Junyoung Choi (Sai)

CTO of IssueHunt, drawing for fun

Pixiv : :

[Junyoung Choi (Sai) @rokt33r]

Drew Kiana
https://t.co/smK0QCQ31l

#honkaiımpact3rd #kiana

[Junyoung Choi (Sai) @rokt33r]

AI Coding tool, https://t.co/w1qfPCBLi3

Internal usage feedback(IssueHunt):
- Dev speed: x2
(Original estimation 5 working days -> finished in 2 days)
- token usage: 30~60% decreased

[Junyoung Choi (Sai) @rokt33r]

Tanstack db looks great. should try in this month.

[Junyoung Choi (Sai) @rokt33r]

@mattpocockuk I have a set of skills to work on a single plan document so I just need /clear. Other agents can instantly start working on the plan doc.

[Junyoung Choi (Sai) @rokt33r]

@sarahwooders Because MCP is compact. It can explain what the interface do and give an exact validator altogether. Otherwise, you have to iterate how the skills written, how API throws an error so agent can trouble shoot by itself. I'd say, MCP first and replace API or CLI with a proper skill.

[Junyoung Choi (Sai) @rokt33r]

Screw loop engineering, screw Claude tag. Been confronting low quality code quite often since Claude start using other GPUs. Low quality is totally fine for me. I just hate the approach, "fire and forget". They should encourage review AI's outputs more easily rather than hiding.

[Junyoung Choi (Sai) @rokt33r]

People should control the context when they use AI, but most don't. They just spray and pray, firing off prompts until one happens to hit the target. I guess even a broken clock is right twice a day. But what a waste of water and electricity. I guess that's how the world breaks.

[Junyoung Choi (Sai) @rokt33r]

@ben_sage I have released an app for it today if you're interested... It works for me for 4 months since I made its prototype.
https://t.co/w1qfPCBLi3
It heavily rely on my issue tracking app though, but it is OSS so you can grasp the gist if you want to hack it your self.

[Junyoung Choi (Sai) @rokt33r]

Released an OSS app for ai coding with Claude Code but not for vibe coding. No more delusion, no more cognitive overload.

APP: https://t.co/ST1w9vEobb
Github: https://t.co/lwRIZe3QBD

#claude #claudecode

[Junyoung Choi (Sai) @rokt33r]

https://t.co/lifZBXQYVt

Looks interesting.

[Junyoung Choi (Sai) @rokt33r]

Next.js is too heavy. Shouldn't have used it for a big project...

[Junyoung Choi (Sai) @rokt33r]

The custom context generation is insanely working well. I can just clear context anytime I want. Still agents know where they are and what they need to do next immediately in every fresh session.

[Junyoung Choi (Sai) @rokt33r]

Working on 2 projects, using 3 session concurrently. Still barely using 30% of daily token limit of Claude's $100 plan. Refactoring code / Context management FTW.

[Junyoung Choi (Sai) @rokt33r]

People should not use /compact of Claude Code.
You must control how to summarize your context. /compact doesn't know what should be summarized and what should be discarded.

[Junyoung Choi (Sai) @rokt33r]

Tried Claude Code desktop app. It SUCKS. Its bash tool cannot resolve profile properly. Better wait till it get more mature.

[Junyoung Choi (Sai) @rokt33r]

Shits are snowballing everywhere. VIVA LOS CODIFICADORES VIBRA!

[Junyoung Choi (Sai) @rokt33r]

Not even 90%. Slops everywhere.
https://t.co/nrI6zI4LFr

[Junyoung Choi (Sai) @rokt33r]

@v_r_a_i_m_e_n_t a코드 너무 더럽다.

[Junyoung Choi (Sai) @rokt33r]

Thought about going back to Korea. Now I'm applying for PR.🫠

[Junyoung Choi (Sai) @rokt33r]

Feels like Github is SO COOKED. Better to evac.