AI & ML

uygnoey Skill

A structured framework for planning and rolling out agentic AI across an enterprise, emphasizing workforce enablement, process redesign with human oversight, and scaling into durable product capabilities.

uygnoey Bundle

Plan a Claude rollout for a financial services firm using the product matrix, the 10 pre-built finance agent templates, and the foundation/pilot/scale adoption phases described in Anthropic's "Deploying Claude across financial services" guide. Use when an AI leader needs to decide which Claude product belongs in which workflow, where to place pre-built finance agent templates, and how to sequence a phased deployment that survives review by risk, compliance, and IT.

uygnoey Bundle

Hand a frontier model a whole job in Claude Cowork instead of prompting it step by step — pick the model and effort level, brainstorm from a bare idea, brief with context rather than rules, delegate the approach/procedure/timing, and review the plan panel while it runs. Use when a Cowork task spans dozens of steps or several days, when a job needs multiple tools and a series of judgment calls, when you are unsure which model or effort setting fits, when a long conversation is burning usage, or when skills and memory files written for an older model may be holding a newer one back.

ksmaster03 Skill

ทำ threat modeling ตั้งแต่ตอนออกแบบ — วาด DFD + trust boundary, ใช้ STRIDE หา threat รายองค์ประกอบ, จัดลำดับความเสี่ยง แล้วแปลงเป็น security requirement + mitigation ที่ส่งต่อให้ทีม dev/test ใช้ได้จริง. Model the system, find what can go wrong (STRIDE-per-element), rank risk, and turn it into actionable security controls. Trigger เมื่อผู้ใช้พิมพ์ /threat-model หรือขอ "threat model / STRIDE / DFD / security design / abuse case / attack surface / วิเคราะห์ภัยคุกคาม / ความเสี่ยงความปลอดภัย".

ksmaster03 Skill

Work the task as a 3-model team — Opus (you, the main agent) plans, decides, reviews and integrates; a Sonnet subagent does the heavy lifting (multi-file coding, design, deep analysis); a Haiku subagent does fast cheap well-scoped work (search, read+summarize, mechanical edits, checks). Delegate via the Agent tool's `model` override, run independent work in parallel, then synthesize. Trigger when the user types /teamagent, or asks to "ทำงานเป็นทีม / work as a team / use haiku+sonnet+opus / แบ่งงานข้ามโมเดล / delegate across models".

uygnoey Bundle

Build agents whose sessions run for tens of minutes to hours, using Outtake's four-stage process and four hard-won learnings from their Recon Agent. Use when context compaction and behavioral drift are real rather than theoretical, when an agent keeps ignoring a system-prompt instruction no matter how it is worded, when deciding whether to stay in Claude Code or graduate to the Claude Agent SDK, when choosing between purpose-built tools and a filesystem plus bash, when manual transcript review no longer scales, or when the agent must operate in an environment that is actively trying to hijack it.

uygnoey Skill

Help an organization map Microsoft 365 Copilot entry points (Researcher, Copilot Studio) to appropriate Claude model choices mentioned in the announcement.

uygnoey Bundle

A practical framework for implementing AI agents in healthcare and life sciences, focusing on interoperability, latency, compliance, and maintaining human clinical authority.

uygnoey Bundle

Use this skill to decide whether to run Claude Managed Agents with self-hosted sandboxes and/or MCP tunnels, and to collect the minimum deployment details and security considerations for each option.

uygnoey Bundle

Helps choose and apply common AI agent workflow patterns (sequential, parallel, evaluator–optimizer). Use when designing multi-step agent systems, deciding whether to parallelize subtasks, or improving output quality with iterative evaluation.

uygnoey Bundle

Patterns and reusable prompt templates for creating “dynamic workflows” in Claude Code to orchestrate multi-agent harnesses for complex, high-value tasks.

uygnoey Bundle

Decide when to use multi-agent architectures (context protection, parallelization, specialization), how to decompose work by context boundaries, and how to add a verification subagent pattern with explicit pass/fail criteria.

uygnoey Bundle

Choose which Claude model class — Mythos/Fable, Opus, Sonnet, or Haiku — to run a workload on, and use effort level to dial in the quality/speed/cost balance. Use when picking a model for a new production workload, when revisiting a model choice after evals show a gap, when a workload is latency- or cost-sensitive enough that a lower class is worth testing, or when deciding whether to pair a cheaper worker model with a more capable advisor. Default recommendation is to start with the most intelligent generally available model and tune down, because cost-per-task is often lower on more capable models even when price-per-token is higher.

uygnoey Bundle

Shift an agentic codebase from "agents emit code" to "agents emit verifiable specifications." Use when agents generate code faster than the team can review it and verification has become the real bottleneck, when control logic is scattered across routes, database constraints, service code, background jobs, and docs so no one can state the system's state machine, when agents need to change operational behavior safely without a full CI cycle, or when designing an autonomous build-and-operate loop and deciding what humans must still coordinate. Based on Datadog's Temper, a "universal machine tool" that verifies agent-emitted specs through four independent layers before anything runs.

uygnoey Skill

Guidance for designing workflows where MCP servers provide tool access and skills encode procedural expertise, sequencing, and output standards.

uygnoey Bundle

Patterns for connecting agents to production systems, including when to use direct APIs vs CLIs vs MCP, and how to design MCP servers and clients for scale, security, and context efficiency.

