On May 4, 2026, a16z co-founder Marc Andreessen posted his personal AI system prompt on X. It opened by telling the model it was "a world class expert in all domains... on par with the smartest people in the world" and closed with this:

"Verify your own work. Double check all facts, figures, citations, names, dates, and examples. Never hallucinate or make anything up. If you don't know something, just say so."

The "never hallucinate" line got him roasted. Tech journalist Karl Bode put it best on Bluesky: "Yes, you can just demand that the LLM not make errors. That's definitely how the technology works."

Here's the thing the dunking missed, though. Andreessen had the right instinct and the wrong mechanism. He's doing real prompt engineering, badly. And the fix is four lines you can paste into any model right now.

Why "never hallucinate" is an instruction with nothing behind it

A language model does not check a fact before it answers. There is no internal ledger of true statements it consults, no verification pass a prompt can switch on. It predicts the next token from patterns in its training data. That's the whole machine.

So "never hallucinate" is a command pointed at a step that doesn't exist. It's like writing "never get a flat tire" on the dashboard. The instruction names an outcome without touching anything that produces it.

This isn't a tuning problem that better models will grow out of. OpenAI's September 2025 paper Why Language Models Hallucinate (arXiv:2509.04664) proved a lower bound: the hallucination rate is at least twice the misclassification rate, minus some correction terms. An earlier result, Calibrated Language Models Must Hallucinate, showed the floor holds regardless of architecture or how clean your data is.

The core finding is the part worth internalizing. Models guess because guessing scores well. On most benchmarks, "I don't know" earns the same zero as a wrong answer, so a confident wrong guess is never penalized more than an honest abstention. The model learned to bluff because bluffing was free.

You can't scold that away. You can change the incentive at the prompt level.

The naive version vs. the scaffold

This is what most people write:

Answer accurately. Never hallucinate or make anything up.
If you don't know, say so.

It feels responsible. It does almost nothing, because "don't know" still has no consequence and the model still has no reason to prefer it over a fluent guess.

Here's the replacement. Four lines, each one wired to a specific failure mode.

The Prompt:

Use ONLY the sources I paste below. If the answer isn't in
them, reply exactly: "Not in the provided sources."

Every factual claim must cite the source it came from, by
name or number. No citation means you don't make the claim.

Tag each claim with your confidence: [high] [moderate]
[low] [unknown]. Base it on how directly the source supports it.

If the question is outside what these sources cover, say so
and stop. Refusing is a correct answer, not a failure.

SOURCES:
[paste your text here]

QUESTION:
[your question]

Why This Works: Each line removes a different path of least resistance. Grounding ("use only these sources") stops the model from blending remembered training data into your answer. Cite-or-abstain forces traceability, so a claim with no source attached never gets written. Confidence tags give you a filter you can scan. And explicit permission to refuse converts "I don't know" from a penalty into an accepted move, which is the exact lever the benchmarks were missing.

Expected Output:

The contract's termination clause requires 60 days' written notice [Source 1, high]. Either party may terminate for material breach after a 30-day cure period [Source 1, high]. The governing law appears to be Delaware [Source 2, moderate] though the clause references "the State" without naming it explicitly, so treat that as inferred.

Not in the provided sources: whether the auto-renewal can be waived. That term isn't covered in what you pasted.

Notice what the model did at the end. It refused, by name, on the one thing it couldn't support. That refusal is the product working, not breaking.

What the numbers say about each line

The lines aren't equal. Ranked by how well the research backs them:

Grounding is the heavy lifter. Forcing the model to answer from provided context instead of memory cuts hallucinations 30 to 50 percent across typical business use. Retrieval-backed setups report 70 to 80 percent fewer. Phrasing matters more than you'd expect: "answer using ONLY the provided text" beats "answer using the text below," because the softer version reads as an invitation to sprinkle in remembered facts.

Permission to abstain is the sleeper. On the AA-Omniscience benchmark, Claude 4.1 Opus reportedly hit a 0% hallucination rate by refusing to answer when it wasn't sure. A 2025 Nature study found prompt-based mitigation cut hallucinations by about 22 percentage points. The mechanism is the incentive fix: once refusing is allowed, the model stops needing to bluff.

Cite-or-abstain works, with one sharp caveat. Asking for citations only helps when the sources are real and in front of the model. Point it at its own memory and adversarial tests have pushed citation fabrication as high as 94 percent in worst-case conditions. The model will happily invent a plausible-looking source. Cite-or-abstain is powerful glued to grounding and dangerous without it.

Confidence tags are useful but not self-sufficient. A model doesn't reliably know when it's hallucinating, so its self-reported confidence isn't gospel. Treat the tags as a triage filter on top of grounding, the thing that tells you where to double-check, not a guarantee.

This is not theoretical hygiene. The Damien Charlotin AI Hallucination Cases database now tracks more than 1,600 court filings built on AI-fabricated citations, growing five to six a day. In Oregon this year, two lawyers drew a $110,000 sanction over 23 fake citations and 8 invented quotations. And 59 percent of the documented cases involve people representing themselves, the ones who turned to AI because they couldn't afford a lawyer and then had no way to catch the confident fabrication. The people leaning hardest on these tools are the least equipped to spot when they're being lied to. That's the whole reason the scaffold matters.

The fifth line you'll be tempted to add (don't, or do it carefully)

There's an obvious candidate for a fifth line: "estimate the answer yourself before you look at any number I give you, so you don't get anchored." It sounds rigorous.

The research is genuinely mixed on it, and I'd rather tell you that than sell you a clean five-step list. Across models and prompts, answers drifted toward a supplied anchor 88 percent of the time. "Estimate first" prompting actually increased mean bias by up to 81 percent on one smaller model, while barely helping others. Stronger models aren't more resistant; they get swayed by "expert" framing even when told to ignore it.

If estimate-before-anchor moves your numbers, keep it. But it's model-dependent and unreliable on its own. The four lines above do the real work. Don't let a tidy-sounding addition crowd out the parts that are actually load-bearing.

That's the throughline. You can't command honesty out of a next-token predictor. You can engineer the conditions where guessing stops being the easy path: take away the reason to guess (let it abstain), fence in the input (ground it), and demand traceability (cite or stay silent). Andreessen wanted the outcome. The scaffold builds the conditions that produce it.

If your team is shipping AI features and keeps getting bitten by confident-but-wrong output, this is a skill you can train into people in an afternoon. Want hands-on prompt engineering training for your team? Connect with Kief Studio on Discord or schedule a session.

confidence prompting

Marc Andreessen Told His AI to 'Never Hallucinate.' Here's the Four-Line Scaffold That Actually Works