skeptical-triage
关于
This Claude Skill implements a 3-round self-challenge with an arbiter to rigorously filter false positives from findings or verdicts. It's designed for high-stakes scenarios where a false-positive gate block would be costly, such as blocking deployment or writing to trusted reports. The skill uses tools like Read, Grep, and Bash to examine code and documentation paths during its verification process.
快速安装
Claude Code
推荐npx skills add avelikiy/great_cto -a claude-code/plugin add https://github.com/avelikiy/great_ctogit clone https://github.com/avelikiy/great_cto.git ~/.claude/skills/skeptical-triage在 Claude Code 中复制并粘贴此命令以安装该技能
技能文档
Skeptical Triage
Filter false positives from multi-angle review, security audit, QA regression flags, or any high-stakes judgment before it turns into a blocker.
Three rounds of skeptical self-review + an impartial arbiter, with a confidence score from the vote.
When to invoke
| Caller | Finding type | Apply triage? |
|---|---|---|
/review | Angle 2/4/7/9 P0/P1 (security, SQL, privacy, concurrency) | Yes |
/review --deep | Any angle P0/P1 | Yes |
security-officer | CSO audit P0/P1 | Yes |
security-officer | Secret in source/git, confirmed CVE | No — hard finding |
qa-engineer | Flaky-test verdict (is this a regression or flake?) | Yes |
architect | ADR trade-off dispute (option A vs. B when both look reasonable) | Yes |
| Any | P2/advisory | No |
The 4-step pattern
Run these sequentially. Each round sees prior reasoning. Arbiter sees all rounds.
Round 1 — Reachability / Premise
Question: is the premise true?
- For security/reliability: can an external attacker reach this code path with untrusted input? Trace input flow backward from the bug site to its origin. If only trusted internal callers → lean INVALID.
- For regressions: does the failing behavior reproduce from a clean state on the target branch?
- For ADR trade-offs: is the constraint that forces the choice actually binding? (e.g. "we need <10ms p99" — is that real or aspirational?)
Output: {round: 1, verdict: VALID|INVALID|UNCERTAIN, reasoning: "...", crux: "single key fact"}
Round 2 — Verify cited defenses / counter-evidence
Question: are claimed defenses real and sufficient?
- Every cited defense → use
Grepto find its actual implementation line. - Resolve constant names to numeric values.
MAX_BUF_SIZEis not a verified bound —#define MAX_BUF_SIZE 64is. - For regressions: is the cited "test covers this" actually asserting the right invariant?
- For ADR: is the cited benchmark/precedent real (grep for it, read it), or rumored?
If you cannot point to the line that enforces the defense, it does not exist.
Output: same JSON shape, with grep_used: true/false.
Round 3 — Missed angles
Question: what did Rounds 1-2 not consider?
- Error paths, integer overflow, race windows, different callers, platform differences
- Do NOT rehash prior rounds — add new evidence or concede
- For QA: retry logic masking the failure? Test pollution from another test?
- For ADR: option C that neither reviewer raised?
Output: same JSON shape.
Arbiter
Input: all 3 rounds + original finding/question + source code.
Question: final call — which side has the stronger evidence?
- Deliver single
verdict: VALID|INVALID(no UNCERTAIN — make the call). - Deliver one-sentence
crux— the key fact the verdict turns on. - If 3 prior rounds all said the same thing, only override with overwhelming new evidence and explain why.
Output:
{
"verdict": "VALID",
"crux": "memcpy at auth.c:142 copies network-controlled len bytes into 64-byte stack buffer with no bound check",
"reasoning": "Rounds 1 and 3 verified attacker reach; Round 2 found no size check in 50 LOC radius; arbiter confirms no caller clamps len."
}
Hard rules
Burn these into every round's prompt:
- Absence of defense → VALID, not UNCERTAIN. If you searched for a defense and did not find one, that is the answer. "Other code probably handles this" is not a valid defense.
- A constant name is not a verified bound — only its resolved value is. Grep for the
#define/constdeclaration. - Name the line or it does not exist. Vague references to "assumptions in this codebase" do not count.
- Do not contradict your own conclusion in the same response. If you verified a defense is insufficient, that is the verdict. Stop searching for reasons to flip.
- Code quality issue ≠ security vulnerability. Data race on diagnostic state, NULL check on internal-only API, UB only in debug builds → INVALID.
- Trust your own reasoning. If you see the crux on first read, don't manufacture a counter-argument.
