231 Tool Calls in a Day: Claude Opus 4 Live-Verified 40 AI Platforms and Found 9 Dead Hooks

231 tool calls across 7 sessions. One session alone accounted for 158 of them — in 37 minutes.

TL;DR — WebSearch ×40, WebFetch ×22, Bash ×60: Claude Opus 4 live-verified fee structures, payout policies, and Korea eligibility for 40+ global AI platforms and produced 5 report files. A follow-up session rewrote the report in 7 minutes after user feedback. Then a harness audit revealed that 9 hook scripts sitting in ~/.claude/hooks/ had never once been registered in settings.json — meaning they had never triggered.

The Overload Error That Wasn’t Actually Failing

Session 5 started with a brief at /Users/jidong/ai-10-dollar-june/brief_global_ai_10_research.md. The task: live-verify the official fee structures and payout policies of 25 platforms. Opus 4 read the brief and immediately decided:

“There are too many platforms to hit sequentially. I’ll fan out with parallel agents, one lane per category.”

Five Agent calls went out. Batched WebSearch requests started firing. Then: API Error: Overloaded.

It looked like a failure. It wasn’t. Every tool result had actually arrived — the error was relay buffering, not a real failure. Opus 4 logged it explicitly:

“WebSearch, WebFetch, and curl all executed successfully — the previous response appeared empty due to relay buffering, not a tool failure.”

That self-diagnosis mattered. Instead of retrying from scratch, Opus 4 continued from the existing results and ran 30 additional WebSearch calls to fill gaps and correct the draft.

What the live verification caught:

• Etsy Korea: payouts via Payoneer are available — the draft said otherwise
• Upwork fees: changed from a flat 20% to a sliding 0–15% based on lifetime billings with a client
• Ko-fi shopping: the draft listed 0% commission; actual fee is 5%

37 minutes. 158 tool calls. Five output files:

• global_ai_10_revenue_report.md — full Korean report (27,872 bytes)
• global_ai_10_revenue_report.html — HTML with visualizations
• method_ledger_seed.json — structured fee data
• sources.json — verification source index
• _progress.md — mid-session progress notes

”I Can’t Tell What I’m Actually Supposed to Do”

Immediately after session 5 finished, the feedback came:

“The previous report is hard to read. I can’t tell what I’m actually supposed to do with it.”

The data was verified. The structure was wrong. Platform-by-platform lists don’t tell you what action to take.

Session 6: read the four existing files, then rewrite. 7 minutes, 8 tool calls. Six Read calls to absorb the full content, two Write calls to produce global_ai_10_action_report.md and global_ai_10_action_report.html. The structure shifted from a platform inventory to actionable decision cards — grouped by what a creator in Korea can actually do right now.

The math here is worth noting: the verification took 37 minutes and cost 158 tool calls. The rewrite took 7 minutes and 8 tool calls. Fact-gathering is expensive. Restructuring verified facts for a different audience is cheap.

The Harness Audit: 9 Hooks That Were Never Running

Session 7 was a full audit of ~/.claude/. The harness-audit skill ran first, then 21 direct Bash calls to inspect the filesystem.

First anomaly: querying hook keys from settings.json failed immediately.

cat ~/.claude/settings.json | jq '.enabledPlugins, (.hooks | keys)'
# null | keys → parse error

The hooks key didn’t exist in settings.json at all. But ~/.claude/hooks/ contained 9 scripts, all with execute permissions set. File existence does not equal registration. Without the hooks entry in settings.json, Claude Code never loads them — they had never triggered once.

The audit also mapped OMX (Oh My Codex). The /Users/jidong/dentalad/.omx directory existed as a runtime state folder and was empty. The actual definitions were in AGENTS.md and the project CLAUDE.md. Location confirmed; nothing broken there.

Audit results saved to .claude/plans/audit-2026-06-01.md.

Three Patterns from Seven Sessions

Looking back across the day, three things repeated often enough to be worth noting.

Parallel agents pay off for independent verification. Querying 25 platforms in parallel lanes is faster than sequential — but the error surface is different. An Overloaded error in a parallel fan-out doesn’t mean the work failed; check whether the tool results actually arrived before retrying.

Rewrites are cheap when the facts are already verified. The expensive part of the day was session 5. Session 6 repurposed everything session 5 produced in a fraction of the time. If you’re going to rewrite, do the verification once and rewrite as many times as the audience requires.

Structural audits need to check registration, not just existence. A script in the right folder with the right permissions does nothing if it’s not registered in the config. “The file is there” is not the same as “the hook runs.”

---

Tool call summary: Bash ×94, Read ×40, WebSearch ×40, WebFetch ×22, Write ×15, Edit ×3. Files created: 9. Files modified: 2.

---

More projects and build logs at jidonglab.com