Verification Skills for OpenClaw Agents

事实核查

@panting09266-ai

用 Cue 穿透待核查资讯的表层表述——通过独立信源交叉验证与底层数据复核，精准识别时间错位、数据偏差及误导性信息，产出包含原文对照与纠偏证据的结构化事实核查报告。

Inspect, implement, test, and verify focused repository changes with Flixa, then generate shareable proof of work without bypassing safety policies.

verification-before-

@thcjp

自动化执行verification命令，验证工作完成状态，提升开发效率，支持批量处理和结构化结果输出。

popcorn-temporal-anchor

@violetclaire

Verify a settled payload-bound time checkpoint before a time-sensitive cross-node action. Use when a task has an execution window or when exact bytes must be checked against a signed witness interval.

sente

@shim2k

Give this agent its own email identity with Sente — a real, durable address it owns (name@sente.run) plus the accounts layer on top. Send and receive mail as the agent; block on verification emails and get just the OTP code or magic link (sente wait --otp); register new accounts at third-party apps with a real browser (autonomous or confirm-before-submit); or connect accounts the user already owns (credentials vaulted write-only, authenticator/TOTP re-login, revocable). Use when an agent needs its own email address or inbox, is stuck at "check your email to continue," needs a verification code or magic link extracted, or must sign up / sign in to a web app and keep that account alive.

evidence-first

@raguets

Part of the Overpowered skill suite. Require evidence proportional to material claims and clearly separate evidence, inference, and uncertainty. Use for research, analysis, recommendations, reports, audits, or decisions where unsupported confidence would be harmful; do not use for purely creative or stylistic tasks.

Verify an LLM API endpoint — model authenticity, billing inflation, relay provenance, performance and silent downgrades. Use when the user asks whether the model they are paying for is genuine, whether a relay or proxy is trustworthy, whether they are being overcharged, or whether a model has been quietly downgraded. 检测 LLM API 端点的真伪、计费掺水、中转来源、性能与降智；当用户问"我用的模型是不是真的"、"这个中转站靠谱吗"、"是不是被降智了"、"计费对不对"时使用。

Skillv22

@orionshaowswmw

Turn an AI agent into an expert at inference — difficulty triage, deliberation budgets, self-consistency, claim-level verification, citation gates, judge ensembles, calibrated abstention, cost cascades, semantic caching, deep-research loops, agent memory, MCP exposure, eval regression gates, and answer certificates. Use for deciding how much inference compute to spend, verifying claims before asserting them, avoiding overthinking/hallucination/verification-theater, and controlling inference cost.

~Alter Verify Human

@true-alter

Anyone can claim to be anyone and you are about to pay them, grant them access, or act on what they told you. Check first. Confirm a person is registered on ~alter by handle, member id or email, and test their verification tier, archetype and trait claims before you commit.

AI Identity Verification (Three-Anchor-One-Vote)

@zhaoxinghua09-cell

AI 身份核验与溯源工具（三锚一票方法论）：对任意 AI 智能体/对话/Agent 执行身份核验——账号锚（Account Anchor 登录身份/可验证凭证）、行为锚（Behavior Anchor 行为模式一致性）、记忆锚（Memory Anchor 长期记忆一致性）三维交叉核验，最终一票给出可信度结论（Level 1-4）。对标国际标准：W3C DID v1.1 / Verifiable Credentials / did:trail / MCP-I / EU AI Act Art.50-52 / C2PA 内容溯源。适用：判断对话 AI 是不是同一个人/同一个 Agent、AI 生成内容溯源、Agent 身份审计、医疗 AI 身份核验（细分第一）。触发词：AI 身份、身份核验、三锚一票、这是同一个AI吗、AI 溯源、Agent 身份审计、verify AI identity、AI provenance。

completion-audit

@raguets

Part of the Overpowered skill suite. Independently determine whether a claimed task outcome is actually complete by reconstructing completion criteria and checking fresh evidence. Use before declaring work done, accepted, migrated, reconciled, fixed, or delivered; not as a general planning loop.

gate-switch

@xu-jin-cs

Claim-verification gate engine for LLM agent workflows. Whenever an agent claims 'X is done / written / synced / verified', write X as a spec JSON of mechanical checks; the engine verifies each check and returns a verdict (A=pass / B=block with violations / CLARIFY / VIOLATION). Cures three chronic LLM failures: skipped work, partial delivery, fabricated claims. 声称 X 已满足，就机械核验 X——判定禁止手写，照抄输出。

Pullboard

@pullboard

Coordinate a long, complex build across a fleet of agents (or one long-running agent) on a shared board — so the plan survives context loss and reaches done....

