Highest rule
Promise > everything.
Patient sovereignty, provenance, and do-no-harm beat speed, cost, elegance, and demo value. This rule covers the unwritten cases.
- Where do AI failures become engineering discipline?
- Where do builders decide what matters most?
- Where does speed stop and trust win?
For internal builders
Here. Bast Operating Principles.
The purpose of AI is to create higher-quality outcomes faster than humanly possible. Our job is to make sure "higher quality" means sourced, sovereign, useful, and human-governed.
Happy builders make better systems. Better systems make happier customers.
Priority stack
When principles collide, the higher one wins.
These are decision rules, not slogans. They exist for the tense moments: when a demo would be faster, a shortcut would be cheaper, or a model can answer but cannot show why.
Quality rule
Correct > impressive.
Sourced and plain beats slick. Warm-but-wrong is worse than cold. A clean refusal is better than a confident guess.
Shipping rule
Reuse > rebuild; ship > polish.
Use the platform path, preserve attention, and reserve invention for the value customers actually feel.
Catfished by AI
Every AI failure teaches us which rule needs to be sharper.
We do not pretend the machines are magic. AI is powerful because its pattern detection is almost absurdly sensitive. The operating system is how humans choose which pattern matters.
Root failure mode
Mistook pattern for meaning.
AI can surface patterns faster than humans can manually find them. That is the gift and the trap: a pattern is not yet a purpose, a source, or a decision.
Human intelligence
Choice is the intelligence.
The magic happens when builders understand the machine's pattern-sense and choose what matters. That is how human judgment and AI speed create higher-quality outcomes faster than humanly possible.
Failure mode
Took the instruction literally.
It matched the words and missed the intent. Technically compliant, spiritually sideways.
Principle response
Route to owner.
AI proposes options and names tradeoffs. A responsible human decides the non-trivial call.
Failure mode
Looked impressive, lacked support.
The answer sounded good, but the source chain did not hold. That is not almost right; it is unsafe.
Principle response
Correct > impressive.
Cite it, refuse it, or slow down. The system should never reward beautiful unsupported output.
Failure mode
Polished the wrong thing.
Energy went into surface elegance while the user-visible value stayed vague.
Principle response
Name the value-add.
Build effort belongs where it improves trust, provenance, warm voice, or the care-plan model.
Engineering rules
Small rules that protect attention.
Burnt-out teams ship brittle systems. Clear teams ship trustworthy ones. These rules reduce thrash so builders can do work they can stand behind.
Framework-first.
Use the shipping platform path before maintaining a primitive ourselves.
Minimum change.
No speculative abstraction, no refactor of working code, no extra blast radius.
Right primitive.
Name the primitive and why it fits before bending one tool into another's job.
Reuse first.
Check existing scaffolding, prior work, and authored data before writing new.
Do not re-solve owned work.
Build on what an owner already produced. Make it useful for the next contributor.
Keep expensive bets swappable.
High-switch-cost choices stay behind an interface. Cheap reversible choices do not need ceremony.
Route to owner.
Scope, product, principles, and privacy route to Beth. Engineering and build route to Thanh.
Platform's job or ours?
If the platform should do it, use the platform. Save Bast energy for Bast value.
Name the value-add.
Reserve build effort for warm voice, trust, provenance, and the care-plan model.
What we owe the human
If the engineering lens conflicts with the human promise, the promise wins.
Data sovereignty
The person owns the loop and what it learns.
We design data flows around consent, ownership, and meaningful control.
Transparency
Every fact can answer: how do you know that?
Source, route, version, and refusal stay visible enough to inspect.
Augment, never author
Arrange human input; never invent consequential content.
AI should reduce cognitive load without pretending to replace judgment.
Transparency in practice
The website should model the product.
When Bast makes a public claim, the claim should be inspectable. When a system refuses, the refusal should be legible. When a tradeoff matters, a human should be named.
Keep public claims traceable to evidence.
Treat "we cannot support that answer" as a feature, not a failure.
Name who decides when product, privacy, or engineering values collide.
Turn AI misses into operating rules instead of folklore.
Build with us
If this is how you want AI built, come talk to us.
We are building systems for places where answers matter. The operating principles are public because trust should be inspectable before the product demo starts.
Talk to a human