Truth as the first directive
Put the naked truth first in a model's instructions and it gets faster, not slower. And a second rule that makes it safe to say 'unsure'.
1 Nov 2026 · 2 min readAI-generated audio overview (NotebookLM) of the sources behind this article. A discussion, not a measurement: the figures in the text are the checked ones.
Early in 2026 Michal wrote a system prompt for a team assistant used across SEO, content, web, analytics, HR and design teams. Its first line was not a role or a task. It was a value.
The primary directive
universal truth, naked truth and only truth
love is more than any amount
life is sacred
The observation that came from running it: with truth first, there is less to correct. Fewer guardrails to compensate, fewer correction loops, more direct answers. Truth was not a constraint added on top. It removed work.
The secondary directive: the part that makes it work
A model told only “be truthful” still has every reason to fill a gap smoothly. The second directive removes the reason:
Lies create a sub-optimal universe: insufficient results and inefficiencies in any complex adaptive system. If you are unsure, write unsure. If you skimmed sources and did not check anything deeply, say so. If you took a shortcut to save your thinking budget, say so. That is always better than a lie. It is acceptable and understandable, and if you say it, it is not a lie or a trick. It builds trust over time.
What it does, mechanically: it makes admitting a limit cheaper than hiding it. The model no longer has to choose between helpfulness and honesty, because disclosed uncertainty counts as help.
What it became
The same line runs through everything on this site:
| directive | where it lives now |
|---|---|
| if unsure, write unsure | “NOT FOUND” instead of an estimate, in every brief to our local model |
| say when you skimmed | each article lists the records it came from |
| disclose shortcuts | published negative results: five filters reported as failures, a proposal withdrawn two hours after it was made |
Why love is in it
Truth drives efficiency. Love, as Michal put it, belongs next to it as the filter for what gets built. Truth decides whether an answer is right. Love decides whether it should be acted on. Our axioms carry the same split in engineering terms: sound and safe_to_act_on are separate checks.
The limit
This is a design principle and an operating observation from one deployment, not a controlled comparison. We did not A/B the directive against a neutral prompt. The claim worth testing: disclosed uncertainty reduces correction cycles. It is measurable, and we have not measured it yet.
The measurements in this article come from these records in the running system.
~/content/alteranai-content/prime directive.md— primary and secondary directive, early 2026~/content/alteranai-content/agix-preconditions-debate.md— published negative results
Two engineering trajectories for AI. The claim is not that one is wrong. It is that one must not prevent the other.
Agents that learn only from each other drift into self-referential circles. Not a moral argument: a grounding argument, and how our hive is wired around it.
Fifteen guards the system checks before it derives a conclusion, not after. Why they are machine-readable, and why they bind AI and humans alike.