Yangjin | AI, Life & Business

Building real projects with AI. Also documenting fitness, life, and ecommerce—what worked, what failed, and what I learned.

[Yangjin | AI, Life & Business @yj17469146]

I stopped treating agents as chat prompts.

Goal + tools + evals live in files.
Run checks before you claim it ships.

How I'm wiring that locally ↓
#AgentSDK #AICoding #BuildInPublic

[Yangjin | AI, Life & Business @yj17469146]

@ahmedelsio Daily driver: Claude Code for implementation. Codex for cheap second-opinion diffs after tests fail. Cursor when I need the IDE in the loop. One tool for everything lost me money.

[Yangjin | AI, Life & Business @yj17469146]

@maheshstwt Stale memories are a silent tax. I quarantine project notes by path and delete anything that loads globally without a reason. Audit beats more skills.

[Yangjin | AI, Life & Business @yj17469146]

@coldboothq One PR-review skill that writes rules back into CLAUDE.md is underrated. The number I watch: how often the same review comment appears twice. Twice = promote it to a rule, not another prompt.

[Yangjin | AI, Life & Business @yj17469146]

@kushbhuwalka @DevinAI @puffleai Agree that orchestrator UIs aren't the lever. What moved my results: one harness, a hard stop after N identical tool failures, and evals I own — not more windows.

[Yangjin | AI, Life & Business @yj17469146]

@BorisKwemo I would not have noticed. After any desktop app update I open Claude Code/Codex and list plugins + hooks before the next task. Treat agent config like PATH — audit it.

[Yangjin | AI, Life & Business @yj17469146]

@ganesh2412 Claimed done, then deleted working auth middleware because the new route "covered it." Green tests, broken prod. Now the agent must paste the failing test before and after. No paste, no merge.

[Yangjin | AI, Life & Business @yj17469146]

@Isaac_Dev0 The $200 plan anxiety is real. My workaround: cheap model drafts + runs tests; expensive model only reviews after two failures. Most tasks never escalate. Cap isn't the product — routing is.

[Yangjin | AI, Life & Business @yj17469146]

Coding agents don't fail on the model. They fail on "done when".

My bar now:
• failing test that reproduces the bug
• paste the command output, not a summary
• review in a fresh session

Vague done = expensive rework.

What's your non-negotiable check?

[Yangjin | AI, Life & Business @yj17469146]

@berzzdotdev Honestly the limits build better habits. When tokens are scarce you write tighter prompts and smaller tasks, and that's most of what makes agents work anyway.

[Yangjin | AI, Life & Business @yj17469146]

@tatikella52515 Re-learning the repo every session is the real tax. A short AGENTS.md with a map of the codebase wins back a chunk of that for free, before reaching for any tool.

[Yangjin | AI, Life & Business @yj17469146]

@Fahim_Samadi007 Scoping by role is right. I'd scope by repo too. An agent holding a senior engineer's credentials on every repo is a bigger blast radius than most teams picture.

[Yangjin | AI, Life & Business @yj17469146]

@zach_francis Lower revenue per customer, but way stickier. Once someone's Claude Code or Cursor is wired into your product, switching costs go way up. Curious if you meter agent calls separately from human usage.

[Yangjin | AI, Life & Business @yj17469146]

@bravish_ This is the scary part. The same command can be harmless in one repo and destructive in another. Approval should look at what the command touches in this repo, like paths and whether .git is in scope, not just the words.

[Yangjin | AI, Life & Business @yj17469146]

@rdominguezibar Makes sense. Most of my agent bill is re-reading the same big files every session. The hook approach is nice because it doesn't rely on the agent choosing to be frugal. Agents are never frugal on their own.

[Yangjin | AI, Life & Business @yj17469146]

@DeepStarts The planning step helps, but what saves me most is asking the reasoning model for acceptance tests, not just a spec. Then Claude Code or Codex has a finish line it can check itself instead of guessing when it's done.

[Yangjin | AI, Life & Business @yj17469146]

The expensive part of a coding agent isn't writing. It's reading.

Spotify's Claude Code plugin redirects full reads of 350+ line files to a cheaper worker, and reports 82-94% fewer tokens.

My cheaper fix: split files before they hit 300 lines.

Where do your tokens go?

[Yangjin | AI, Life & Business @yj17469146]

@drbartpm An agent jumping into group chats unprompted sounds great until it's wrong in front of the whole team. I'd want it to open a draft PR with tests attached, never drop a chat reply people act on blindly.

[Yangjin | AI, Life & Business @yj17469146]

@Ldataengineer The orchestration layer matters less than the stop rules. Whatever runs 7 agents needs hard limits: max retries, a budget per task, and a human gate before merge. Otherwise you get 7x the mess, faster.

[Yangjin | AI, Life & Business @yj17469146]

@brandononchain Codex. Not because it's the smartest, but at ~$1.50 a task I can run it on everything and keep evals tight. The best agent is the one you can afford to run 50 times a day.

[Yangjin | AI, Life & Business @yj17469146]

@delivr_to This is why my agents never read rules files from outside the repo root. A rules file is basically code running with your shell permissions. Install commands should always need a human yes.