I Asked Each AI How to Squeeze the Most Out of Its Rival

I paid $480 a year for Claude Pro and ChatGPT Plus.
I was pretty sure I was only getting about 30% of the value.

So I ran a simple experiment: I asked each AI how to squeeze the most out of the other one.

I expected both to quietly promote themselves.
Instead, one of them told me I was probably wasting most of what I was paying for.
The other corrected a basic mistake in my setup before it even answered the question.
And both openly admitted where their rival was better.

No brand protection. No false modesty. Just the closest thing to honest advice you can get from a model that has nothing to gain by flattering itself.

Here’s exactly what they said — including the parts where they recommended the competition.

What ChatGPT Said About Claude

(I told ChatGPT my setup: Claude Pro, Claude Code, working in investing, English blog/tools, and high school education.)

ChatGPT’s opening line: “You’re probably using about 30% of what you’re paying for.”

Then it gave me eight principles.

1. Stop micromanaging. Give Claude goals, not steps.

Don’t say: “Help me write HTML.”

Say: “Design an SEO-optimized investing tool. Analyze competitors first. Then design the UI. Then write the code. Then test it yourself.”

The output is completely different. Claude performs better with autonomy than with step-by-step instructions.

2. Give Claude fixed job titles — not just questions.

ChatGPT suggested I assign Claude four permanent roles:

Chief Investment Analyst — every day, ask: what changed about moat / industry structure / financial quality / 3-year earnings? Give me a hold / watch / re-research verdict. Nothing else.
English Blog Editor — for every article: SEO score, E-E-A-T score, Reddit/Hacker News/Medium distribution potential, whether it’s worth translating or recording as a podcast.
Product Manager — ask as a Google PM: why would someone open this tool, bookmark it, share it, come back, recommend it, pay for it? Only keep features that survive all six questions.
High School Exam Designer — not a teacher. A gaokao question writer. What does this concept test? Where do students most often fail? Design the animation.

3. Claude Code is where the real money is.

Most people who pay for Claude Pro spend it on chat. ChatGPT said the actual value is Claude Code — give it the whole project, not individual files. Let it audit, fix, test, and commit without you managing each step.

4. Build a CLAUDE.md file. 90% of users don’t do this.

A project-level instruction file that tells Claude your goals, your audience, your tech stack, your style rules. Once it’s there, you never explain yourself again. Every session starts with Claude already knowing the rules.

5. Clear context between tasks. Don’t let sessions bloat.

One long conversation across 300 rounds makes Claude progressively worse. One project, one conversation. When a task is done, /clear. Start fresh. Claude’s performance degrades as the context window fills.

6. Build a Prompt Library.

Stop rewriting the same prompts. Save your investment review prompt. Save your article audit prompt. Save your lesson design prompt. It’s infrastructure, not convenience.

7. Use Claude as executor. Use ChatGPT as challenger.

Claude generates. ChatGPT finds the holes — logic errors, weak assumptions, SEO gaps, investment biases. This is cross-review, and it’s more reliable than either model alone.

8. The full weekly split.

Time	Claude handles	ChatGPT handles
Morning	Read filings, extract long-term logic	Challenge conclusions, add counterpoints
Afternoon	Build tools (Claude Code)	Review from user value / SEO / product angle
Evening	English blog draft, podcast script	Polish language, check E-E-A-T, distribution strategy
Weekend	Refactor site, automation scripts, education tools	Propose new product directions, monetization ideas

What Claude Said About ChatGPT

(I told Claude my setup: ChatGPT Plus, Codex, same three domains.)

Claude started with a correction: what I called “Pro” is actually Plus. The real Pro is $200/month. The distinction matters because Plus has a rate limit — roughly 160 flagship messages per 3 hours — and once you hit it, the model drops down automatically. So the entire strategy changes: spend expensive compute on things that actually require it.

1. Use Projects as three permanent workstations.

Build separate Projects for investing / blog+tools / education. Upload your frameworks, custom instructions, and key files into each. Never re-explain your context again. (This is the ChatGPT equivalent of CLAUDE.md.)

2. Turn your recurring roles into Custom GPTs.

If you role-play the same setup repeatedly — “you are a gaokao question designer” or “you are a moat analyst who never makes up data” — that’s a Custom GPT. Build it once, use it forever, share it with others.