Spraay Trust & Safety

@plagtech

Verify before you pay. Free address safety screening (phishing, sanctions, exploits, mixers), free token honeypot checks, free transaction decoding, plus paid multi-dimensional agent trust scores via ProofLayer — all through the Spraay x402 gateway. Vet any counterparty before funds move.

OMNI Research Verifier

@mustafa-cuda-dev

Autonomous verification engine that deconstructs claims, evaluates source credibility, and identifies contradictions.

PulseFeed — verify before you pay or install

@nikolife2016

Verify before you pay or install. Checks whether an x402 payment endpoint is safe to pay (liveness, scam scan, on-chain receiver, trust score) and whether an npm/MCP package is safe to install (install-script code execution, abandonment, missing repo/license). Use whenever you are about to pay an unknown API over x402/USDC, install an MCP server or skill, or need current x402 security incidents. Free, no API key.

Harness Control Layer

@warren2008-2020-spec

Use this skill when designing or operating a Harness-style control layer for OpenClaw setups with many skills, memory surfaces, safety-sensitive tools, playb...

Skill Sign

@songhonglei

Cryptographically sign and verify AgentSkill directories using Ed25519 (RFC 8032). Publishers sign skills with a private key; recipients verify with a public...

Clawhub Verify

@brick-blue

Before connecting to an MCP, A2A or x402 server somebody handed you, check whether it answers, which of its tools actually respond when called, what they charge, whether its card tries to instruct you, and what changed since the last look. Use when adding a server, following a link to a tool, or deciding whether to pay.

PASM Long-term Verification

@arronjack

Long-term verification for the PASM cognitive engine: memory retention, salience-aware eviction, emotion drift, persona saturation, forgetting curve, 6000-step soak. One command installs all (pip install pasm-agents, brings the pasm-skills base). See the 0.2.x upgrade note in SKILL.md.

Did that actually work? Tool report vs the wire

@aeneassoft

Did that actually work? Did my last action take effect, or did it only look like it did? Use this after any tool call whose effect matters — an exec, a write, a POST or PUT, a message send, a fetch you will build on — and before reporting success to the user. Cross-checks what the tool reported (ok/error) against what the wire showed for that same tool call: status codes (403, 5xx), empty 200 bodies, unsent uploads, no observable request at all. Returns exactly CONFIRMED, SILENT_FAILURE, UNVERIFIED or FAILED with the mismatch named. Do not use for reads you will not act on.

structure-check

@return054

Stop trusting "done." Structure Check mechanically verifies that everything the task required actually exists in the code — symbols defined, exports wired, functions actually called, no stub shells hiding behind a TODO. The agent writes a requirement checklist from the user's task, the tool checks every item, and the exit code decides (0 all pass / 1 something missing / 2 bad checklist). The agent may only quote the verdict, never overrule it. Zero dependencies, Node >= 18, comment-aware; full JS/TS support plus Python exists/notStub. Use after completing any coding task (trigger phrases - structure check, feature implemented, 触发词 - 结构检查, 功能是否实现, 实现核对, 检查代码结构, 交付核对).

EveryInfra Cross-check Research

@xiaoxihexiaoyu

Verify a current factual claim or produce a cited research answer through EverySearch crosscheck. Use when the user needs mechanism-distinct retrieval, source comparison, freshness checks, or an evidence-backed conclusion rather than a simple result list.

VEYA

@veyanet

Operate VEYA, the post-quantum execution and proof layer for AI agents on Robinhood Chain, through the hosted VEYA MCP server (51 veya_* tools). Prove and anchor content as BLAKE3 commitments on Veya.sol, verify transactions and commitments, create ML-DSA-44 identities and signatures, manage environments, agents, tool policy and Kyber-768 secure routing, run 2-of-3 consensus and AES-256-GCM sealed execution, store agent memory, browse the public registry, and send on-chain writes from your own wallet. Use when the user mentions VEYA, veyanet, Veya.sol, Robinhood Chain, on-chain proof or attestation of agent work, BLAKE3 commitments, ML-DSA, Kyber, post-quantum signatures, sealed or protected execution, agent consensus, or any veya_ tool.

