Impeccable explained: design guidance inside an AI coding workflow
Impeccable explained: design guidance inside an AI coding workflow
See what its skill, commands and detector do, then decide where human review remains necessary.
What you will learn
- Understand the three surfaces
- Know what the detector sees
- Set the review boundary
Before you start
- One screen and its user task
- Authority to inspect an agent skill and project hook
Produce an evidence-backed design change that another reviewer can accept or reject.
Key takeaways
- The skill and detector solve different parts of the workflow.
- Product facts belong apart from surface styling.
- A clean scan is evidence, not design approval.
Understand the three surfaces
Impeccable packages design guidance as an agent skill, exposes named commands and runs deterministic checks on frontend work. Its README lists 24 commands and 61 detector rules at the inspected revision. Those counts describe the upstream release, not a quality score.
The initial `init` flow records durable product facts in PRODUCT.md. A later surface can have its own visual direction in DESIGN.md. Keeping audience and constraints apart from transient styling prevents an agent from treating a screenshot as product strategy.
Know what the detector sees
The CLI and browser extension can run rule-based checks without an LLM key. An agent critique may add judgment, but a clean deterministic scan cannot prove that a page works for its audience or meets accessibility needs.
Begin with one existing screen. Write down its task, constraints and current design system, then compare a detector finding with the rendered page. Accept or reject a suggested change for a stated reason.
Set the review boundary
The skill can shape an agent’s edits, and hooks can surface findings during those edits. Neither should silently approve a production release. The review should still cover copy, interaction, responsive behavior and actual user evidence.
This series inspected a fixed repository revision. It did not run the binary, browser extension or a user study; commands and integration behavior are attributed to the upstream files.
Decision guide
| Criterion | Option A | Option B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
Implementation steps
- 1
Pick one screen and record the user task in PRODUCT.md.
- 2
Inspect its current visual system before using a command.
- 3
Compare one detector finding with the rendered page.
Copy-ready example
product truth: audience + task + constraints
surface direction: layout + type + interaction
review: detector finding + rendered evidence + human decisionFrequently asked questions
Is Impeccable a design system?
It supplies guidance and checks; the product still needs its own design decisions.
Does it require a model API key?
The README says deterministic detection can run without one; agent-driven critique depends on the host you use.
Sources
- Impeccable / README.mdSource checked 2026-09-29
- Impeccable / .agents/skills/impeccable/SKILL.mdSource checked 2026-09-29