keepsimple logo
Language switcherru
picLog In
W O L F ’ S · T E R M I N A L · 天

Thinking

Vibecoding skills

  1. Stage I. Solo: you and one run
  2. Thinking

    Wrong perspective

    It knows the answer.

    Right perspective

    It predicts a likely answer, and a likely answer can be wrong.

    Vibecoding skills

    You decided to check what ChatGPT, Claude or Gemini says before you use it.

  3. Thinking

    AI is something I ask for advice.

    AI is something I hand work to.

    Vibecoding skills

    You decided to run Claude Code or Codex in a terminal with bypass mode on and let it act.

  4. Thinking

    It remembers what I told it last time.

    It knows only what is written down for it.

    Vibecoding skills

    You decided to keep a global CLAUDE.md and one per project.

  5. Thinking

    The more I give it, the smarter it gets.

    Past a point, more context makes it dumber.

    Vibecoding skills

    You decided to keep CLAUDE.md under 200 lines and move the rest into documents it opens when needed.

  6. Thinking

    A bad answer means a weak model.

    A bad answer usually means a bad question.

    Vibecoding skills

    You decided to put one goal in each request and say what the answer must contain.

  7. Thinking

    It looks right, so it works.

    Nothing works until a test says so.

    Vibecoding skills

    You decided to write tests and bring in a Judge agent that checks the first one.

  8. Thinking

    I found my model.

    The best model for my task changes every few months.

    Vibecoding skills

    You decided to write a script that sends the same task to several vendors and scores each with your evals.

  9. Stage II. Orchestration: you and a system of runs
  10. Thinking

    One smart agent can do the whole thing.

    A big problem is a chain of small runs.

    Vibecoding skills

    You decided to start subagents and pass files from one run to the next.

  11. Thinking

    Give every agent everything, just in case.

    What a run does not see cannot mislead it.

    Vibecoding skills

    You decided to keep a list per step of what the run receives and where it may write.

  12. Thinking

    I check the result at the end.

    Every step needs its own check, or the first mistake spreads through the chain.

    Vibecoding skills

    You decided to put an eval between runs, so a failed step passes nothing on.

  13. Thinking

    More AI makes the system better.

    Wherever a step is predictable, code does it better than a model.

    Vibecoding skills

    You decided to turn routine steps into ordinary functions and hooks that check the output.

  14. Thinking

    Each session is a fresh start.

    What the system learned has to outlive the session, if it passed an eval.

    Vibecoding skills

    You decided to give each agent a memory file and every run a shared store to search.

  15. Thinking

    I have to drive every step myself.

    My job is the decisions, and the agents handle the handoffs.

    Vibecoding skills

    You decided to write an onboarding protocol and a message format your agents share.

  16. Thinking

    My work is doing the task.

    My work is shaping tasks the system can deliver without me.

    Vibecoding skills

    You decided to set up schedules and watchers, so the system runs overnight and leaves logs.

  17. Thinking

    I run a set of agents.

    I run something closer to a company.

    Vibecoding skills

    You decided to group your agents into departments, each with its own Judge, and let a product manager agent run whole projects through them.

  18. LVL 15???

At that point you have stopped renting a model to talk to and started owning the system around it.