Skip to content
Back to skills

Independent Review

ASecurity

Use when security, auth, billing, concurrency, data integrity, or another high-stakes artifact needs adversarial review by a provider other than its author.

  • 7 stars
  • 0 votes
  • 0 copies
  • 5 views
  • Added August 31, 2026
ai-agentsrustgobashgitsecurity

Works with

  • cli

Security analysis

A100/100

Pro scans all 2 files and shows the line behind each finding

Scanned October 1, 2026

npx -y skills add lawzava/megapowers --skill independent-review --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Independent Review?

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

Security grade badge for Independent Review
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/lawzava-independent-review/badge)](https://www.skillsdirectory.com/skills/lawzava-independent-review)

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: independent-review
description: Use when security, auth, billing, concurrency, data integrity, or another high-stakes artifact needs adversarial review by a provider other than its author.
when_to_use: "Trigger phrases: get a second opinion from another model, external review, have another vendor review this, adversarial review of auth, billing, or migration code before merge."
metadata:
  short-description: Cross-provider adversarial review of a high-stakes artifact
---

# Independent Review

Use model review for residual uncertainty after executable checks. Same-provider
review gives only context separation, not independence. The lead owns
remediation and reruns the oracle.

## Trusted review path

Take the reviewer family and exact command from the capability registry or
user instruction; never guess. Inspect a file or immutable commit range before
any external call. Set `review_tool` to the installed
`scripts/megapowers-review.go` beside this `SKILL.md`, not a project copy.

```bash
reviewer_cmd='<reviewer CLI and arguments>'   # {prompt_file} and {scratch_dir} expand
intent='<intended behavior and acceptance boundary>'
go run "$review_tool" inspect --file path/to/file --intent "$intent" \
  --provider <reviewer-family> --provider-command "$reviewer_cmd"
go run "$review_tool" inspect --base <base-revision> --head <head-revision> \
  --intent "$intent" --provider <reviewer-family> --provider-command "$reviewer_cmd"
```

The inspection prints the intent, binary hash, command, chunk package hashes,
and one `approval_token`. Approve only that exact package, intent, binary, and
command:

```bash
go run "$review_tool" review --file path/to/file --intent "$intent" \
  --provider <reviewer-family> --provider-command "$reviewer_cmd" \
  --author <author-family> --approve-external "$approval_token"
```

Author and provider labels must differ and name real vendor families. The
prompt arrives on stdin unless the command names `{prompt_file}`. Credentials
pass only through `--provider-env NAME`. A preflight probe fails fast on login,
usage-limit, or stall before any artifact bytes leave the machine. Any file,
intent, binary, or command change requires a new inspection and token.

## Declare the payload once

External dispatch sends artifact bytes to another vendor. Before sending, write
one declaration: the file list or immutable commit range, total bytes, and
destination provider. Ask approval with that same declaration and dispatch
that identical scope, so any harness approval reviewer sees one payload. A review mandate alone does not authorize disclosure. If the
artifact changes after approval, declare and approve again. If the harness's
own reviewer still denies the send, report its reason and stop; do not retry
with a different payload. Surface a permission prompt or provider stall
instead of waiting silently.

Receipts are advisory, not an approval gate; they stay under private Git
metadata unless `--out` names an existing absolute directory outside the
repository. Use `--retain-transcript` only after a sensitive-data decision.
Treat the review verdict as a claim. Record credible findings and explain
dismissals against the artifact intent. Fix as the single writer and rerun
acceptance tests after every material change. Bound correction rounds:
re-review only fixes, affected boundaries, and new evidence. Join every
requested review before completion; approval cannot settle a queued review.

Files in this skill

  • SKILL.md2.7 KB
  • scripts/megapowers-review.go37.4 KB

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.