Impeccable explained: design guidance inside an AI coding workflow
Impeccable architecture: product context, command and detector
Follow a design request through the skill, runtime and review surfaces.
What you will learn
- Product truth is durable input
- Commands are an interface to work
- Hooks add timing, not authority
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
- Context and visual direction have different lifetimes.
- Rules and agent critique provide different evidence.
- Hooks do not grant design authority.
Product truth is durable input
The README describes `init` writing PRODUCT.md for audience, constraints and evidence. Design direction for a particular surface belongs separately in DESIGN.md. A command can use both, but a new layout should not overwrite the product’s factual context.
This separation is useful when one product has a landing page and a dense dashboard. Their composition can differ while the intended users and operating constraints remain the same.
Commands are an interface to work
The skill presents commands such as `audit`, `critique`, `polish` and `distill`. Some ask an agent to reason and edit, while the detector rules produce reproducible findings. A command name is not a guarantee that every suggestion is mechanically checkable.
The README distinguishes a composition-first path from a code-first path and documents live iteration. Choose the path that fits the current state of the screen, then preserve the evidence for the visual decision.
Hooks add timing, not authority
Provider-native hooks can run around file edits and surface detector output. They change when feedback appears. The user and host still decide whether to accept edits and grant project trust.
An architecture diagram should mark the hosted agent, local launcher, engine binary and project files as separate boundaries. This review does not claim to trace every platform adapter in the monorepo.
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
Record PRODUCT.md and a surface-specific direction separately.
- 2
Trace one command from the skill to its output.
- 3
Identify where a hook runs and who approves it.
Copy-ready example
PRODUCT.md -> agent skill command -> engine or critique
DESIGN.md -> surface direction ----^
rendered screen + detector output -> reviewer decisionFrequently asked questions
Where are product facts stored?
The documented init flow writes PRODUCT.md.
Does every command run the same checks?
Read the command description and engine contract; command names alone do not establish identical behavior.
Sources
- Impeccable / README.mdSource checked 2026-09-29
- Impeccable / .agents/skills/impeccable/SKILL.mdSource checked 2026-09-29
- Impeccable / docs/ENGINE.mdSource checked 2026-09-29