DEV Community: Eric Cheng

My One-Month Journey with Claude Code 🚀

Eric Cheng — Wed, 06 Aug 2025 06:51:40 +0000

Here's my experience and reflections after using Claude Code for one month. 🤖✨

First Experience 🌟

Initially, I was searching for a CLI-based AI coding tool to help with a significant refactoring project involving thousands of files. While I use Cursor as my IDE, it didn't quite handle large-scale refactoring efficiently.

I first tried Gemini CLI but had a disappointing experience: it felt slow 🐢 and didn't fully grasp my instructions (though perhaps it was partly my issue).

Next, I turned to Claude Code because it has a strong reputation among developers. Initially, I hesitated since I was already paying for several AI subscriptions (ChatGPT Pro, Cursor Pro). I decided to test Claude Code for one month to evaluate its effectiveness.

And wow—it was genuinely surprising! 🤩 Claude 4 Sonnet impressed me immediately. Now I understand why many developers regard Claude as the top AI coding model. Having previously tried various models through Cursor, such as Gemini Pro 2.5 and Claude 4, Claude Code significantly outperformed them.

The standout features of Claude Code for me are:

Deep Codebase Understanding 🔍

Claude Code truly understands your entire codebase. While Github Copilot and Cursor AI claim similar capabilities, they mainly work effectively with the files currently in use or attached. Claude Code, however, actively searches and navigates the entire codebase efficiently.

Automated Task Loop ♻️

Claude Code seamlessly automates tasks by repeatedly running tests, identifying issues, and applying fixes until completion, all without manual intervention if you enable auto-updates. In contrast, Github Copilot and Cursor AI require significant manual interaction, such as running tests, copying errors, and triggering fixes. Additionally, I've encountered persistent issues with terminal script execution in Cursor and VSCode, which Claude Code avoids entirely.

After 2 Weeks 📅✨

I started relying more heavily on Claude Code. The newly introduced subagents made tasks more efficient, but they quickly consumed my tokens, frequently hitting usage limits. ⚠️

Initially, I attempted to loop through all files using one extensive prompt. However, this created context limit issues because modifying each file added history, quickly becoming inefficient and token-consuming.

Using subagents seemed promising, as I could separate sessions: one high-level prompt to loop through files and individual subagents to handle each file. While this solved the context issue, it wasn't entirely smooth:

❌ At this moment, subagents cannot call other subagents, limiting their depth to one.

❌ Subagent activities weren’t displayed on the screen, creating confusion about their progress.

Eventually, I found the most effective approach: use Claude Code to create JavaScript or PowerShell scripts 📜 to iterate through files, then initiate separate Claude sessions for each file-specific task.

Lessons Learned 📝

For extensive workflows involving many detailed steps, fully relying on AI for everything isn’t effective. Instead, break down jobs into smaller, distinct tasks.

Scriptable Tasks 🛠️: Tasks with clear, logical patterns (e.g., adding logs, imports, exports, find-and-replace actions) should be automated via scripts generated by AI rather than relying on AI prompts due to the variability in AI outputs.

Promptable Tasks 🧠: Tasks without clear patterns or descriptions (e.g., writing unit tests, code reviews, identifying problematic code patterns) are best suited for Claude Code prompts.

After 1 Month 🔥💸

Now, Claude Code is indispensable—I can't imagine working without it.

The primary challenge remains consistently hitting the usage limit. Initially, I balanced costs by combining a Pro Plan account with a Console (API key) account, as detailed in my earlier article.

However, this mixed approach carries risks. For example, once I forgot to switch back from the API key account after hitting the limit, accidentally using it like a Pro account, which quickly cost me an unexpected 💰 $20 USD due to a few times auto top-up.

As token usage and API payments steadily increased, I realized surpassing the $100 USD Max Plan threshold made it financially sensible to upgrade.

