Several patterns share adjacent territory. Here is what came before, how SEW differs, and what appears to be genuinely new.
The closest prior art
CrewAI Task Guardrails (2024–2025) is the nearest existing pattern. It supports embedded validation criteria, natural-language checks checked by an LLM, and automatic retry up to 3 times on failure — the same default as SEW.
| CrewAI Task Guardrails | Self-Enforcing Workflows | |
|---|---|---|
| Who calls the checker | The framework, automatically | The skill itself, explicitly |
| Where criteria live | Passed by caller at task creation | Inside the skill's own definition file |
| Universal auditor | No — each task has its own guardrail | Yes — one auditor reads any skill's DoD |
| Quality system audits itself | No | Yes — meta-skill has its own DoD |
| Repairs bad criteria | No | Yes — silently, via review-definition-of-done |
The most important difference is who calls the checker. In CrewAI, the framework weaves in the guardrail automatically — the task author doesn't write the call. In SEW, the skill author explicitly includes the audit call in the skill body. This makes the quality contract visible and auditable by anyone reading the file, and it makes the skill's commitment to self-verification explicit rather than implicit.
Other adjacent patterns
In Scrum, a DoD is a shared checklist a team uses to determine when an item is truly complete. SEW borrows the term and concept directly. The key difference: in Agile, the DoD is external to the work item and checked by humans. In SEW, it is embedded inside the skill and checked automatically.
Software components specify preconditions, postconditions, and invariants as part of their interface. A skill's DoD is essentially a postcondition — the skill must satisfy it before completing. The difference: DbC postconditions are formal, machine-checkable assertions in code. Skill DoDs are natural-language criteria checked by an AI agent.
A quality gate is a checkpoint in a pipeline that must pass before the build proceeds. audit-workflow-deliverables is functionally a quality gate. The difference: CI/CD gates are configured externally in the pipeline. SEW's gate criteria are internal to the skill itself.
AOP separates cross-cutting concerns (logging, security) from business logic by weaving them in automatically. The audit call at the end of every skill is a cross-cutting concern — applied uniformly, regardless of what the skill does. The difference: AOP weaves concerns externally. In SEW, the skill author writes the audit call explicitly.
What appears to be new
No single feature here is novel in isolation. The combination is:
What a skill looks like
--- name: my-skill description: Does one job well. Invoke when the user asks to... allowed-tools: Read, Bash, Write, Edit --- # Skill: my-skill ## 1. Do the work ... skill procedures ... ## Definition of Done - [ ] Criterion one — specific, verifiable, pass/fail - [ ] Criterion two — covers a distinct output or state - [ ] Criterion three — atomic, one thing per criterion [Final step: Call audit-workflow-deliverables with this skill's path]