1 min lesson
The anti-pattern taxonomy
Use "Every AI-in-the-SDLC failure mode rhymes" to describe the practical rule.
Step 1 of 3
Every AI-in-the-SDLC failure mode rhymes. Learn the taxonomy and you can name what's going wrong in a customer's org in one sentence, then point at the guardrail. For each one: the failure, who loses trust and the control that contains it.
Learn more
Advanced table
Anti-pattern
- Anti-pattern
- Vibe-merges
- Failure → who loses trust
- Merging on a green-checkmark feeling without reading the diff → reviewers, then prod when it breaks
- Guardrail
- Required author self-review plus code-owner approval; AI review stays an independent signal, never the merge trigger
- Anti-pattern
- Mega-diffs
- Failure → who loses trust
- One PR touches 80 files mixing refactor and behavior → reviewers can't hold the blast radiusHow much breaks if a change goes wrong; the scope of potential damage. Press Enter for the full definition., so they rubber-stamp
- Guardrail
- Enforce scope: split refactor from behavior, keep PRs small and coherent, flag oversized diffs in review
- Anti-pattern
- Prompt-and-pray
- Failure → who loses trust
- Fire a vague prompt, accept whatever comes back unverified → the author, whose name is on it
- Guardrail
- Specify intent, read the output, test it. The author owns the agent's work, full stop
- Anti-pattern
- Fabricated confidence
- Failure → who loses trust
- Model states a wrong answer fluently (hallucinated API, invented behavior) → whoever trusted the fluent tone
- Guardrail
- Verify against ground truth: docs, types, a running test. Tone is not evidence
- Anti-pattern
- Hidden generated code
- Failure → who loses trust
- Large AI-generated blocks merged with no signal they're generated → future maintainers and auditors
- Guardrail
- AI-code tracking and attribution; honest PR descriptions; review density scales with how much is generated
- Anti-pattern
- Context rot
- Failure → who loses trust
- Long agent session drifts off the original task; later edits contradict earlier ones → the author, silently
- Guardrail
- Short, scoped sessions; re-ground the agent; review the final diff against intent, not the conversation
- Anti-pattern
- Secrets / prompt-injection
- Failure → who loses trust
- Secrets pasted into context or hostile content in a repo or issue hijacks the agent → security, the whole org
- Guardrail
- Secret scanning, Privacy ModeCursor's setting that guarantees code data is not used for training by Cursor or its model providers, and that an admin can enforce org-wide; data-retention terms are a separate, contractual layer. Press Enter for the full definition. and ZDRZero Data Retention. A contractual guarantee that the model provider won't store your code or train on it. Press Enter for the full definition., terminal sandboxing, least-privilege MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition., treat external content as untrusted
- Anti-pattern
- Excessive agent permissions
- Failure → who loses trust
- Agent granted broad repo, tool and prod access 'to be convenient' → security and SRESite Reliability Engineering. The team and practice that keeps production reliable: monitoring, on-call and incident response. Press Enter for the full definition.; the blast radiusHow much breaks if a change goes wrong; the scope of potential damage. Press Enter for the full definition. is now huge
- Guardrail
- Least privilege: model/MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition./repo allowlists, RBACRole-Based Access Control. Granting permissions by role rather than configuring each person individually. Press Enter for the full definition., terminal sandboxing, isolated VMs, scoped tokens
- Anti-pattern
- Volume-as-success / mandated usage
- Failure → who loses trust
- Measuring lines-of-AI-code or mandating usage → leadership credibility, then engineers who game the metric
- Guardrail
- Measure outcomes (DORADORA metrics. Four widely-used delivery measures: deployment frequency, lead time for changes, change failure rate and time to restore service. Press Enter for the full definition., throughput, defect rate), not AI volume; drive adoption via mentorship, not mandate
| Anti-pattern | Failure → who loses trust | Guardrail |
|---|---|---|
| Vibe-merges | Merging on a green-checkmark feeling without reading the diff → reviewers, then prod when it breaks | Required author self-review plus code-owner approval; AI review stays an independent signal, never the merge trigger |
| Mega-diffs | One PR touches 80 files mixing refactor and behavior → reviewers can't hold the blast radiusHow much breaks if a change goes wrong; the scope of potential damage. Press Enter for the full definition., so they rubber-stamp | Enforce scope: split refactor from behavior, keep PRs small and coherent, flag oversized diffs in review |
| Prompt-and-pray | Fire a vague prompt, accept whatever comes back unverified → the author, whose name is on it | Specify intent, read the output, test it. The author owns the agent's work, full stop |
| Fabricated confidence | Model states a wrong answer fluently (hallucinated API, invented behavior) → whoever trusted the fluent tone | Verify against ground truth: docs, types, a running test. Tone is not evidence |
| Hidden generated code | Large AI-generated blocks merged with no signal they're generated → future maintainers and auditors | AI-code tracking and attribution; honest PR descriptions; review density scales with how much is generated |
| Context rot | Long agent session drifts off the original task; later edits contradict earlier ones → the author, silently | Short, scoped sessions; re-ground the agent; review the final diff against intent, not the conversation |
| Secrets / prompt-injection | Secrets pasted into context or hostile content in a repo or issue hijacks the agent → security, the whole org | Secret scanning, Privacy ModeCursor's setting that guarantees code data is not used for training by Cursor or its model providers, and that an admin can enforce org-wide; data-retention terms are a separate, contractual layer. Press Enter for the full definition. and ZDRZero Data Retention. A contractual guarantee that the model provider won't store your code or train on it. Press Enter for the full definition., terminal sandboxing, least-privilege MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition., treat external content as untrusted |
| Excessive agent permissions | Agent granted broad repo, tool and prod access 'to be convenient' → security and SRESite Reliability Engineering. The team and practice that keeps production reliable: monitoring, on-call and incident response. Press Enter for the full definition.; the blast radiusHow much breaks if a change goes wrong; the scope of potential damage. Press Enter for the full definition. is now huge | Least privilege: model/MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition./repo allowlists, RBACRole-Based Access Control. Granting permissions by role rather than configuring each person individually. Press Enter for the full definition., terminal sandboxing, isolated VMs, scoped tokens |
| Volume-as-success / mandated usage | Measuring lines-of-AI-code or mandating usage → leadership credibility, then engineers who game the metric | Measure outcomes (DORADORA metrics. Four widely-used delivery measures: deployment frequency, lead time for changes, change failure rate and time to restore service. Press Enter for the full definition., throughput, defect rate), not AI volume; drive adoption via mentorship, not mandate |
The two that masquerade as wins
Volume-as-success and mandated usage are dangerous precisely because they look like leadership 'driving adoption.' The metric goes up while trust goes down. Engineers game the number and the org learns that AI means theater you're forced to perform.
The Box case study is the counter-model: 85%+ daily active and 30–50% throughput gains came from mentorship, not from a mandate (+75% usage in 6 weeks via peer enablement). Pull, not push.
Learn more
Optional practice
Practice: The anti-pattern taxonomy
QA VP proudly reports that AI now writes 40% of the company's code and they've mandated daily Cursor usage. Name the anti-patterns and what you'd recommend instead.