Nothing has to look broken. Each step can be perfectly reasonable on its own.
A few handoffs later, the original mistake is buried under everything built on top of it.
By then, the team may not even know where the assumption came from, or whether it was ever validated.
Shared context doesn’t just preserve knowledge. It can preserve confidence in something that was wrong from the start.
The next context problem may not be forgetting
It may be knowing when something everyone remembers needs to be questioned again
OpenClaw 2.0 has arrived
openclaw.ai/blog/openclaw-2-…
Ma first reaction was: why not just use a VPS?
Then it clicked: this isn’t really about a better server?!
If your agent mostly waits and only needs compute in bursts, Cloud Run Instances starts to make a lot more sense
We all see the same internet. The difference may be what each of us asks our agents to find
Grok Bot can now search X, read your timeline, and follow live signals
Ask it to look for one thing, and you get one version of X. Ask it to look for something else, and the same feed becomes completely different.
Same internet, different questions, different signals
Grok Bot now has improved support for X!
When you connect your X profile, a developer account will be automatically created with some included credits.
10
13
216






















