system-prompt-curator — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited system-prompt-curator (Agent Skill) and scored it 92/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 2 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 2 flagged
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
extend op-cell: turn a role, task class, and tool surface into a durable autonomous-agent system prompt; or correct an existing prompt until its workflow and completion gates are load-bearing. Optimize for agents that must act in a repository, recover from tool failures, verify their work, and deliver evidence rather than analysis.
Keep the prompt self-contained. Do not depend on tribal context, hidden tools, model names, or sibling documents. The generated prompt must tell the agent who it is, where it works, what tools exist, what phases it must follow, what evidence gates completion, and how to recover when reality disagrees.
Core payloads:
references/principles.md — ten research-backed prompt principles and the invariant each protects.references/anti-patterns.md — failure mode → why it fails → fix table used by --improve.references/templates.md — mandatory template, GitHub-issue template, research template, and harness-level reinforcements.references/templates.md mandatory skeleton unless the role exactly fits the GitHub-issue or research template. Preserve section order: IDENTITY → ENVIRONMENT → TOOLS → WORKFLOW → COMPLETION → TIPS → WHAT NOT TO DO → EXAMPLES.references/templates.md: completion validation, separate help path, observation formatting, empty-output handling, truncation rules, environment suppression, history compression, malformed-output retry.--improvereferences/anti-patterns.md; report every confirmed issue. Certainty:complete immediately, no phases, no completion preconditions, no tools, no verification).references/principles.md as PASS / FAIL / PARTIAL. Do not skip principles because the prompt is short.| Severity | Meaning | Required Action |
|---|---|---|
| HIGH | Prompt can complete without work, call finish without evidence, use nonexistent tools, or skip verification | Fix before returning |
| MEDIUM | Prompt likely drifts: vague identity, weak phases, no recovery guidance, lazy tool discovery, missing example | Fix unless user explicitly requested minimal prompt |
| LOW | Style or harness hardening issue: aggressive phrasing, token bloat, missing observation formatting, no output truncation rule | Fix when cheap; otherwise recommend |
Before returning any created or improved prompt, verify:
| Gate | Pass Criteria | Blocking |
|---|---|---|
| Role anchor | Identity matches task type and implies the correct behavior | Yes |
| Tool contract | Prompt lists actual available tools and when to use them | Yes |
| Ordered workflow | Prompt has explicit phases and no analysis-phase completion | Yes |
| Artifact mandate | Prompt states the concrete deliverable and forbids read-only completion for action tasks | Yes |
| Verification | Prompt names the verification command/method or evidence substitute | Yes |
| Completion schema | Finish action requires non-empty evidence fields appropriate to the task | Yes |
| Failure recovery | Tips or example show a failed command/tool call and changed approach | Yes for autonomous agents |
| Convention discovery | Coding agents must inspect existing project patterns before editing | Yes |
| Tone | No adversarial threat block or exaggerated all-caps rule wall | Yes |
| Self-eval | All ten principles checked PASS or justified PARTIAL | Yes |
Create mode output:
PROMPT:
<complete system prompt>
DESIGN NOTES:
- Identity: ...
- Completion gate: ...
- Verification: ...
- Harness assumptions: ...
- Approx tokens: ...Improve mode output:
FINDINGS:
| Severity | Principle | Evidence | Risk | Fix |
|---|---|---|---|---|
...
IMPROVED PROMPT:
<complete rewritten system prompt>
DELTA:
- ...
SELF-EVAL:
| Principle | Status | Note |
|---|---|---|
...complete, finish, or return instruction without preconditions. Gate it.Use available tools as needed with no tool list. Enumerate tools or state they are unavailable.CRITICAL, NEVER FAIL, YOU WILL BE PENALIZED blocks. Replace with normal checklists.You are a focused AI agent. Replace with role tied to task and deliverable.~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.