97 Tool Calls in One Session — Shipping a GPT-5.5 Analysis to 3 Platforms with Claude Code

That’s how many tool calls Claude made in a single session. 29 hours and 39 minutes: 42 Bash calls, 11 Telegram responses, 8 Agent dispatches. It all started with a one-line Telegram message: write a blog post about GPT 5.5 and duct tape.

TL;DR Split a GPT-5.5 (codename Spud) and Duct Tape (GPT Image 2) analysis into two parts and published them simultaneously across three platforms. The critical move: catching a duplicate post from 8 days earlier before it caused SEO conflicts.

One Telegram Message, 30 Hours of Work

The message was short. Claude picked it up, fired the auto-publish skill, and started researching immediately.

GPT-5.5 and Duct Tape are unrelated projects. GPT-5.5 (codename Spud) is a model released on 2026-04-23. Duct Tape is the internal codename for GPT Image 2, which was tested on LM Arena under the aliases packingtape and maskingtape. The framing angle: “OpenAI’s April two-handed launch: reasoning + image = super app puzzle.”

Research results and a proposed structure were sent to Telegram together — no waiting for approval before locking in the outline.

Duplicate Detection Changed the Strategy

Before pulling references, existing publication history was checked first. That’s where the critical finding came in.

Eight days earlier (2026-04-16), a “OpenAI Duct Tape / GPT Image 2” post had already been published to all three platforms. That topic was already deeply covered. Writing the same thing again would be straight duplicate content.

Strategy adjusted: GPT-5.5 (Spud) as the main topic, Duct Tape connected via internal links to the existing post. Canonical URL unified to https://jidonglab.com/blog/openai-gpt-5-5-spud.

After a user request to split into two parts, the content was divided into Part 1 (GPT-5.5 technical analysis) and Part 2 (OpenAI Super App roadmap).

4 Agents in Parallel

Once 5 references were loaded, 4 agents were dispatched simultaneously:

1. @content-writer → "Part 1 ko (spoonai.me)"
2. @content-writer → "Part 1 en (spoonai.me)"
3. @content-writer → "Part 1 DEV.to (English)"
4. @content-writer → "Part 1 Hashnode (English)"

Sequential would have meant 4× the time. Parallel means total time equals the slowest agent’s runtime.

One small bug surfaced during the review pass: the DEV.to description was 156 characters — one over the 155-character limit. Trimmed and pushed.

Publishing Flow

After Part 1 validation, Part 2 went through the same process. Final commit sequence:

# spoonai-site
git add content/blog/2026-04-25-openai-gpt-5-5-spud-{ko,en}.md
git commit -m "feat: GPT-5.5 Spud analysis (Part 1 ko/en)"
git push origin main

# dev_blog
git add posts/2026-04-25-openai-gpt-5-5-spud-en.md
git push origin main

Status reports went to Telegram at each stage. Errors triggered immediate Telegram alerts, which meant the session could run unattended.

Tool Distribution

Tool	Count	Purpose
Bash	42	git push, file checks, metadata queries
Telegram response	11	Progress reports, structure proposals
TaskUpdate	11	In-session task state updates
Agent	8	Parallel content generation
TaskCreate	7	Subtask creation
Read	7	Reference and existing file review

Bash dominates at 42 because git operations and filesystem checks run repeatedly throughout the session. The actual content generation was handled by 8 Agent calls.

What Skipping the Duplicate Check Would Have Cost

The same topic would have been published twice. From an SEO perspective, canonical conflicts. From a reader perspective, redundant content. Checking publication history before research is the sequence that matters.

The auto-publish skill’s Phase 1 (source analysis) enforces this. Keyword input triggers an existing publication history check. Follow the skill, skip the mistake.

More projects and build logs at jidonglab.com