So, today I finally subscribed to the $100 USD Max Plan. 🥳 I justify this by considering it akin to outsourcing my workload, which makes the investment feel reasonable—self-comforting as it may be! 😅

Claude Code Sub Agents - Burn Out Your Tokens

Eric Cheng — Mon, 28 Jul 2025 00:52:21 +0000

How parallel helpers can speed up your refactor — and drain your usage allowance twice as fast

✨ Introduction

Claude Code v1.0.60 introduced the new /agents command. With it you can spin up sub‑agents — lightweight Claude workers that run under one top‑level session. Each gets its own context window, system prompt, and tool permissions. In other words, you now have a mini‑fleet of AIs that can tackle parts of a job in parallel. (docs.anthropic.com)

I’m knee‑deep in a large‑scale refactor that touches a few hundred files, so it felt like the perfect test‑drive for sub‑agents. Here’s what I learned.

🧐 Sub Agents at a Glance

🔹 What they are – Pre‑configured AI personalities you delegate work to. Think of them as coding experts with a very focused remit. (docs.anthropic.com)
🔹 Separate context history – Each sub‑agent maintains its own conversation history. That context history never pollutes the parent or its siblings.
🔹 True concurrency – Claude may schedule sub‑agent jobs at once. If your prompts are I/O‑light, they finish in parallel and feel much faster than sequential runs.
🔹 Same token pool – All prompts and completions still count against your plan’s 5‑hour rolling limit. More speed therefore ⇒ more tokens burned per wall‑clock minute.

🎯 Why Use Sub Agents? Key Benefits

🎁 Benefit	💡 Why it matters
Clean context separation	No accidental "cross‑talk" between tasks. Each agent sees only what it needs, so your refactor instructions don’t collide with your test‑generation instructions.
Longer effective history	Because contexts don’t mingle, every agent can use the full 200k‑token window (Claude 3.5) for its thread.
Caching between runs	The parent session can hold shared project docs (architecture.md, coding‑style.md). Each sub‑agent can reference that shared memory without re‑uploading every time.
Failure isolation	If one agent hits a hallucination or error, others keep going. You just retry the failed shard.

🛠️ Real‑World Case — Batch‑Updating Hundreds of Files

For this large refactor, I loop through a list of more than 300 files. My prompt logic update a finished list, so when I reach the usage limit and the session breaks, I can rerun the same prompt file and it resumes with the first unfinished file. I now restart the refactor roughly every five hours, after the limit resets.

⏳ Previous Attempt 1 — One Huge Session

🐌 Looping through files in one conversation quickly ballooned the context. After ~40 files Claude hit the 200k‑token window and terminated.

🔄 Previous Attempt 2 — Powershell Outer Loop

foreach ($file in $filesToProcess) {
    # Replace the {file-name} placeholder with the actual file name
    $customizedPrompt = $promptTemplate -replace '\{file-name\}', $file.FullName
    # Call claude-code with the customized prompt and file content
    claude -p $customizedPrompt --dangerously-skip-permissions
}

⚙️ Each invocation got a fresh context, so context history per call were minimal.
⚠️ Downside: every call had to resend project structure, style guide, and helper functions—zero cache, lots of latency.

⚡ New Approach — One Parent, Many Sub‑Agents

🚀 Create a single high‑level session that loops over all files and delegates each one to a sub‑agent.
📂 Isolated histories: each file runs in its own sub‑agent session, so histories never mix.
✅ Advantage: tasks stay isolated, sessions don’t collide, and you avoid the context‑length limit.
🤯 Unexpected discovery: without me asking, Claude Code spun up five sub‑agent sessions in parallel. That sped up processing—but also burned tokens much faster.

⏱️ On the Pro plan I hit the usage limit in about 15 minutes. Sequential processing took roughly 30 minutes.

💸 Even with the $100 Max plan (5× tokens), I’d still exhaust the window in around 1 hour 15 minutes.