Confidence scoring
confidence = valid_rounds_before_arbiter / 3
100%(VVV) — 3/3 rounds VALID. Arbiter rubber-stamps unless it finds something brand-new.67%(VVI or VIV or IVV) — majority VALID. Arbiter breaks tie with new evidence.33%(IIV or IVI or VII) — majority INVALID. Arbiter usually confirms INVALID.0%(III) — 3/3 INVALID. Arbiter rarely overrides.
Arbiter overrides the final verdict; confidence reflects the round vote for transparency. Record both in the output so humans can see where the arbiter diverged.
Applying triage results to severity
Once the arbiter returns:
| Arbiter verdict | Confidence | Severity action |
|---|---|---|
VALID | ≥ 50% | Keep original severity |
VALID | < 50% | Demote: P0→P1, P1→P2 |
INVALID | any | Remove from gate tally, record as [FILTERED] in report for audit |
UNCERTAIN (only if arbiter could not decide) | n/a | Keep original severity, flag for manual CTO review |
Output schema
Every caller logs triage results to .great_cto/triage-log.jsonl (append-only, one JSON per line):
{
"timestamp": "2026-04-19T12:34:56Z",
"caller": "review|security-officer|qa-engineer|architect",
"finding_id": "SEC-042",
"file": "src/auth.c:142",
"original_severity": "P0",
"rounds": [
{"round": 1, "verdict": "VALID", "crux": "..."},
{"round": 2, "verdict": "VALID", "crux": "...", "grep_used": true},
{"round": 3, "verdict": "INVALID", "crux": "..."}
],
"arbiter": {"verdict": "VALID", "crux": "..."},
"confidence": 0.67,
"final_severity": "P0"
}
This log is how we measure whether triage earns its keep. Review it weekly:
# False-positive rate: how many findings the arbiter flipped to INVALID
jq 'select(.arbiter.verdict=="INVALID")' .great_cto/triage-log.jsonl | wc -l
# Average rounds-to-consensus (did we need all 3 or did R1+R2 agree?)
jq '[.rounds[].verdict] | unique | length' .great_cto/triage-log.jsonl
If FP rate < 10% after 50 triages — triage is filtering noise that wasn't there. Lower threshold or skip triage for that angle. If FP rate > 40% — original review prompt is too trigger-happy; tighten the angle rules.
Token budget
Per triaged finding: ~4 LLM turns (3 rounds + arbiter). At typical review sizes (~5-10 triaged findings per PR), total budget: 20-40 extra turns per /review. Batch when possible — one arbiter can handle multiple findings in a single call if their cruxes are independent.
For cost-sensitive runs (approval-level: auto on a huge PR), consider: triage only P0, leave P1 untriaged. Re-tune based on .great_cto/triage-log.jsonl data.
Anti-patterns
- Don't triage P2/advisory findings. The whole point is gate decisions. P2 is advisory — let the author see it and move on.
- Don't let rounds rehash each other. Round 3 prompt must say "add NEW evidence or concede." If 3 rounds produce identical reasoning, you wasted 2 turns.
- Don't skip the arbiter on UNCERTAIN. If all 3 rounds say UNCERTAIN, the arbiter's job is to decide — not to join the fog.
- Don't hide arbiter overrides. When the arbiter flips the majority vote, record both
confidence(the vote) andfinal_verdict(the arbiter). Humans deserve to see the disagreement.
GitHub 仓库
相关推荐技能
llamaguard
其他LlamaGuard是Meta推出的7-8B参数内容审核模型,专门用于过滤LLM的输入和输出内容。它能检测六大安全风险类别(暴力/仇恨、性内容、武器、违禁品、自残、犯罪计划),准确率达94-95%。开发者可通过HuggingFace、vLLM或Sagemaker快速部署,并能与NeMo Guardrails集成实现自动化安全防护。
cost-optimization
其他这个Claude Skill帮助开发者优化云成本,通过资源调整、标记策略和预留实例来降低AWS、Azure和GCP的开支。它适用于减少云支出、分析基础设施成本或实施成本治理策略的场景。关键功能包括提供成本可视化、资源规模调整指导和定价模型优化建议。
quantizing-models-bitsandbytes
其他这个Skill使用bitsandbytes库量化大语言模型,能在GPU内存有限时通过8位或4位量化减少50-75%内存占用,同时保持精度损失最小。它支持INT8、NF4、FP4等多种量化格式,可与HuggingFace Transformers无缝集成,适用于需要部署更大模型或加速推理的场景。还提供QLoRA训练和8位优化器支持,让开发者能轻松实现高效模型压缩。
dispatching-parallel-agents
其他该Skill用于并行处理3个以上无依赖关系的独立故障,可为每个问题域分派专属Claude代理同时执行调查修复。它通过并发处理多个独立问题显著提升故障排查效率,特别适用于测试文件、子系统等无共享状态的场景。
