Six months ago I started using Claude Code seriously for side projects.
I'm 16, Class 12 ISC, Guwahati. I've been building stuff since Class 9 — a crop disease classifier, a React Native app, various web projects. Claude Code was a game-changer for how fast I could ship.
The Anthropic bill was also a game-changer. Final number after six months: $13,631.
That's about ₹11.4 lakh at current exchange rates. More than my family spent on groceries in the same period. I'm not exaggerating.
The maddening part: the Anthropic billing console showed me "total tokens: ~40M" with basically no breakdown of why. I couldn't answer the most basic question: which sessions were eating the most? Which projects? Which patterns were costing me money that a different workflow would eliminate?
So I built a tool to find out. It's called Burnd. Here's what I found.
How Burnd works
Claude Code writes a .jsonl file for every session to ~/.claude/projects/. Each line is a structured event — tool calls, responses, token counts. I wrote parsers for 8 waste patterns that kept showing up in my data.
To run it:
npx getburnd
Nothing leaves your machine. No account. No API key. It reads those local JSONL files and prints your top 3 leaks.
The 8 patterns — and what they cost me
1. Long Bash output (~$31/month)
The biggest leak in my data. Test runners, build systems, and package installs were dumping 10,000–50,000 bytes of output into context on every call. That output stays in context for the rest of the session.
The fix is one line: | head -100 at the end of your bash commands. I now save around $30/month from this single change.
2. Repeated file reads (~$10/month)
Burnd found one session where the same file was read 31 times. The pattern: the agent reads a file, writes something new, then reads the same file again to verify. And again. And again.
The fix: trust Edit over Read → Write → Read. One read, one write.
3. Tool error storms (~$40/month)
When Claude Code hits a broken environment — missing dependency, wrong Node version, broken test suite — it thrashes. It calls the same failing tool 15–20 times trying different variations. Every call costs tokens.
I now spend 5 minutes fixing my environment before starting a session. Saved ~$40/month.
4. Tool overuse (Bash trap) (~$15/month)
One session had 80% of its tool calls as Bash. It was using cat to read files, find to locate things, grep for search. All of these have cheaper native alternatives in Claude Code (Read, Glob, Grep). The Bash calls bring 10× more output into context than necessary.
5. Late-night coding (~$180/month — my biggest single fix)
This one surprised me. My 00:00–05:00 sessions cost 2.5× more per task than my daytime sessions. Not because Claude charges more at night. Because I write worse prompts when I'm tired, cause more re-dos, generate more error storms.
After I started refusing to use Claude Code after midnight, I saved $180 in a single month.
6. API retry storms (~$8/month)
These are invisible from the UI. When Claude Code hits a rate limit or network hiccup, it retries. Each retry is a full context resend. Hidden in system records, impossible to see without parsing the JSONL directly.
7. Skills firing too aggressively (~$12/month)
I had a skill with a trigger pattern that matched basically every message. 42% of tool calls in certain sessions were from that one skill. The fix: tighten trigger conditions, or scope skills to specific projects.
8. Project outliers (~$30/month)
One project was costing 3.2× my session median. Burnd flagged it immediately. The culprit: a 40,000-token CLAUDE.md that was being loaded into context on every session. Trimmed it to 3,000 tokens, saved $30/month on that project alone.
Total waste in my data: $76/month
That's $76 in flagged, fixable waste from 6 months of bad habits. Real money, real fixes, all tested.
To be clear: this doesn't mean Claude Code wasn't worth it. It was worth it. But $76/month in waste from fixable habits — that's a ₹6,300/month return on spending 30 seconds running a tool.
Try it
npx getburnd
Free, MIT, local. Prints top 3 leaks. No account, no data uploaded.
For the full dashboard with all 8 detectors, spend charts, and per-project breakdown:
npx getburnd serve
# open http://localhost:4711
GitHub: github.com/garvitsurana271/burnd
Landing: getburnd.vercel.app
If you use Claude Code, run npx getburnd on your own data. Tell me what your top leak is — I'm curious whether the patterns generalize or whether my late-night-coding problem is uniquely mine.
— Garvit
Top comments (0)