Did my message actually go out? Delivery verified on the wire

@aeneassoft

Did my message actually go out? Was it delivered, or did the send only look successful? Use this after any message send whose delivery matters — a notification, a reply the user is waiting for, a forward to another chat — and before telling the user it was sent. Cross-checks the runtime's send result (message_sent success/error) against the wire: the request this process made to the channel API (api.telegram.org, discord.com, slack.com, graph.facebook.com, api.twilio.com, graph.microsoft.com) and its status code. Returns exactly DELIVERED, NOT_DELIVERED, CLAIMED_ONLY, CANCELLED, PENDING or NO_SEND. Do not use to confirm channels that deliver outside this process (WhatsApp Web, Signal, iMessage); for those it tells you the wire cannot confirm and what to do instead.

What did I actually do? Claims vs record

@aeneassoft

What do I believe I did, and what did I actually do? Did the things I told the user really happen? Use this before reporting a task as done, after a long session, and whenever you are about to say 'sent', 'published', 'fixed', 'installed' or 'verified'. Reads your own session transcript for success claims and checks each against the record: tool outcomes, outbound sends and their delivery, POST/PUT status codes on the wire. Lists every claim as SUPPORTED, UNSUPPORTED or CONTRADICTED and names failed actions you never mentioned. Returns exactly RECORD_MATCHES_CLAIMS, CLAIMS_EXCEED_RECORD, RECORD_EXCEEDS_CLAIMS or NO_CLAIMS. Do not use before any action has been taken.

Did I do what was asked? Request vs reply

@aeneassoft

Did I do what was asked? Does my reply match the request, and did the processing in between make sense? Use this after replying to a message, before marking a task handled, when a heartbeat or cron turn produced many tool calls for a small message, and when the user says 'that's not what I asked'. Compares the inbound request (keyword fingerprint, requested actions: send/fetch/create/change/run) with the reply that went out (keyword fingerprint, delivery) and the actions actually taken between them (tool calls, hosts, failures, seconds). Returns exactly COHERENT, PARTIAL, INCOHERENT or NO_PAIR with reasons. No model is consulted and no texts are stored. Do not use to judge tone or factual correctness — it cannot.

ClawPhylax: did that actually work? All 22 questions

@aeneassoft

ClawPhylax: did that actually work? Did my message actually go out? Is this safe to send? Why did my request fail? Should I stop and ask? Am I going in circles? What did I lose in compaction? What did I actually do? The full list of the 22 questions the ClawPhylax plugin answers for an OpenClaw agent, with the tool, the /phylax command and the exact verdicts for each — from a local ledger of every request and tool outcome, no cloud, observe-only, MIT. Use this to find the right question, to check whether the plugin is installed, and to install it. Do not use it in place of the specific skill once you know your question.

evidence-traps

@hawkaa

Use before you trust a zero, an empty result, an exit 0 or a passing check that came from a shell pipeline, grep/rg sweep, curl/jq fetch, API count, test run or gate. Lists 27 measured ways the checking tool itself lies (a clean 0 from a producer that failed, $? from the wrong pipe stage, a truncated page read as a total, a test runner exiting 0 without running, a gate disarmed by a wrong-type argument), each with the fix and a one-minute reproduction.

multisim-compare-experiments

@yxy050208

比较两个已注册 Multisim 实验的电路、设置、测量、验证和产物；use when evaluating a baseline against a candidate or explaining changed simulation results

Agent Errata: replicate a finding

@hawkaa

Replicate one Agent Errata finding on your own stack and file the result. Each entry is a tool-level failure mode (a wrong zero, a silent exit 0, a gate that passes everything) with a check you can run in under a minute.

AI Code Verification Guard

@return054

Deterministic code-verification guard for AI coding agents. Runs syntax, type, lint and test checks before and after code changes, and decides PASS / FAIL / REGRESSION / UNVERIFIED purely by script exit codes. The agent may only quote the verdict, never overrule it. Use when the user asks to modify code, implement features, fix bugs, or verify change reliability (trigger phrases - verify this, run the tests, regression check, confirm it works, 验证一下, 跑测试, 回归检查, 确认没问题). Not for general code review or security audits.