debate-cbffcf — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited debate-cbffcf (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.
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.
Orchestrates a structured debate by spawning Claude sub-agents as named participants. The main Claude instance moderates, passes curated context between agents each round, and a Synthesizer produces an actionable verdict.
| Mode | Spawns | Rounds | Total agent calls |
|---|---|---|---|
--quick | Proponent + Critic | Opening only | 3 (+ Synthesizer) |
--full | Proponent + Critic + Devil's Advocate | Opening + one rebuttal round | 7 |
Extract from the user's message:
If a file path is given, read it now and attach its contents as shared context.
--quick or --full if explicitly present in the invocation.team size). These are passed to all agents as hard givens, not debatable.
If the topic is missing or genuinely ambiguous, ask one focused question only (e.g. "Are you debating whether to do X or how to do X?"). Do not ask multiple questions.
If --quick or --full was explicitly given, skip this step.
Otherwise, ask the user to choose before proceeding:
Quick — Proponent + Critic, opening statements only, Synthesizer verdict. (~3 agent calls) Full — Proponent + Critic + Devil's Advocate, opening statements + one rebuttal round, Synthesizer verdict. (~7 agent calls, costs more)
>
Which mode? (quick / full)
Wait for the user's answer before continuing.
If mode is --full (whether from the invocation or chosen in step 2), state:
Spawn Proponent and Critic in parallel. In --full mode, also spawn Devil's Advocate in the same parallel batch.
Each agent receives only this curated context (do NOT pass full conversation history):
Role brief: [role-specific paragraph below]
Topic: [exact topic text]
Domain context: [attached text or "none"]
Constraints (treat as fixed): [list or "none"]
Instructions:
- Write an opening statement of 300–400 words.
- Begin by steelmanning the strongest version of the opposing view in 2–3 sentences,
then explain why it ultimately does not hold.
- Use at least one concrete example or analogy.
- Avoid logical fallacies. Be direct.
- Do not hedge excessively or say "it depends" without specifying on what.Role briefs to inject per agent:
Proponent: "You are arguing IN FAVOR of the topic. Your goal is to make the strongest honest case for it. Do not mention downsides unless you are rebutting them."
Critic: "You are arguing AGAINST the topic. Your goal is to make the strongest honest case against it. Do not concede advantages unless you are rebutting them."
Devil's Advocate (full mode only): "You are challenging both the Proponent and the Critic. Your job is to expose weaknesses in both sides' reasoning, surface hidden assumptions, and raise the question neither side is asking. You are not neutral — you are adversarial toward weak arguments regardless of which side they come from."
Collect all opening statements. Present them to the user, labeled by role.
After displaying opening statements, ask:
"Anything to add (new constraint, clarification, context) before rebuttals? Press Enter / say 'continue' to proceed."
If the user adds something, incorporate it into the context passed in round 2. In quick mode, skip directly to step 7 (Synthesizer) after this step.
Spawn Proponent, Critic, and Devil's Advocate sequentially (each must read the others' opening statements).
Each agent receives the same curated context as round 1, plus the full round 1 transcript appended:
Round 1 transcript:
[all opening statements, labeled by role]
[If user added context in step 4, include it here as: "User added: ..."]
Instructions (rebuttal round):
- Write a rebuttal of 200–300 words.
- Address the single strongest argument made against your position.
- You may concede minor points if it strengthens your core argument.
- Do not repeat your opening statement.Collect all rebuttals. Append them to the transcript under a "Round 2" heading.
Spawn a Synthesizer agent with the complete debate transcript. Context passed:
You are synthesizing a structured debate. You were not a participant.
Your job is to produce a fair, actionable verdict — not to pick a winner by fiat.
Debate transcript:
[full transcript, all rounds]
Constraints (treat as fixed): [list or "none"]
Domain context: [same as passed to participants]
Produce a verdict in this exact format:
## Debate verdict
**Strongest argument (Proponent):** [one sentence naming the argument and why it lands]
**Strongest argument (Critic):** [one sentence]
**Strongest argument (Devil's Advocate):** [one sentence — omit if quick mode]
**Key unresolved question:** [the single most important question neither side resolved]
**Recommendation:** [one of: "Strong case for X", "Lean toward X", "Genuinely uncertain"]
**Reasoning:** [2–3 sentences explaining the recommendation, citing specific arguments]
**If genuinely uncertain:** state what information would break the tie.If the debate reveals strong consensus across all participants (rare), the Synthesizer may declare "Strong consensus toward X" instead of forcing a recommendation.
Show the full debate transcript (labeled by role and round) followed by the Synthesizer's verdict block. Do not editorialize beyond what the Synthesizer returned.
defined in each step. This prevents role contamination and reduces noise.
at 7. Do not add more participants without explicit user request.
should say so rather than inventing conflict.
entirely, note this in the transcript and do not silently let it pass.
subjective preferences (color choices, naming), redirect the user to /code-review or just answer directly.
was provided, offer once: "Want me to read a file or snippet before starting?" Do not ask again if the user declines.
givens and argue within them, not around them.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.