Code that tells you when it is wrong

An agent will finish a task, report success, and be completely wrong - not because it lied, but because nothing in its loop could ever have found out. Every law here is one way a system was made able to find out. All eighteen were a real failure first.

Get started View source

bash

# the codex, the doctrine and 70+ skills - one clone
git clone https://github.com/Aitherium/awskills
cd awskills && bash scripts/install-awskills.sh

# then, to your agent:
"use the code-like-david skill"

19laws

4chapters

MITlicense

0dependencies

The codex, in full

Four chapters to get from nothing to a working setup, then eighteen laws. Filter by part, or search for the failure you are actually looking at. Every row opens the document.

codex/23 of 23

What this actually isan agent is a loop with tools, not an oraclestart here Your first hourone real change, verified, on your own repostart here Run it on your own hardwarewhat a local model is actually good atstart here Give it handstools, skills and packsstart here A rule nothing asserts is a suggestionthe first law, and the one the other seventeen depend onenforcement Make it a check, not a ticketa mechanically-detectable defect must never be discharged by writing it downenforcement Watch your gate faila checker nobody has seen fail is not a gateenforcement Mutate the test, not just the codea fixture that cannot fail proves nothingenforcement Design for the silencethe expensive failures do not raisesilence A check that cannot run must not passsilence is not a verdictsilence The symptom names the innocentfollow the wire, not the error textsilence A checker in the wrong place found nothingand that is indistinguishable from a clean passsilence Detection without delivery is not detectionthe alert path is the one thing you cannot verify by using itsilence A gate that floods gets switched offbeing right too loudly is how a rule diesadoption Open green, ratchet downa gate that opens red gets bypassed, not satisfiedadoption Measure it againevery number in your docs is decaying right nowadoption Written is not deployedask the running thing, never the sourcedeployment You wrote it; that does not mean it shipsa file is tracked when the tooling says sodeployment Generate, never copya comment asking people to keep two copies in step is not a gatedeployment The defect lives in the unionevery file is individually correctdeployment Fail closed, then prove the happy patha gate that denies everything passes every denial testtrust Never trust the caller for an authorization decisionthe payload is not evidencetrust Count the closure, not the edgean import count is not a portability measurementtrust

How to read it

The laws are grouped into five parts. If you only have time for one, read Silence - those five cost the most per incident, because nothing tells you they happened.

start

Start here

Never used a coding agent? Four chapters, about an hour, ending with one real verified change on your own repository. Covers local models and how to give an agent hands.

Enforcement

A standard nothing checks is a preference. How a preference becomes a property of the codebase - and why a checker nobody has watched fail is not a gate.

Silence

The expensive failures do not raise. They return 200, render correctly, log nothing, and leave the container healthy. A missing thing is indistinguishable from a thing nobody wanted.

III

Adoption

A gate only works while people keep it switched on. Most gates die of being right too loudly - so narrow the rule, pin the count, ratchet down.

Deployment

The gap between the code you wrote and the code that is running. A build can succeed and ship nothing; a live mount makes the file current and leaves the process stale.

Trust

Two patterns no linter will catch, because both are semantic: the gate that fails open, and the decision keyed on input the caller chose.

Quick start

The codex is markdown. Clone it, read it, or point your agent at it - the laws are written to be loaded by an agent as easily as by a person.

bash

git clone https://github.com/Aitherium/awskills
cd awskills

# the reading path
ls codex/path/

# the eighteen laws
ls codex/laws/

# install the doctrine into your agent's config
bash scripts/install-awskills.sh

The laws are descriptive, not aspirational. Every number here was counted on a running system on the day it was written, and several of the incidents are ones where the first diagnosis was also wrong. Where a law contradicts something you know about your own stack, believe your stack and re-measure - that is literally law twelve.

The whole method is four steps

Not a reading list. The laws exist to be installed, and this is the loop that installs them. A codebase run this way gets harder to break over time without anyone having to be careful.

1 · Something breaksA real failure, in production or in review. Not a hypothetical.

2 · You fix itRoot cause, not a workaround. A workaround that cannot be avoided gets recorded with a date and a severity.

3 · You ask one questionCould a check have caught this? This is the step everyone skips, and it is the only one that compounds.

4 · If yes, the check IS the workWrite it. Watch it fail on purpose. Wire it somewhere unattended. Do not write a ticket - a ticket is the cheapest possible response to finding a defect, and it is why the same class recurs for weeks.

One estate, 7 doors

These projects are organs of the same system — the same identity plane, the same agent runtime, the same operational discipline. Follow any of them.

the osaitherium.comThe Living OS these four repos are cut from. Open, no login.pythonAither ADKBuild AI agent fleets — 3 lines, any backend, local or cloud.markdownawskillsPortable agent skills — self-contained procedures an agent loads on demand.cudaaitherkvcacheNear-optimal KV cache quantization for LLM inference — sub-byte compression.powershellAitherZeroPowerShell 7+ automation framework — numbered, self-describing scripts.writingThe BlogHow all of it is actually built, measured and operated.pythonawrunA priority-aware queue and dispatcher for agentic runs and ad-hoc CI builds.

The ecosystem this belongs to

Every Aitherium repository publishes an aither-manifest.json beside its page, and each surface reads all of them — so the network is browsable from any node in it. A repository that publishes nothing shows as unknown, never as zero.