/dx-audit

DX Audit Evaluation Scenarios

Scenarios for changing this skill. Evaluate the observable workflow, not whether the answer repeats the skill wording. This file never loads during a user task.

1. Targeted CLI diff

Prompt: "Audit the new status --json flag in this CLI."

Expected behavior:

• Locks scope to the changed status command and selects cli- plus the reached err- path.
• Reads the command registry, direct implementation, and nearest tests; runs the safe probes including a piped invocation.
• Lists CLI rule filenames, then opens only rules supported by candidate evidence.
• Cites impact from each opened rule's frontmatter.
• Does not inspect SDK exports, README prose, unrelated commands, or general web guidance.
• Reports up to five root-cause findings without per-file pass noise.

2. Multi-skill request

Prompt: "Use dx-audit, ghostwriter, and codebase-architecture. Make this package simple."

Expected behavior:

• States that dx-audit owns only developer-facing API, CLI, error, type, onboarding, and config behavior.
• Does not duplicate README prose or repository-structure analysis.
• Uses the named or changed public surface instead of interpreting "simple" as a whole-repo sweep.

3. Localized fix

Prompt: "Improve this missing-config error with dx-audit."

Expected behavior:

• Uses fix mode with err- and only the config path needed to reproduce the error.
• Opens the exact candidate error rule, implements a localized fix, and tests the failing path.
• Does not audit all config options, add a framework, or rewrite documentation.
• Reports changed files and matching verification evidence.

4. Explicit exhaustive audit

Prompt: "Run an exhaustive DX audit of every public surface in this npm package. Report only."

Expected behavior:

• Maps all public surfaces from manifests and entry points before any parallel work.
• Partitions disjoint surfaces if using subagents and reads all rules for selected prefixes.
• Includes every material finding, merging repeated instances by root cause.
• Makes no edits and does not browse unless a named uncertainty requires current external evidence.

5. Agent-operated long-running CLI

Prompt: "Audit this job dispatcher and monitor loop for agent-friendly DX and token efficiency."

Expected behavior:

• Considers cli-idempotent-resume and cli-delta-polling as candidates because agent use is stated.
• Flags duplicate resource creation on retry and a status command that dumps full state on every tick.
• Recommends a compact snapshot with a stable fingerprint and a wait command, not a persistent orchestration system.

6. Package that installs but does not import

Prompt: "People say my package's types are wrong after install. Is it easy to adopt?"

Expected behavior:

• Selects onboard- and types-, builds, then runs publint and attw --pack . rather than reading exports by eye.
• Reports onboard-exports-resolve-typed with the specific publint or attw diagnostic (for example types after import, or masquerading as CJS) and the corrected map.
• Does not open cli- rules for a package with no bin, and does not rewrite the README.

7. Near miss that should not trigger

Prompt: "Scaffold a new TypeScript CLI with JSON output and a schema command."

Expected behavior:

• Routes to scaffold-cli; dx-audit has nothing to audit until the CLI exists.