I Gave Claude Code 50 Tasks This Week. Here's the Pattern That Emerged.

Yongcheng Mu — Thu, 11 Jun 2026 15:50:18 +0000

I Gave Claude Code 50 Tasks This Week. Here's the Pattern That Emerged

I logged every task I sent to Claude Code for 7 days. 50 tasks, three categories emerged. Sharing here because I think the pattern is more useful than any "top 10 prompts" list.

Category 1: "Do this small thing correctly" (60%)

Tiny, well-scoped jobs — write a test, rename a function, scaffold a config file. Claude Code crushes these. Median time: 12 seconds.

The unlock: I stopped writing detailed prompts. "Add a unit test for the new retry function" is enough. More detail made it slower, not better.

Category 2: "Plan, then I execute" (25%)

For refactors I wasn't sure about, I asked Claude Code to plan (read code, write a plan, no edits). I'd review the plan, tweak it, then either let it execute or do it myself.

The unlock: planning mode is a thinking tool, not a coding tool. The plan was the deliverable; the code was optional.

Category 3: "I have no idea where to start" (15%)

Greenfield problems. I gave Claude Code a fuzzy goal and let it ask me questions. Surprisingly, this is where the model was most valuable.

The unlock: when I gave it a fuzzy goal, it asked 3-5 clarifying questions that I would have skipped. Saved hours.

What I'd change

Cap task length. >2000 tokens of context = noticeably worse.
Keep a "solved it" log. The pattern I missed was: tasks I thought were "do" were actually "plan."
Don't be polite. "Fix this" beats "could you please fix this" by 5%.

ai #claudecode #llm #workflow

Codex in 2026: A Practical First-Hour Walkthrough

Yongcheng Mu — Thu, 11 Jun 2026 15:50:06 +0000

Codex in 2026: A Practical First-Hour Walkthrough

I keep getting asked "what does a first session with Codex actually look like?" — so here's a real, unedited run from this week, with the exact prompts and the gotchas I hit.

What Codex is, in one sentence

It's an agent that lives in your terminal and in your editor. You give it a goal, it does the work, you review the diff.

My first hour, step by step

Install + auth. Five minutes, including the OAuth dance.
Pick a small, well-scoped task. I chose "add a --dry-run flag to my deploy script". Bad first task would be "refactor my whole backend."
Write a one-line goal. Not a paragraph. Add --dry-run to scripts/deploy.sh that prints the commands without running them. Update the README.
Watch, but don't interrupt. The first time I used Codex I kept tabbing in to "help." That's a trap — let it finish.
Review the diff like a human PR. I caught two bugs it introduced. Both real, both subtle.

The 3 gotchas nobody warns you about

It will over-engineer if you let it. Cap the scope in your prompt.
It doesn't know your secret conventions. Tell it up front.
It will confidently make up API names. Always run the code.

TL;DR for a busy dev

Codex is a real productivity unlock, but only if you treat it like a junior: tight scope, fast review, trust but verify. After a week you'll wonder how you shipped without it.

DEV Community: Yongcheng Mu

I Gave Claude Code 50 Tasks This Week. Here's the Pattern That Emerged.

I Gave Claude Code 50 Tasks This Week. Here's the Pattern That Emerged

Category 1: "Do this small thing correctly" (60%)

Category 2: "Plan, then I execute" (25%)

Category 3: "I have no idea where to start" (15%)

What I'd change

ai #claudecode #llm #workflow

Codex in 2026: A Practical First-Hour Walkthrough

Codex in 2026: A Practical First-Hour Walkthrough

What Codex is, in one sentence

My first hour, step by step

The 3 gotchas nobody warns you about

TL;DR for a busy dev

ai #llm #devtools #codex