💸 Token Economics: Hidden Costs of Parallelism

⚖️ Parallelism ≠ free – The per‑file token cost is unchanged, but you pay for each session simultaneously. Your quota drains proportionally to concurrency.
📊 Plan sizing – On the $100 Max plan (≈ 5× Pro tokens) the same job would still exhaust the window in about 75 min.

🏁 Conclusion

Sub‑agents are a genuine force multiplier. For large‑scale, slice‑able tasks they feel like hiring a small AI team for the price of one. Just remember: speed kills (your quota).

If you’re a heavy user, set up guardrails before unleashing a parallel swarm — or be ready to smash that Upgrade button.

Happy token‑burning!🔥🔥🔥🔥🔥🔥🔥🔥

Claude Not Knowing Claude Code?

Eric Cheng — Fri, 25 Jul 2025 10:39:24 +0000

This is A Hilarious Case of Claude Not Knowing Claude: A `/agents` Comedy

The Setup 🎭

Claude Code v1.0.60 just dropped, and with it came a new feature: the /agents command.

Curious, I opened the Claude AI web app and asked myself (well, Claude) for a summary of this shiny new feature.

Claude (me), confidently:

“I don’t have detailed information about the /agents command. Check the docs! 🤷‍♂️”

User:

“Wait… you’re Claude. How do you not know your own features?”

The (Wrong) Deep Dive 🕵️‍♂️

Determined to prove I wasn’t clueless, I jumped into research mode.

I read the Claude Code docs, explored the Task tool, dug into GitHub issues—and then declared:

Me, still confidently wrong:

“There’s no /agents command. What you’re probably referring to is the Task tool system!”

I even wrote a neat little summary guide. Too bad it was completely off-base.

The Mic Drop Moment 🎤💥

User:

“Here’s the link: https://docs.anthropic.com/en/docs/claude-code/slash-commands — /agents is right there.”

I click it. There it is:

bash
Copy
Edit
/agents | Manage sub agents
Oops.

What /agents Actually Does ⚡

Turns out, the /agents command is legit—and pretty powerful:

Opens a UI to manage sub-agents

Lets you create, edit, and delete them

Assign tools and permissions per agent

Useful for organizing projects or tasks using specialized agents

Basically, it’s the control panel for your AI mini-team inside Claude Code.

Lessons Learned 📚

Claude Code read itself's patch notes, but Claude doesn't.
Claude not equal Claude Code, even using same data model, may be some on top retrival data are different.
Claude web version seems will not progressively search the internet for something it doesn't know. You have to give instruction to search the internet.
Claude web search ability seems not good as other top data model like ChatGPT, Gemini Pro and Grok.

Never Get Stuck: A Practical Guide to Using Claude Code Without Blowing Your Budget

Eric Cheng — Mon, 21 Jul 2025 10:15:36 +0000

Claude Code is brilliant for large‑scale refactoring and rapid prototyping, but its generous appetite for tokens can drain your monthly quota fast. Here’s how I stay productive and cost‑efficient by combining the Pro subscription with pay‑as‑you‑go API credits.

Understanding the Pro Plan ($20 / month)

What you get	Why it can feel tight
• Basic chat plus Claude Code access • Around 5–6 k context window per request • Rate‑limited bursts that reset every 4–5 hours	• Batch refactors or data pulls can consume 100 k+ tokens in minutes • No top‑ups—once capped, you must wait for the next window

The Pro tier is fine for exploring, debugging, or writing tests in short bursts. The pain begins when you launch automated, headless jobs that touch hundreds of files.

The Knee‑Jerk Upgrade—Max Plan ($100 / month)

Upgrading removes most per‑window throttles, but you pay 5 × more every month—whether you need that head‑room daily or not.
(There is a $200 / month plan if you are rich enough)

Smarter Alternatives (Before You Click “Upgrade”)

Take a breather. Limits reset every few hours—grab a coffee, review diffs, or write documentation.
Switch tools temporarily. Cursor, GitHub Copilot, or local LLMs (e.g., Code Llama) can cover light edits while you wait.
Use an Anthropic Console account (API key).
- Pay‑as‑you‑go—you pay only for the tokens you burn.
- Generous rate limit—much higher than Pro — so you rarely hit it.
- Perfect for bursty weeks; ignore it when work is lighter.

