Engineering Principles

Status: Draft v0.1
Product: AIES — AI Engineering Standard
Updated: 2026-07-16

Purpose

These principles are the constitution of AIES. Every prompt, skill, agent, pattern, cookbook, and control should be consistent with them. If a recommendation violates a principle, the principle wins until the community revises it.

The ten principles

#	Principle	Meaning in practice
1	Human first	Humans own intent, approval of high-impact effects, and accountability. AI assists; it does not replace judgment.
2	AI assisted	Prefer AI for draft, search, and toil — with harnesses that validate outputs before they become production state.
3	Reproducible	Same inputs, pinned versions, and recorded evidence should yield reviewable results. Chat folklore is not a process.
4	Observable	Traces, logs, evals, and audits must explain what the system did and why a gate passed or failed.
5	Testable	Prompts, agents, and retrieval paths have evaluation cases. Model behavior is treated like code.
6	Deterministic where it matters	Probabilistic steps sit inside deterministic gates (schemas, hooks, policy, CI).
7	Vendor neutral	Patterns and controls must not require a single model vendor, cloud, or IDE. Integrations are adapters.
8	Secure by default	Untrusted content cannot grant authority. Least privilege, injection resistance, and approval for side effects.
9	Context rich, noise poor	Assemble the minimum high-signal package. Dumping the monorepo is not context engineering.
10	Documentation driven	The repo is the source of truth. If it is not versioned here, it is not part of the standard.

How to use principles in reviews

When reviewing a PR, skill, or agent definition, ask:

Where is the human gate for irreversible effects?
What evidence proves the change (tests, evals, plan)?
Could another vendor/runtime adopt this without rewriting the idea?
What is excluded from context on purpose?

Non-goals

Not a model benchmark league table
Not a certification body (yet)
Not an endorsement of any cloud or foundation model vendor

Compliance shape (for every AIES page)

Prefer ending operational docs with:

Section	Intent
Required	Must for the claimed scope
Recommended	Should for most teams
Optional	Risk/profile dependent
Anti-patterns	Explicitly forbidden or discouraged
Automation	Hooks, CI, CLI that enforce the above
References	Repo paths and external citations