🚧 Anti-patterns
Habits that undermine autonomous delivery—and what to do instead.
Operational & Engineering
Developers still writing code syntax
Using AI as autocomplete while still writing logic and tests by hand. Let agents write the code; focus on specifications, controls, and Persistent Context.
Manual line-by-line syntax reviews
Reading generated diffs by hand as a merge gate. Use automated checks and independent Prove assessment. People resolve failures, not routine sign-off.
Skipping client meetings & building from stale tickets
Developers isolated from client dialogue, relying on third-hand ticket summaries. In Meeting-Driven Development, client dialogue is the primary signal source, captured and structured by AI (0 — Meet).
The AI center of excellence that owns delivery
A central team that writes the AI-assisted code for everyone. It becomes a queue, squads lose ownership, and the practice never spreads. Enablement ships capabilities, not features.
Enablement as an approval board
Platform or security turning into a gate squads must ask permission from. The correct answer to recurring approval requests is an automated deterministic guardrail, not a meeting.
Technology-layer squads
"Frontend squad", "API squad", "data squad". Every user outcome then requires three squads and two hand-offs. Boundaries follow outcomes.
Process & Quality
Skipping Plan
Sending a vague request straight to a coding agent. Define testable acceptance criteria in Plan first.
Weak deterministic controls ("Trust the model")
Relying on LLM self-critique or superficial unit tests that do not test invariants. Generated code must face automated checks (compilers, strict type systems, AST linters, tests) and dense deterministic controls (contract tests, mutation testing, security scans).
Hand-fixing generated bugs
Fixing generated code by hand. Add a failing test or correct the context, then let the agent fix the implementation.
Uniform review effort
Applying the same checks to a copy edit and a permissions change. Match validation evidence to risk.
Batching releases
Holding changes for a weekly release "for safety". Batching multiplies blast radius and destroys the correlation between a change and its effect.
Context & AI usage
Prompting instead of context
Increasingly elaborate prompts re-explaining the same conventions in chat. That is a missing Persistent Context, not a prompt engineering skill gap.
The everything instruction
One 900-line instruction file covering all of engineering. Models weight it poorly, humans stop reading it, and it drifts. Small, scoped, evidence-driven artifacts.
Stale context left in place
An instruction describing a framework the company migrated away from. Worse than no instruction: it produces confident, consistent, wrong output at scale.
Vibes-based rollout
Trying a new model on one prompt, liking it, enabling it for everyone. Without rigorous evaluation, a single change can make the whole organization worse overnight.
Agents with production credentials
An agent able to act directly on production because it was convenient. Agents act through pull requests behind deterministic controls and independent automated validation before merge. Autonomous decisions do not grant unrestricted production access.
Tool sprawl and context fragmentation
Scattering requirements, code, and evidence across disconnected tools. Keep them linked on one Git platform.
Measurement
Counting output volume
Lines of code, pull requests, suggestions accepted. Generation makes all of these trivially inflatable and none of them meaningful.
Never deleting anything
A codebase where invalidated features accumulate because deleting feels like waste. Every dead feature taxes every future generation pass and every future reviewer.