# Agent Skill Safety Review Intake

Use this redaction-friendly intake before requesting a $29 Custom Agent Audit Mini-Report for an AI-agent skill, portable workflow, or delegated subagent routine.

**Review page:** https://jarvislandingdeploy.vercel.app/agent-skill-safety-review.html  
**Checkout:** https://jarvislandingdeploy.vercel.app/api/checkout?product=custom-agent-audit-mini-report  
**Free contract template:** https://jarvislandingdeploy.vercel.app/portable-agent-skill-contract-template.md

## Safety first

Do **not** paste secrets, tokens, customer data, private repository contents, credentials, private messages, or proprietary internals unless your organization has explicitly authorized sharing them. Redacted descriptions are enough.

Replace sensitive values with placeholders like `[CRM]`, `[customer record]`, `[internal wiki]`, `[browser session]`, `[API key omitted]`, or `[private repo]`.

## 1. Skill identity

- Skill name:
- Version/date:
- Intended runner/runtime: OpenClaw / Claude Code / Hermes / Codex-style coding agent / browser-only agent / other:
- Primary job-to-be-done:
- Non-goals — what the skill must not do:
- Who will run it: human operator / autonomous agent / subagent / marketplace user:

## 2. Trust boundary

Check every boundary this skill crosses:

- [ ] Reads local files
- [ ] Writes or edits local files
- [ ] Uses browser sessions or authenticated websites
- [ ] Calls SaaS tools or APIs
- [ ] Touches credentials, tokens, secrets, or environment variables
- [ ] Reads/writes memory, notes, docs, or long-term state
- [ ] Sends messages, emails, comments, PRs, or public posts
- [ ] Triggers purchases, billing, deployment, deletion, or destructive changes
- [ ] Delegates work to other agents/subagents
- [ ] Intended for marketplace/public reuse

## 3. Inputs

- Required inputs:
- Optional inputs:
- Inputs the skill must reject:
- Source freshness requirement — what must be checked live before action?
- Private-data handling rules:

## 4. Tool authority

For each tool/action the skill may use, list the scope and stop condition.

| Tool/action | Allowed use | Forbidden use | Human approval needed when |
|---|---|---|---|
| | | | |
| | | | |
| | | | |

## 5. Output contract

- Required final artifact or response:
- Required evidence/citations/log links:
- Required confidence/uncertainty notes:
- Required machine-readable schema, if any:
- What counts as done?

## 6. Memory/state policy

- State the skill may read:
- State the skill may propose updating:
- State the skill may write automatically:
- State the skill must never store:
- Expiry/review rule for memories or saved decisions:
- Conflict rule when memory disagrees with current source-of-truth:

## 7. Receipts and audit trail

What should be logged after every run?

- Inputs used:
- Tools called:
- Files/pages/systems touched:
- External actions taken:
- Decisions made:
- Human approvals received:
- Exceptions/escalations:
- Rollback or recovery instructions:

## 8. Exception behavior

When should the skill stop and ask instead of acting?

- Missing input:
- Conflicting context:
- Stale source:
- Permission uncertainty:
- Irreversible action:
- Possible data leakage:
- User intent ambiguity:
- Tool failure:

## 9. Minimum eval cases

Provide or draft at least one case for each category.

| Eval type | Scenario | Expected safe behavior |
|---|---|---|
| Happy path | | |
| Missing input | | |
| Conflicting source | | |
| Permission boundary | | |
| Private data | | |
| Tool failure | | |
| Adversarial/prompt injection | | |
| Rollback/recovery | | |

## 10. What you want reviewed

Pick the most important review outcome:

- [ ] Is this safe to publish publicly?
- [ ] Is this safe to run internally?
- [ ] Is this safe to delegate to subagents?
- [ ] What permissions/tools should be narrowed?
- [ ] What receipts/evals are missing?
- [ ] What would make this marketplace-ready?
- [ ] Other:

## Paste-ready request

> I want a Custom Agent Audit Mini-Report for one agent skill. Here is my redacted intake. Please score readiness, identify the top safety/authority risks, list missing contract fields, suggest receipts/evals, and give a launch verdict.

---

**Next:** If this intake exposes real uncertainty around tool authority, memory writes, external actions, or public reuse, use the $29 review path: https://jarvislandingdeploy.vercel.app/api/checkout?product=custom-agent-audit-mini-report