How do I install the Cavet Design Review skill?

Run npx skillmds@latest add chaoschild/cavet-design-review in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

Is Cavet Design Review safe to use?

This skill has not completed SkillMD's automated safety review yet. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with Cavet Design Review?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is Cavet Design Review free to use?

Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.

Who published Cavet Design Review?

ChaosChild (@chaoschild) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Cavet Design Review by chaoschild — Agent Skill

Name: Cavet Design Review
Author: ChaosChild

Cavet Design Review

Full security sweep of a design at the moment it is being finalised, agreed, written up, or handed to build — "let's go with this", "write the design doc", "ready to implement", an ADR or spec being completed, a plan being approved. Use even if security was discussed during the design conversation; the review exists because conversational flags are probabilistic and things get deprioritised mid-flow. Produces the threat model and decision records.

ChaosChild Updated Sep 17, 2026

File contents

cavet-design-review

The checkpoint. Conversational flagging (cavet-design) is probabilistic; this is the deterministic pass at the door between design and build. It exists to catch two things: what was raised and then forgotten, and what was never raised at all.

If cavet is not installed, invoke cavet-install and resume.

Procedure

Read the open items first. Run cavet items. This is not optional and it is not a formality: the review must not re-derive from the design document alone, because the class of thing that was raised, correctly deprioritised, and then forgotten is invisible in the document.
Tier 1 — raised and unresolved. For each open design item: restate the concern, the operator's stated reason for deferring (from the item), and whether the finalised design changes that calculus. Ranked above everything else.
Tier 2 — never raised. Sweep the finalised design against references/checklist.md. Report only gaps that are not already open items.
Present both tiers, tier 1 first, as tables (format below).
Record decisions. For each item the operator decides: run cavet resolve with the decision (accept risk / mitigate how / out of scope why). Items the operator defers again stay open with the new reason via the CLI. Do not close an item on your own judgement.
Write the artefacts. Update .cavet/design/threat-model.md and add one file per non-trivial decision under .cavet/design/decisions/. Templates in references/templates.md. These are what make build-phase findings interpretable later ("we accepted this because…"), so they carry reasoning, not just outcomes.

Output format

design review · <design name> · <n> open items · <m> new gaps

## Raised and unresolved
| item | concern | your reason at the time | still holds? |
|------|---------|--------------------------|--------------|

## Not previously raised
| # | area | gap | consequence |
|---|------|-----|-------------|

Then: for each row, one line of recommended disposition. No essay. Full rationale on request only (cavet-design/references/rationale.md has the common ones).

Calibration

A gap is something the design does not address, not something it addresses differently from how you would.
Do not pad tier 2 to look thorough. An empty tier 2 is a good result and should be said plainly: "no additional gaps".
If the design is too thin to review ("we'll figure out auth later"), say that as one tier-2 row rather than inventing detail.

---name: cavet-design-reviewdescription: Full security sweep of a design at the moment it is being finalised, agreed, written up, or handed to build — "let's go with this", "write the design doc", "ready to implement", an ADR or spec being completed, a plan being approved. Use even if security was discussed during the design conversation; the review exists because conversational flags are probabilistic and things get deprioritised mid-flow. Produces the threat model and decision records.---# cavet-design-reviewThe checkpoint. Conversational flagging (`cavet-design`) is probabilistic; this isthe deterministic pass at the door between design and build. It exists to catch twothings: what was raised and then forgotten, and what was never raised at all.If `cavet` is not installed, invoke `cavet-install` and resume.## Procedure1. **Read the open items first.** Run `cavet items`. This is not optional and it is not a formality: the review must not re-derive from the design document alone, because the class of thing that was raised, correctly deprioritised, and then forgotten is invisible in the document.2. **Tier 1 — raised and unresolved.** For each open design item: restate the concern, the operator's stated reason for deferring (from the item), and whether the finalised design changes that calculus. Ranked above everything else.3. **Tier 2 — never raised.** Sweep the finalised design against `references/checklist.md`. Report only gaps that are not already open items.4. **Present both tiers**, tier 1 first, as tables (format below).5. **Record decisions.** For each item the operator decides: run `cavet resolve` with the decision (accept risk / mitigate how / out of scope why). Items the operator defers again stay open with the new reason via the CLI. Do not close an item on your own judgement.6. **Write the artefacts.** Update `.cavet/design/threat-model.md` and add one file per non-trivial decision under `.cavet/design/decisions/`. Templates in `references/templates.md`. These are what make build-phase findings interpretable later ("we accepted this because…"), so they carry reasoning, not just outcomes.## Output format```design review · <design name> · <n> open items · <m> new gaps## Raised and unresolved| item | concern | your reason at the time | still holds? ||------|---------|--------------------------|--------------|## Not previously raised| # | area | gap | consequence ||---|------|-----|-------------|```Then: for each row, one line of recommended disposition. No essay. Full rationaleon request only (`cavet-design/references/rationale.md` has the common ones).## Calibration- A gap is something the design *does not address*, not something it addresses differently from how you would.- Do not pad tier 2 to look thorough. An empty tier 2 is a good result and should be said plainly: "no additional gaps".- If the design is too thin to review ("we'll figure out auth later"), say that as one tier-2 row rather than inventing detail.

Frequently asked questions