Skip to content
Back to skills

Scope Guard

ASecurity

Judgment check before calling substantial work done or proposing a commit. Use before "done", before committing more than a trivial change, and whenever you notice yourself building something nobody asked for. The judgment layer above the automated gates, not a duplicate of them.

  • 2 stars
  • 0 votes
  • 0 copies
  • 1 view
  • Added September 26, 2026
ai-agentsrustgo

Security analysis

A100/100

Scanned September 26, 2026

npx -y skills add Tradebaas/Groundwork --skill scope-guard --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Scope Guard?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Scope Guard
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/tradebaas-scope-guard/badge)](https://www.skillsdirectory.com/skills/tradebaas-scope-guard)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: scope-guard
description: Judgment check before calling substantial work done or proposing a commit. Use before "done", before committing more than a trivial change, and whenever you notice yourself building something nobody asked for. The judgment layer above the automated gates, not a duplicate of them.
---

# scope-guard: before you say "done"

The gates (`checks/check.mjs`, typecheck, lint, tests) are the floor and must be green. This
checklist catches what they can't see. Answer honestly; the point is the answers, not the ritual.

1. **Trace.** Does every part of this diff trace to a spec, an SC-item in BRIEF.md, or an
   explicit request? Anything that doesn't → out of the diff, into INTAKE.md.
2. **Ladder.** Did something get built that already existed: in the repo, the platform, or a
   dependency? Rebuilt-in-disguise counts.
3. **Minimum.** Is there code here for cases nobody asked for: options, abstractions,
   "flexibility"? Deletion is still cheap now; a month from now it's archaeology.
4. **While-I-was-here.** Any refactors, renames, or cleanups outside the task? Revert them or
   split them into their own proposed change; never smuggle them along.
5. **Safety floor intact?** Validation at trust boundaries, authorization, data-loss error
   handling, no silently swallowed failures, secrets, accessibility: none simplified away
   (AGENTS.md hard rule).
6. **Deferrals marked.** Every deliberate simplification has a `defer:` marker with a ceiling
   and an upgrade trigger, and a DEBT.md row if it will outlive this change.
7. **Done means verified.** Has `verify` actually been run on this change, behavior exercised,
   not just green tools? If not, that's the next step, not this one.
8. **Good is good.** Now that it passes: stop. List improvement ideas in INTAKE.md if they're
   worth keeping; build none of them unasked.

Then update what the change touched: STATE.md (always), the spec status, DEBT.md, stale docs
(retired wording → denylist). Report the check's outcome in one or two lines: findings and
what you did about them, or "clean". ⚓

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.