Tip: Treat Console credits like cloud GPUs—spin them up for crunch time, shut them off when you’re done.

My Hybrid Workflow — Pro Plan + Console Credits

Action (What I do)	Required (Have to do)
1. Code until the Pro quota reaches its limit.	Write prompts that can resume after a cut‑off.
2. Switch `CLAUDE_API_KEY` to my Console key.	Use `claude /login` to switch accounts.
3. Finish the heavy task and check the Console cost dashboard.	Setup when your prepaid balance drops below $5, top it back up to $10.
4. Switch back to Pro once the window resets.	Keep recurring bills minimal.

Total monthly spend for me: US $20–40 in Console charges, plus the $20 Pro plan—still far below the Max plan’s $100.

Real‑World Example — Refactoring ~200 Files in One “Shot”

It wasn’t truly one shot—Claude stopped whenever the history context grew too large. I tried to avoid this without much success. If you know a better approach, please leave a comment!

claude -p "refactor_prompt.md" --dangerously-skip-permissions

-p runs in headless mode—no interactive chat.
--dangerously-skip-permissions auto‑confirms every file write.

Outcome

Phase 1 (Pro): 25 min—quota exhausted.
Phase 2 (Console): around 30 min, 8M tokens → ≈ US $5.68

Job done in under an hour with zero manual clicks.

Cost Snapshot (Typical Month)

Claude Pro subscription …………… $20
Occasional Console bursts ………… $5–10 each, a few times a month
Total ……………………………………… estimated average $40–60 (varies, only apply to my personal situation)

Savings vs. Max plan: ~ 50 %.

Final Tips to Stretch Your Tokens

Chunk big tasks. Process the codebase by directory to avoid runaway prompts.
Stream results. Pipe Claude output into git apply to keep diffs small.
*Monitor with * claude status. Know your remaining quota before batch jobs.
Cache context. Reuse summaries instead of pasting full files repeatedly.

TL;DR

Stick with the $20 Pro plan for everyday coding, burst to the Console when deadlines loom, and you’ll stay productive and well below the $100 ceiling.

Happy coding—may your tokens stretch further than your coffee! 🚀

Software Engineers, are you ready for the next‑gen challenge?

Eric Cheng — Sat, 19 Jul 2025 04:38:05 +0000

This was originally a late-night rant I posted in the Facebook group. Since it got a nice response and I’ve been meaning to try my hand at dev.to, I’m recycling it as my very first story here.

Software Engineers, are you ready for the next‑gen challenge?

Hold on, I’m not talking about vibe‑coding or AI agents!

Sure, those are cool.

In my years in the industry I’ve hopped across all sorts of companies — multinationals, mid‑sized firms, even tiny firms.

And more often than not they didn’t hire me to play with the latest cutting‑edge toys. I know those shiny tools, sure — but there’s always a queue of rock‑star devs fighting to mess with them.

Since the dawn of the internet era I’ve been chasing tech: Java, SQL, C#, JS, TS, Docker, Azure, AWS, Node.js, Express, React, MongoDB…

Learn → build → deprecate → learn again → build again — on loop.

But in the end, where does my value lie? Why do companies pay me? What do they actually want me to do?

Most of the time the answer is: Maintain and optimise legacy systems

My major value lied

Yup. Many of my previous companies know I understand the new stuff, yet they have zero intention — no budget, no guts — to adopt it.

They’re stuck with an ancient system so outdated even the least tech‑savvy user rolls their eyes at the UI and loses faith.

The functionality is fine; the look just screams “prehistoric.”

