Devibify is a Codex skill for implementing, refactoring, and auditing web interfaces that need product-design judgment. It keeps work within scope, grounds decisions in product evidence, preserves coherent design systems, covers reachable states, and reports what verification supports.
Use it for UI work such as:
- auditing a product surface without changing files;
- fixing a focused interaction or accessibility defect;
- implementing a page, form, table, dashboard, or flow;
- refactoring shared UI without changing accepted behavior;
- translating an approved visual reference into code;
- replacing generic AI-generated structure with product-specific hierarchy.
Devibify does not target backend-only work, mechanical frontend edits, copy proofreading, or exact token substitutions that require no UX decision.
Place this repository folder in the Codex skills directory:
$CODEX_HOME/skills/devibify
With the default Codex home directory, the path is:
~/.codex/skills/devibify
Keep SKILL.md at the top of the installed devibify folder.
The repository file agents/openai.yaml stores skill interface metadata inside the skill folder.
Install the whole repository folder. You do not need to remove or rearrange any files. During normal use, Codex starts with SKILL.md and opens supporting references when relevant. It does not automatically load the tests or eval data.
Call the skill by name when you want to force activation:
Use $devibify to audit this checkout flow. Return findings only, cite the evidence for each finding, and do not change files.
The skill also allows implicit invocation for matching frontend work. Its description excludes backend work and mechanical edits.
Devibify selects one primary mode before broad inspection or modification:
| Mode | Purpose |
|---|---|
AUDIT |
Inspect and report without modifying product files |
NARROW_FIX |
Correct one defined surface or defect |
FEATURE_IMPLEMENTATION |
Add or change a page, flow, component family, or behavior |
NET_NEW_DESIGN |
Design a new interface without an established reference |
REFACTOR |
Improve structure while preserving accepted behavior and visual intent |
REFERENCE_FIDELITY |
Implement against an approved screenshot, design, or specification |
Each mode uses the same scope, evidence, and verification rules with effort proportional to the task.
SKILL.mdcontains activation boundaries, guardrails, modes, workflow gates, and reference routing.agents/openai.yamlcontains Codex interface metadata and implicit-invocation policy.references/contains task-specific guidance loaded through the routing table.scripts/scan-ui-smells.pyprovides an optional, read-only source scanner. Its findings require human or agent review.
scripts/validate-devibify.pychecks package structure, metadata, links, encoding, eval schemas, and release counts.tests/covers the validator and every scanner rule with standard-library unit tests.evals/records trigger expectations, behavior cases, critical gates, and the scoring rubric.
The eval corpus specifies expected behavior. This repository has no runtime Codex evaluator, so it does not claim that a model has passed those cases. The validator checks eval structure and coverage. These maintenance files belong in the source repository even though Codex does not need them during a skill run.
Regular users do not need to run the assurance harness. It is included for maintainers, contributors, and anyone who wants to verify the repository.
Run the repository checks from its root:
python scripts/validate-devibify.py .
python scripts/validate-devibify.py . --json
python -m unittest discover -s tests -p 'test_*.py'
python -m py_compile scripts/validate-devibify.py scripts/scan-ui-smells.py
python -m json.tool evals/trigger-cases.json >/dev/null
python -m json.tool evals/behavior-cases.json >/dev/nullScan supported UI source files with:
python scripts/scan-ui-smells.py <path>
python scripts/scan-ui-smells.py <path> --json
python scripts/scan-ui-smells.py <path> --strictThe scanner reports findings without editing files. Default mode returns status 0 when it finds issues. Strict mode returns status 1 when it finds a high-confidence issue.
See CHANGELOG.md for the v1 and v2 feature summaries.
Cyxzdev's Uncodixfy inspired Devibify.
See LICENSE.md.