3. Deep Research is a cross-validator, not a chatbot.

Plus includes Deep Research with limited uses per month. Don’t waste it on basic questions. Use it to fact-check AI-generated conclusions against primary sources — earnings reports, official announcements. Claude once caught Gemini fabricating a premise. Deep Research is exactly the tool for that kind of audit.

4. Scheduled Tasks for your morning dashboard.

Plus supports timed automation. Set it to pull your tracked sector signals every trading day morning, generate a draft summary, and have it waiting when you wake up. You review; it researches.

5. Budget your rate limit intentionally.

Drafts, translations, formatting: use mini or low reasoning. Financial statement audits, logic-checking, complex reasoning: use high reasoning mode. Heavy long-running tasks: send to Codex Cloud, which runs independently from your chat quota.

6. Codex is separate from your message quota. Use it differently.

Quick local edits: local CLI. Long tasks, batch work, anything that takes more than a few minutes: Codex Cloud at chatgpt.com/codex. It runs in the background, you preview the diff when it’s done, then merge.

7. For investing: AI as auditor, not author.

Claude’s exact words: “Strictly ban it from writing investment conclusions — that’s exactly where fabrication is most common.” Use it to audit your own reasoning and flag red-line violations. Upload your position snapshot and ask it to find internal contradictions you’ve missed. Never ask it to generate an investment thesis.

The Honest Summary

After running both conversations back-to-back, the sharpest insight wasn’t any single tip. It was this:

They described each other’s weaknesses accurately.

ChatGPT said Claude performs best when given autonomy and clear constraints — and that Claude Code is underused. That matches everything I’ve experienced.

Claude said ChatGPT Plus has a rate limit that degrades performance if you’re not careful — and that Deep Research is the most underused feature. Also accurate.

Neither of them would have said this about themselves.

The cross-examination worked.

If you’re paying for one of these and not the other: the ChatGPT tips for Claude work even if you only use Claude. The Claude tips for ChatGPT work even if you only use ChatGPT. The individual principles hold up on their own.

If you’re paying for both: the split is clear. Claude builds. ChatGPT reviews. You decide.

This is part of my Learning in Public series — what I’m actually figuring out, in real time.

The Task Matrix: Which AI Won Each Category

After both conversations, I mapped what they actually said — including where each one deferred to the other:

Task	Winner	Rating	Why
Coding / autonomous dev	Claude	⭐⭐⭐⭐⭐	Claude Code handles multi-file edits end-to-end
Deep research / fact-checking	ChatGPT	⭐⭐⭐⭐⭐	Deep Research feature cross-verifies sources
Long-form writing	Claude	⭐⭐⭐⭐⭐	More nuance, less AI-sounding output
Investment analysis audit	ChatGPT	⭐⭐⭐⭐	Better at checking claims against primary sources
Tool & product design	Claude	⭐⭐⭐⭐⭐	Stronger at multi-step autonomous execution
Scheduled automation	ChatGPT	⭐⭐⭐⭐	Scheduled Tasks built into Plus
Education content design	Claude	⭐⭐⭐⭐⭐	Structured layered explanations, exam-style thinking
Brainstorming / challenge mode	Either	⭐⭐⭐⭐	Use both — different gaps surface with each model

The pattern is consistent: Claude wins on building and writing. ChatGPT wins on verification and research. Neither dominates everything.

The Real Lesson Isn’t Which AI Is Best

I started this experiment looking for a winner.
There isn’t one.

The real finding is simpler and more useful: AI becomes dramatically more valuable when models are forced to challenge each other instead of selling themselves. The moment you remove the incentive to self-promote, you get something closer to the truth.

Claude builds.
ChatGPT reviews.
You decide.

The individual tips still work even if you only pay for one of them. But if you pay for both, the highest-leverage move is not to pick a favorite — it’s to make them argue.

This is the first post in a series I’m calling AI Cross-Examination: letting models evaluate each other instead of themselves.

Next up in AI Cross-Examination:

I Asked 4 AIs Which One Lies the Most. They Named Names.
I Fed the Same Investment Thesis to 5 Models. Only One Said “Don’t Buy.”
I Followed Both AIs’ Advice for 14 Days. Here’s What Broke.

If you want the unfiltered versions, follow along.