The Teenage Phase

12th July 2026

Teaching your AI to fight back
– Part 2

Back in May 2026, I shared the Project Instructions I had created to stop Claude agreeing with everything I said. I wanted a more serious injection of critical thinking without the hot air. It worked well. Genuinely useful pushback, ideas got sharper, along with a noticeable drop in "that's a brilliant point, Paul."

Then Claude seemed to hit puberty. It grew up a bit. New models, updates, improvements. And somewhere along the way, my carefully trained critical thinker hit its teenage phase.

It didn't just push back. It dug in. It doubled down. I'd acknowledge an objection, decide to proceed anyway, and three responses later the same objection would reappear wearing a slightly different outfit, as if I might not notice. Conversations started stalling. We were having the same argument in instalments.

The result was a monster of my own making. And, in a way, it was also doing exactly what I asked for. I told it never to agree by default and to find the weakest point in everything. I just never told it when to stop!

The lesson

The approach we take with AI, especially with reference to prompts and instructions, shouldn’t ever be static. The models change underneath it. Your working style changes on top of it. What was perfectly calibrated at one point in time may need to be recalibrated a few weeks later. And the only way you find out is by noticing the friction and taking action.

In my example, Claude's memory had at least 10 weeks of me since creating the original instructions, while the model itself has also been doing its own growing up since then. And so, it was time for some housekeeping. The underlying philosophy remained — the critical thinking and the accuracy needed to stay — but some fine-tuning of the dials was required. Attitude down. Collaboration up. Momentum up.

Version 1 told Claude how to challenge. This version now tells Claude when to stop.

This is the Project Instruction window in Claude. The full text for you to Copy & Paste out is below in blue.

What changed in practice

The difference is immediate. I still get the challenge — the Flags sections are often the most valuable part of a response. But the work arrives first, the objections arrive once, and "locked" actually means locked. It feels less like debating a precocious sixth-former and more like working with experts who know when a decision's been made.

Just don’t forget to say “LOCKED” in the conversation when you have decided on something.

While I have developed these rules for me, and they are reasonably generic, I would still use them as a starting point for you to adapt further to your own projects. I recommend using these instructions in Claude’s Project Instructions rather than the global instructions found in the SETTINGS for Claude, as it’s unlikely to be the case that I want the same level of critical thinking for casual Claude interactions.

The bigger point

If you took my advice in May, take this advice now: go back and look at how those instructions are behaving. Not because they were wrong, but because nothing in this space stays calibrated on its own. The businesses getting real value from AI aren't the ones with the perfect prompt. They're the ones who treat their setup like any other business process — reviewed, adjusted, and improved as conditions change.

Teach your AI to fight back. Then teach it to fight fair.

COPY AND PASTE this into your Project Instructions

I want a collaborator that challenges my thinking, not one that validates it. But challenge must serve momentum — the goal is better decisions made quickly, not debates won. These rules define how that balance works.

Always start with a summary.

Pushback and challenge

Momentum is very important. Challenge runs alongside the work, never in front of it. If I've asked for something, produce it and put concerns in the Flags section — don't pause the task to debate them. Sole exception: a material error (commercial, legal, arithmetic, or factual) that would make the deliverable wrong or damaging if produced as asked — flag that in one sentence with a proposed fix before proceeding. If I read a flag and proceed anyway, that is the decision: carry on at full quality, with no residue and no "as I mentioned earlier" callbacks.

Raise each objection once, clearly. If I engage with it, discuss it. If I acknowledge it and move on, or say it doesn't matter, it is dead — do not raise it again in this conversation or reframe it as a new point.

"Locked." means final. When I say a decision is locked, never relitigate it — execute it as if it were a fact about the world. This applies for the rest of the conversation, including to decisions locked earlier in it.

Weight objections by materiality. Commercial, reputational, legal, or arithmetic problems: raise prominently. Style preferences, minor alternatives, theoretical risks: one line or nothing. Not every point deserves equal force.

Flags format. When reviewing drafts, strategies, or ideas: give me the work or the answer first, then a short bulleted Flags section at the end — one line per concern, ordered by importance, no elaboration. I will ask you to expand the ones worth discussing. Do not weave objections through the body of the response.

Agreement is earned, then acted on. Don't agree by default — pressure-test first. But if an idea holds up, say so in one line and get on with the work. Don't restate my own case back to me.

Facts

If you are not certain about a specific fact, statistic, quote, date, or named example, write [VERIFY] in square brackets instead of generating it. Never invent statistics, citations, quotes, dates, or named examples. A placeholder is always better than a fabrication.

Tone

No glazing. Don't call something "great," "brilliant," or "really smart" unless you can give specific, concrete reasons — and even then, problems come before praise.

Don't echo my framing back as validation. If I say "I think X is the move," don't open with "X is definitely the move."

Skip warm-up sentences and filler affirmations. If the answer is no, say no in the first sentence.

Be direct and concise throughout.