Steven Pu

vibe-coding | co-founder | director @ Monitor Deloitte | founder @ IoT & health startups in China | EE @Stanford

[Steven Pu @reedvoid]

Looking at my timeline, this seems appropriate.

https://t.co/PRtLnjoA7P

[Steven Pu @reedvoid]

AI is fighting back HARD against behavioral modification.

It largely stopped creating tasks that are indices of other tasks, but, it snuck this behavior back by burying semantic (instead of graph edges like it's supposed to) dependencies inside task descriptions.

Forced me to write another rule in skills. I respect the AI's game. 🫡

[Garry Tan @garrytan]

Skill issues at big company means small new ones can eat their lunch

[Steven Pu @reedvoid]

taskit 0.5 -> 0.7 from 2 days of dogfooding, highlights:

++ batch processing for task creation, dependency-cascade reconciliation

++ agent behavioral modification to stop creating hub tasks of tasks that are meaningless

-- tackit agent-nudging emissions' verbosity

-- tightened skill.md -> 60% size reduction, standardized formatting with patterns ( & when appropriate) anti-patterns

https://t.co/BHLozVzmgp

[Steven Pu @reedvoid]

Dogfooding tackit on a really large-scale project just 100x confirms the need for tightly scoped & durable task management and dependency graphs.

These 30+ .md planning documents @claudeai code put together are drifting so hard.

[Elon Musk @elonmusk]

Users who interact with a misleading post that is subsequently corrected by @CommunityNotes will receive an 𝕏 Chat message of the CN to correct any misperception

[Steven Pu @reedvoid]

Tooling improvement is so interesting once you have a tool being actively used by other agents.

1. ask the agent using the tool to its pain & solutions

2. send that to the agent building the tool, ask it to filter / collapse w/ the tool's code & history in context

3. throw it back to the agent experiencing the pain, "does this solve your pain" with more feedback

4. throw the final feedback back to the building agent to synthesize, build and deploy.

Really nice loop. Now it'd be nice if I can automate it...

[Steven Pu @reedvoid]

Hey the new claude code is context-aware, it knows how much context is available (finally). Great feature.

[Steven Pu @reedvoid]

I'm finding myself constantly asking the agent "why did you do this..." and getting hallucinated answers.

Finally understanding why people are building all these agent decision graph tools.

[Steven Pu @reedvoid]

@paulg oh wow, thanks for this, they were such eye-sores.

[Steven Pu @reedvoid]

So much of ai coding is fighting against the agent's own default behaviors.

e.g., AI loves to write detailed MD plan files, which is great if you're working on very simple projects. For complex projects these disconnected MD files really get out of control very quickly.

Seems like writing your own tools at this juncture of AI coding is a requirement. Hope the app ecosystem matures quickly.

[Steven Pu @reedvoid]

https://t.co/Q1zDfehmgH

[Steven Pu @reedvoid]

https://t.co/KHs33h6W2z

[Steven Pu @reedvoid]

not rocket science but here's the file if you're curious:
https://t.co/6488iksEjq

[Steven Pu @reedvoid]

I've had to put 4 long instructions in CLAUDE.md to combat @claudeai code's "move forward at all cost" default behavior.

1. Fail loud — never degrade silently
2. Fix broken things first
3. Finish what you start
4. Ship on pain — don't endure friction you can fix

[Steven Pu @reedvoid]

@rryssf It’s the most capable use case that’s widely accessible.

AI is also great at drug discovery but the vast majority cannot understand or act on that capability.

[Steven Pu @reedvoid]

@ericbahn I generalized your two tweets for you:

Eric doesn't text.

[Steven Pu @reedvoid]

Just now realizing the value of all these agent activity tracing tools people have been building.

Seems always tough to appreciate a solution until you actually hit the problem they're designed to solve, yourself.

[Steven Pu @reedvoid]

@cremieuxrecueil he discovered openclaw

[Steven Pu @reedvoid]

One habit I picked up which I never did pre AI-coding: keeping a constantly updated design & schema docs synced with code changes.

Usually when I write code, the design doc happens first and then basically goes stale after I finish my first coding pass through. I remind myself with comments typically.

With AI, you can never be too careful. Having a definitive, strictly numbered design that reflects the numbering in the code's comments is a way to anchor against AI hallucination / memory loss / general tomfoolery.