Skills fire, then don't fire, then fire wrong.
Models ignore rules. Skills get loaded but the agent reads the wrong path. Prose constraints dilute under context compaction. The output looks plausible — until someone reads it carefully.
When the rig you built starts breaking silently
Make silent failures visible.
AI work can fail without an obvious error: the output looks plausible, the setup depends on one operator, and the missing record appears only when someone asks what happened.
Three ways the rig breaks
Three failure patterns worth testing for.
The setup is locked in your head.
Your screenshots. Your private docs. Three Discord threads. The agent works because you know how to prompt it. None of that survives the first time someone else has to run it.
You stopped writing skills as prose. You started orchestrating skills with skills.
Lifecycle skills. Trigger Collision. Silent Scope Decay. The fix isn't more orchestration — it's a contract the orchestration is checked against.
The trigger
The first time you have to defend it.
A client asks why the agent ignored their last instruction. A partner asks where the audit trail is. A regulator asks for the chain of custody. An examiner walks in.
A setup that depends on one person's memory is difficult to defend in that conversation. The next step is a written work contract that defines inputs, permissions, stop conditions, review points, and evidence.
What changes
The next move is a work contract.
When skills depend on other skills, define the contract they are checked against: sources, permissions, stop conditions, review points, and required evidence.
The consulting engagement maps that contract and builds the first agreed piece inside your tools. The aim is to surface unresolved work for review instead of hiding it behind plausible output. Scope and fee are set in the engagement letter.
Have us install it.
If you have one workflow to review, bring it into the One-Workflow Plan. If you are still naming the workflow, use Claude Cowork first.