Skip to content

Latest commit

 

History

History
57 lines (33 loc) · 1.51 KB

File metadata and controls

57 lines (33 loc) · 1.51 KB

Common Anti-Patterns

The easiest way to ruin an AI-assisted workflow is to accumulate power without clear boundaries.

Multi-owner chaos

Two or more systems try to define the same task lifecycle.

Symptoms:

  • duplicate plans
  • conflicting instructions
  • unclear stop conditions
  • endless "which system owns this?" confusion

Prompt as architecture

Long prompts are not a substitute for a system design.

If the workflow only works because one giant prompt remembers everything, it will drift, fork, and rot.

Verification theater

The agent says the work is done, but nobody ran diagnostics, tests, or runtime checks.

This produces polished-looking output with weak delivery reliability.

Tool maximalism

Adding more agents, more skills, and more policies does not automatically improve results.

Each additional layer should have a reason to exist.

Internal-language lock-in

A workflow becomes unreadable to outsiders because it is expressed only in private shorthand.

If terms like owner, overlay, routing, or lifecycle are not defined, the playbook becomes a private notebook rather than a reusable resource.

Private draft forever

A repository says it is "almost ready" for open source, but never crosses the line.

Typical causes:

  • no license
  • no onboarding path
  • no case studies
  • no clear contribution model

Better pattern

Good vibecoding playbooks make their own boundaries obvious:

  • what is general methodology
  • what is tool-specific implementation
  • what evidence is required before making claims