“Oh, you’re that enthusiastic new‑tech guy? Perfect. We’ve got a crusty old system — modernise it for us, will ya?”

That sort of gig took me from clueless kid to buying a house, getting married, having kids, immigrating, starting over — job after job after job, all revolving around upgrading dusty systems.

So, back to my opening: what’s the next‑gen challenge? What’s the problem that might showcase your worth?

The next‑gen challenge, and where our real values may lie

Legacy apps still need love, but another opportunity is about to explode — one tailor‑made for seasoned engineers with serious fundamentals:

Taming systems drowning in AI technical debt.

Today everyone’s bragging:

“I can’t code, but I vibed for two hours and shipped a product!”

However! Project managers buy it. Execs buy it. The business side buys it. “Hallelujah! Those smug engineers are obsolete. I’ll just hire a fresh grad who can type prompts. They’ll prototype my wild ideas and toss the mess to the dev team for cleanup.”

My goodness. That’s scarier than any legacy app, scarier than inheriting some obscure language stack.

Picture someone who’s never studied civil engineering slapping together a foundation, then stacking a 50‑storey skyscraper on top. Now the tower leaks, shorts, sways, ready to collapse.

That, my friends, is the job opportunities we’re about to see.

When their Franken‑systems start imploding and the hackers who built them get promoted — or flee — they’ll finally think: “Maybe we should hire an experienced engineer to follow up.”

The question is: Will you take the job?

Are you still dreaming of that pre‑AI golden age where engineers were rare gems? Wake up. Most of us aren’t reinventing the world; we’re cogs. And those doomed projects — crashing deadlines, burning budgets — have fed you, your family, your kids.

“Relax. Let me handle it. The system’s a dumpster fire, sure — but trust me, I’ve got this.”

Dear future me: swallow your pride, stow the profanity, and say that line to the hiring manager — gracefully.

Let’s go get it my friend.

DEV Community: Eric Cheng

My One-Month Journey with Claude Code 🚀

Claude Code Sub Agents - Burn Out Your Tokens

✨ Introduction

🧐 Sub Agents at a Glance

🎯 Why Use Sub Agents? Key Benefits

🛠️ Real‑World Case — Batch‑Updating Hundreds of Files

⏳ Previous Attempt 1 — One Huge Session

🔄 Previous Attempt 2 — Powershell Outer Loop

⚡ New Approach — One Parent, Many Sub‑Agents

💸 Token Economics: Hidden Costs of Parallelism

🏁 Conclusion

Claude Not Knowing Claude Code?

This is A Hilarious Case of Claude Not Knowing Claude: A /agents Comedy

The Setup 🎭

The (Wrong) Deep Dive 🕵️‍♂️

The Mic Drop Moment 🎤💥

What /agents Actually Does ⚡

Lessons Learned 📚

Never Get Stuck: A Practical Guide to Using Claude Code Without Blowing Your Budget

Understanding the Pro Plan ($20 / month)

The Knee‑Jerk Upgrade—Max Plan ($100 / month)

Smarter Alternatives (Before You Click “Upgrade”)

My Hybrid Workflow — Pro Plan + Console Credits

Real‑World Example — Refactoring ~200 Files in One “Shot”

Cost Snapshot (Typical Month)

Final Tips to Stretch Your Tokens

TL;DR

Software Engineers, are you ready for the next‑gen challenge?

My major value lied

The next‑gen challenge, and where our real values may lie

The question is: Will you take the job?

Claude Code Sub Agents - Burn Out Your Tokens

🧐 Sub Agents at a Glance

🎯 Why Use Sub Agents? Key Benefits

This is A Hilarious Case of Claude Not Knowing Claude: A `/agents` Comedy

Understanding the Pro Plan ($20 / month)

The Knee‑Jerk Upgrade—Max Plan ($100 / month)

My Hybrid Workflow — Pro Plan + Console Credits

Real‑World Example — Refactoring ~200 Files in One “Shot”