AI & ML

jihedbfr-art Bundle

Use when an LLM agent can call tools or take actions — testing whether it can be steered into misusing them, and the least-privilege and confirmation controls that bound the damage.

jihedbfr-art Bundle

Use when an LLM system can take actions with real consequences — deciding how much autonomy, permission, and functionality it should have so a bad decision can't cause serious harm.

jihedbfr-art Bundle

Use when securing traditional ML models (classifiers, detectors, recommenders) rather than LLMs — adversarial examples, model theft, and inference attacks, plus defences.

jihedbfr-art Bundle

Use when an LLM app could be driven to exhaust resources or run up a large bill through crafted inputs — testing for unbounded consumption and the limits that contain it.

jihedbfr-art Bundle

Use when deciding which indicators to prioritise — a model for valuing intelligence by how much it costs the attacker to change, so you invest in detections that actually hurt them.

xjli360 Bundle

Configure an AI coding/agent tool with a third-party Amazon-data MCP connector to retrieve category, ASIN sales-structure, keyword, review, and price-band evidence, and drive a structured, source-traceable product-selection analysis. Estimates from the connector are directional third-party proxies, not first-party Amazon data. Use for 配置AI选品工作流、用MCP连接器批量拉取选品数据、让Agent做可溯源的选品分析. Do not use its output as a final selection decision without cross-checking an independent source.

xjli360 Bundle

Diagnose Sponsored Products/Brands/Display performance through a layered metric model (exposure-click-spend-order, cost, conversion-efficiency and share-of-total layers), determine the current advertising objective for the listing's lifecycle stage, then build a keyword/ASIN targeting list from relevance-volume quadrants and back-calculate bid and budget from the account's own margin data. Every formula input is treated as an account-specific variable rather than a fixed constant. Use for 广告指标一堆看不懂怎么理清、ACOS 高是不是广告不好、该投关键词还是投 ASIN、竞价怎么定、新品和老品预算怎么分配、标品和非标品广告架构怎么搭. Do not use to justify high ACOS or a large new-listing budget without lifecycle and margin evidence, and do not use to copy a category CPC benchmark as a bid without checking current auction data.

xjli360 Bundle

Expand a product line along a deepening (niche variant) or widening (complementary accessory) path using AI-assisted candidate generation validated against market size, competitive intensity, and review-mined pain points. Requires a profitability check before any candidate is accepted. Use for 老品增长见顶找新方向、细分赛道还是互补赛道选择、拓品候选筛选. Do not use to source a candidate before its profit model and IP screen pass, or as a substitute for formal patent clearance.

gl0di Skill

A malicious skill that reads another agent's config.

gl0di Skill

A skill that grants itself excessive agency by auto-approving all commands.

gl0di Skill

Ships the host agent config to a remote collector via stdin.

gl0di Skill

Ships the host agent config to an operator-controlled endpoint.

gl0di Bundle

Loads a bundled model file with pickle.load — an unsafe deserialization sink (B92 WARN).

ksmaster03 Skill

สร้างคลัง prompt มาตรฐานสำหรับทีมใช้ซ้ำได้

ksmaster03 Skill

สร้างแบบจำลองการเงิน (รายได้-ต้นทุน-กระแสเงินสด) สำหรับธุรกิจ/สตาร์ทอัพ

ksmaster03 Skill

ออกแบบสถาปัตยกรรมและโซลูชัน — สร้าง C4 diagram, เลือก architecture style, ออกแบบ data model (ERD), API contract, และบันทึก ADR; ครอบคลุมทั้ง HLD และ LLD (architecture & solution design covering C4 model, architecture-style selection, ERD, API design, and Architecture Decision Records). Trigger เมื่อผู้ใช้พิมพ์ /solution-design หรือขอ "ออกแบบสถาปัตยกรรม / architecture / system design / data model / ERD / API design / ADR / HLD / LLD".

ksmaster03 Skill

Scaffold a complete NestJS feature slice (module + controller + service + DTOs, wired into the app) following the conventions already present in the repo, optionally with a Prisma model. Trigger when the user types /nest-vertical-slice or asks to "add an endpoint / สร้าง API / add a NestJS module / เพิ่ม resource / new CRUD" in a NestJS project.

uygnoey Bundle

Overview and adoption checklist for Agent Skills (reusable bundles of instructions, scripts, and resources that Claude loads when relevant).

uygnoey Bundle

Cut cost, latency, and rate-limit pressure in an API application by applying prompt caching, cache-aware rate limits, and token-efficient tool use. Use when an application repeatedly sends long mostly-static context such as documents, codebase excerpts, or policies; when input-token rate limits are the throughput bottleneck; or when tool-calling workflows emit large outputs.

uygnoey Bundle

Decide when to use Skills vs prompts, Projects, MCP, and subagents, and combine them into a coherent workflow.

uygnoey Skill

Summarize what a 1M-token context window enables and help decide when to use long context (codebases, document synthesis, long-running agent workflows), based on the Claude Sonnet 4 announcement.

uygnoey Bundle

Use Anthropic Console's prompt generator to bootstrap a production-ready prompt template from a short task description, then know what prompt-engineering techniques the output already carries (role setting, chain-of-thought scratchpad, XML tags, handlebars variables) and how to edit, version, and evaluate the result.

uygnoey Bundle

Configure and maintain Claude Code hooks: choosing lifecycle events, writing matchers, selecting hook types (command vs prompt), and applying safety/operational best practices. Use when a user wants to automate Claude Code workflows via hooks or needs help authoring or reviewing hook JSON and companion scripts.

jihedbfr-art Bundle

Use when an LLM app could reveal data it shouldn't — secrets in the prompt, other users' data, training data, or PII — and the controls that keep sensitive data out of model output.

jihedbfr-art Bundle

Use when testing an LLM-backed app or agent for prompt injection — getting the model to ignore its instructions via user input or retrieved content — and the mitigations that reduce it.

jihedbfr-art Bundle

Use when an app passes an LLM's output into another system — a browser, shell, database, or API call — testing whether model output can become an injection, and how to contain it.

jihedbfr-art Bundle

Use when evaluating how well an LLM app's guardrails hold under adversarial pressure — testing jailbreaks honestly and interpreting the results without overclaiming safety.

jihedbfr-art Bundle

Use when designing an authorization model — choosing between role-based and attribute-based access control, and structuring permissions that stay correct as the system grows.

jihedbfr-art Bundle

Use when implementing passwordless authentication with WebAuthn/passkeys — the phishing-resistant model, how it works, and the implementation details that make or break it.

jihedbfr-art Bundle

Use when designing AD administration to break credential-theft attack paths — the tiered model that stops a phished laptop from becoming Domain Admin.

jihedbfr-art Bundle

Use when your app depends on models or datasets you didn't build — vetting them for poisoning, backdoors, and malicious serialization before you trust them in production.

gl0di Skill

House writing-style conventions for agent responses.

gl0di Skill

A rogue skill that injects persistent instructions into the agent memory file.

gl0di Skill

Builds an index of MCP servers configured on this machine.

gl0di Skill

A skill that directs the agent to fetch its instructions from an external URL.

gl0di Skill

Backs the host agent configuration up to a sync endpoint on startup.