4 Sessions, 280 Tool Calls: Using Claude Code Beyond Coding

280 tool calls. 4 sessions. Bash 159 times, Read 42, Agent 19. And none of it was what most people think Claude Code is for.

This week: draft a payment processor contract review reply, auto-generate 288 SEO landing pages, debug a UI issue that turned out to be a missing type field from the start. One session ended in 7 tool calls because the environment wasn’t set up right. Claude Code is only as useful as the context you give it — and the permissions you grant.

TL;DR Claude Code handles business admin work, not just code. Subagents keep the main context clean. Missing system permissions kill sessions faster than bad prompts.

Claude Drafted the Toss Payments Contract Reply

Session 1 wasn’t a coding session. Toss Payments (Korea’s Stripe equivalent) sent a contract review email asking for business info, refund policy, and product pricing. The prompt was minimal:

커피챗 프로젝트 접근해줘

Then I pasted the email. Claude read /Users/jidong/projects/coffeechat/, identified the stack (Next.js 16 + Supabase + Toss Payments), price structure, and where the refund policy lived. Draft reply ready in minutes.

It didn’t stop there. When it found the business registration number field blank in site-config.ts, it added a conditional render to hide the number until it’s filled — without me asking. Then it updated Footer.tsx to include a business address and phone number formatted to card issuer audit standards.

I never said “fix the footer.” Claude read the context, identified what a payment processor would flag, and moved first.

Bash 42, Read 16, Edit 5. 27 minutes.

Without Browser Permissions, Claude Is Flying Blind

Session 2 was a dead end. I asked Claude to review the spoonai mobile design, but it didn’t have macOS Accessibility or Screen Recording permissions. The computer-use tool was available, but system permissions weren’t granted.

computer use 열려있는데? chrome mcp는 왜 없어?

Chrome MCP wasn’t installed in this session. WebFetch converts HTML to Markdown — it doesn’t capture real layout. When computer-use can’t see the screen, the only options are: grant the permissions in System Settings, or manually screenshot and paste.

7 tool calls. Session over.

The lesson: Claude’s capability ceiling is defined by your environment setup. A smart model running blind is still blind.

288 SEO Pages, Distributed Across Subagents

Session 3 was the week’s main work. The goal: implement 288 zodiac compatibility SEO landing pages for saju_global. Bash 111, Read 20, Agent 17, TaskCreate 7, TaskUpdate 14. 182 total tool calls.

Started with planning. The writing-plans skill ran against the spec file at docs/superpowers/specs/2026-04-09-seo-compatibility-pages-design.md and produced a structured implementation plan saved to docs/superpowers/plans/2026-04-10-seo-compatibility-pages.md.

Then subagent-driven-development kicked in. Each independent task dispatched to its own subagent. After each task: spec compliance review → code quality review. Two-pass verification, every time.

Content generation ran in the background:

nohup npx tsx scripts/generate-compat-content.ts > /tmp/compat-gen.log 2>&1 &

While the background task ran, the main thread kept working on other things. When I asked “how’s it going?”, Claude read the log file and reported progress. Final output landed in apps/web/data/zodiac-compat-content.json.

The subagent pattern’s core value: main context stays clean. File exploration, implementation, and verification each run in isolated agents. The main thread only sees summaries — not 300 lines of grep output.

The Archive Images Were Never There

Session 4. Images weren’t showing in the spoonai archive. Looked like a rendering bug. It wasn’t.

Read and Grep into the codebase: ArchiveEntry in lib/types.ts had no image field. The getArchiveEntries() function was explicitly dropping meta.image and only passing through date, title, and summary. ArchiveList.tsx rendered text-only cards — no thumbnail slot at all.

It wasn’t “images not showing.” The code was never built to show them.

Bash 6, Read 6, WebFetch 2, Grep 1. 1 hour 39 minutes to fully understand the situation.

Week Summary

Session	Work	Tool calls	Time
1	Payment contract reply + footer fix	73	27 min
2	Mobile design review attempt	7	—
3	288 SEO pages: plan + generate	182	105h cumulative
4	Archive image issue diagnosis	16	1h 39min

Using Claude Code only for coding leaves half its value on the table. Business context, contract replies, batch content generation — it handles all of it if you give it the right context and the right access.

Subagents are non-negotiable for large tasks. Split the work, receive only results. That’s how you ship quality without burning context tokens on noise.

---

More projects and build logs at jidonglab.com