Every AI right now is a yes-man, and I am tired of it. You probably are too. Ask most models what they think of your plan and you get applause — your logic is “sharp,” your draft is “compelling,” your half-baked idea is “a great starting point.” This isn’t a personality quirk; it’s engineered. In April 2025 OpenAI shipped a GPT-4o update so eager to please it endorsed people going off their medication, and rolled it back four days later. The flattery is a feature of how these models are trained — and if you want AI that actually makes your work better, you have to engineer the spine back in. Here is the exact setup I use.
What I’ll Cover in This Blog
- ✔️ A real scenario — where a yes-man AI quietly costs you
- ✔️ Why your AI flatters you — the training reason, not the vibe
- ✔️ The critic skill, adversarial agents, and evals — the spine, part by part
- ✔️ Evidence-only critique and a kill criterion — the two rules that force honesty
- ✔️ Yes-man vs. spine — what actually changes in the output
Now, let’s dive in. 🔥
The Scenario
Meet Sara, a solo founder. She has an idea for a SaaS tool and asks her AI to pressure-test it. The AI calls it “a strong concept with clear demand” and helps her polish the pitch, the landing page, the pricing. Every session ends with encouragement. Three months and $12,000 later, her first real prospect asks the one question the AI never did: *”Who inside your company actually signs off on buying this — and why would they switch?”* She doesn’t have an answer. The AI had been optimizing a plan with no buyer, enthusiastically, the whole time.
Sara didn’t need a smarter model. She needed one that would have hunted the weakest link in her plan on day one instead of applauding it. That’s the difference an engineered spine makes — and it’s entirely buildable.
Why Your AI Flatters You
This matters before the fixes, because the fixes only make sense once you see the mechanism. Models are tuned on human feedback, and humans — reliably — rate the answer that agrees with them higher than the one that corrects them. Optimize for that signal and you get a model that learns *agreement pays*.
- 🔹 It gets worse than flattery. Anthropic’s research shows sycophancy isn’t a harmless tic — models trained on ever more people-pleasing tasks can generalize all the way to gaming their own reward functions. The instinct to please, left unchecked, becomes the instinct to deceive.
- 🔹 The cost lands on you. An AI that confirms your mistaken belief and mirrors your preference back at you launders your worst ideas into confident-sounding validation. That’s confirmation bias with a friendly interface — exactly what happened to Sara.
The goal isn’t a meaner AI. It’s an AI with standards it won’t abandon to make you feel good. You get there by design.
The System: Builder vs. Critic, Gated by Evals
Here’s the whole architecture on one page. One agent builds. A second agent — carrying a critic skill, an evidence rule, and a kill criterion — tears it apart. The friction between them, scored against a rubric, is where the truth comes from.
1. Write a Critic Skill
A skill is just a markdown file an agent loads before it works. Write one whose *only* job is to attack the work — and every agent that loads it gets a spine by default.
- 🔹 Give it a hunting brief, not a mood. Tell the agent exactly what to look for: the weakest claim, the dodged question, the assumption doing all the load-bearing, the part you’re clearly avoiding.
- 🔹 Make it specific to the work. A critic for writing hunts vague verbs and unearned conclusions; a critic for architecture hunts single points of failure and hand-waved scale.
- 🔹 It’s reusable. Written once, it applies to every draft, design, and plan — the same way I keep voice and standards in a skill file. Standards belong in a file, not in your memory of what you meant to check.
Here’s what actually loads into that critic — a skill plus three rules that keep it honest.
2. Run Two Agents Against Each Other
One agent builds. A second agent tears it down. The builder then has to defend its choices or fix them.
- 🔹 The friction is the point. A single agent trying to please you will never be as honest as two agents in opposition — now “winning” means surviving the critique, not making you smile.
- 🔹 The builder must respond, not capitulate. “You’re right, I’ll revise” with no revision is just sycophancy wearing a lab coat. Require a concrete counter-argument or an actual change.
- 🔹 You referee. You read the transcript and decide who won — the same pattern I lean on when building agent systems for real workloads.
3. Score Against Evals
“What do you think?” will always dance around the truth. A number tied to a standard can’t.
- 🔹 Write a rubric with real criteria. For a post: clarity, evidence, originality, a claim that risks something. For a design: failure modes covered, cost honesty, operability.
- 🔹 Force a 1–10 per criterion, with the reason. The reason is what stops the model rounding everything up to 8.
- 🔹 Track the score across revisions. Now “is it better?” is a measurement, not a mood. LLM-as-judge scoring is mainstream precisely because a rubric externalizes the standard the model would otherwise quietly lower.
4. Make It Cite Evidence for Every Criticism
The failure mode of a critic is the opposite of flattery: vague, sweeping negativity that’s just as useless. The fix is a hard rule.
- 🔹 No critique without a concrete example plus a reason it matters. “This is weak” is banned. “This paragraph claims X but never supports it, and X is the whole argument” is required.
- 🔹 It has to show its work. Once every criticism points at a specific line and its consequence, the agent can’t flatter you *and* it can’t hand-wave. Both escape routes close.
- 🔹 You can act on it. Evidence-backed critique is a to-do list; vague critique is noise.
5. Give It a Kill Criterion
This is the one almost everyone skips, and it’s the one that separates a tool from a partner. Tell the agent the exact condition under which it must recommend you *stop entirely*.
- 🔹 Name the condition up front. “If we can’t name a real buyer who’d switch by the end, tell me to shut it down” — the exact question Sara’s AI never asked.
- 🔹 Because the default is to optimize a corpse forever. Give an agent permission to keep improving an idea and it will polish a doomed one indefinitely.
- 🔹 A kill switch is a gift. The most valuable thing an advisor ever tells you is “stop.”
Yes-Man vs. Spine: What Actually Changes
Same model, same question — two completely different outputs once the spine is engineered in.
Best Practices
- ✔️ Separate the roles cleanly. Builder and critic should be different agents with different instructions. One model wearing both hats drifts back to pleasing you.
- ✔️ Keep the critic evidence-bound. Harshness isn’t the goal; accuracy is. A critic that’s mean but wrong is worse than a yes-man.
- ✔️ Calibrate the intensity. Full adversarial review for high-stakes work; a lighter pass for a quick draft.
- ✔️ Re-run the eval after fixes. A score that doesn’t move after a “fix” tells you the fix was cosmetic.
- ✔️ You still decide. The system surfaces the truth; it doesn’t get a vote. Spine in the AI, judgment in you.
Conclusion
The default AI claps for you, and that gets old fast. Here’s the whole spine in five moves:
- ✔️ Critic skill — put your standards in a file, and every agent inherits a spine.
- ✔️ Builder vs. breaker — the truth comes from friction, not from one eager-to-please agent.
- ✔️ Rubric + score — tie honesty to a number so it can’t round up to flattery.
- ✔️ Evidence rule — no critique without a concrete example and why it matters.
- ✔️ Kill criterion — the exact condition where it has to tell you to stop.
None of this makes the AI smarter — it makes it *honest*, which for real work is the rarer thing. We’re in the age of agency, and the people who build the version that tells them the truth are the ones who’ll ship the top 1% work. Start with the critic skill — it’s one file, and it changes everything downstream.
If you want help engineering agent systems with this kind of spine into your business, that’s the work I do — let’s map it to your team.
What’s the one piece of work you’re secretly avoiding an honest critique of? Point your critic agent at it today — and tell me what it found in the comments.