uygnoey Bundle

Decide where to place Claude Code instructions (CLAUDE.md, rules, skills, subagents, hooks, output styles, or appended system prompt) based on load timing, compaction behavior, context cost, and authority.

uygnoey Bundle

Package domain expertise into scalable Agent Skills using progressive disclosure (metadata, SKILL.md, and on-demand references).

uygnoey Bundle

Maintain Agent Skills over time by defining evals, running benchmarks, comparing variants, and tightening skill descriptions to reduce false triggers.

uygnoey Skill

Establish safer, more autonomous boundaries for agentic coding by applying filesystem and network isolation (a "native sandboxing" approach) so the agent can operate with fewer per-action permission prompts.

uygnoey Bundle

Build an analyst agent that works alongside human experts in a high-stakes domain — surfacing insights and forming opinions on exposure, while logging its reasoning, testing actions in sandboxes, and requiring expert approval before decisions count. Use when designing an AI teammate for risk, compliance, or another regulated analytical function; when outputs must be auditable back to their reasoning; when analysis depends on proprietary data plus judgment the firm will not delegate; or when the job is to explain how numbers changed since yesterday rather than answer one-off questions.

uygnoey Bundle

Best-practice playbook for making Claude Code reliable in large repositories by building a harness (CLAUDE.md layering, hooks, skills, plugins, MCP, LSP) and establishing ownership.

uygnoey Bundle

Run a fleet of production agents as code in a git repository — every agent a config file plus a prompt, every change a reviewed pull request, every improvement harvested from human feedback rather than retraining. Use when AI automations are scattered across individual laptops with no shared view of what exists or what it costs, when non-developers need to ship production agents without the dev team becoming the bottleneck, when agents need version history and an audit trail instead of an admin screen, or when you need a defensible way to decide which tasks deserve an agent at all.

uygnoey Bundle

Set up and run a team where people and agents work together in shared channels. Use when agents have workspace access but return shallow answers because context lives in private threads; when deciding which work an agent owns and where a human reviews; when defining roles for several specialized agents instead of one general assistant; when adoption is stalling and needs peer demonstration rather than mandate; or when leadership asks for evidence of value and only usage metrics are available.

uygnoey Bundle

Keep prompt cache hit rate high in long-running agents by following the patterns Claude Code's team uses — lay out static-then-dynamic, push updates into messages instead of editing the system prompt, never change models or tools mid-session, model state transitions as tools (Plan Mode), defer tool loading instead of removing tools, and implement compaction as cache-safe forking. Treat cache breaks as incidents.

uygnoey Bundle

Rightsize the context you hand a newer-generation model — system prompt, CLAUDE.md, skills, tool descriptions, and references — by removing rules the model no longer needs and replacing them with judgment, interface design, and progressive disclosure. Use when a system prompt or CLAUDE.md has grown long, when instructions contradict each other, when a skill over-constrains the model, when tool guidance is duplicated between the system prompt and the tool description, or when migrating an agent built for an older model generation.

uygnoey Skill

A practical playbook for running a short pre-launch model evaluation window that blends automated evals with hands-on “vibe checks.”

uygnoey Bundle

Build AI output that a professional can put their name on — grounded in authoritative content, checked for citation validity before a human ever reviews it, and transparent enough to defend when the stakes are high. Use when the work product has to hold up under professional review (legal, tax, accounting, compliance), when deciding what to require of a model before trusting it with that work, or when deciding what the human in the loop is actually accountable for.

uygnoey Bundle

Decide whether a new model is actually better by putting it through a real day of work with your highest-taste engineers and asking whether the code is something they would keep, rather than trusting a benchmark score. Use when a model aces a benchmark but you are unsure it will hold up in practice, when building an anti-slop internal benchmark, when a score jump needs corroboration, or when a team keeps arguing for weeks about whether a release is an improvement.

uygnoey Bundle

Run a sales inbox as an hourly drafting loop — a thin system prompt, a curated knowledge base of common questions and best answers, and a per-rep voice profile — so every thread needing a reply gets a draft the rep reads, edits, and sends. Use when inbound volume is consuming hours a day on repeated questions; when replies must stay factually anchored to approved product answers; when drafts should sound like the person sending them; or when adding lighter monitors for no-shows, prospects going dark, and new leads that need a first touch.

uygnoey Bundle

Rebuild a product from sequential, separate skills into a single agent that plans and orchestrates across tools in real time, and test candidate models on the ability that architecture depends on — planning and effective tool use. Use when a product is a collection of one-shot features that users have to sequence themselves, when evaluating whether a model can carry an agent across hundreds of tools, or when deciding what to measure about an AI initiative beyond cost per task.

uygnoey Bundle

Decide whether a new model is worth adopting by running it against a domain-specific benchmark head-to-head with the model it would replace, rather than reading public scores — build tests that replicate the real jobs your users do, expand the benchmark with each release, and separate retrieval quality from synthesis quality when reading the results. Use when a new model ships and someone asks whether to switch, when building an internal eval for a high-accuracy domain, or when a benchmark result jumps and you need to know what actually improved.

uygnoey Bundle

Hand a model the work that was previously reserved for your most senior or most specialized engineers — core changes touching interdependent parts — by briefing the goal and the why rather than every step, front-loading a question round, letting it run unattended, and gating on human review before shipping. Use when a core system change is blocked waiting for a specific engineer to free up, when deciding whether a non-engineer can own a piece of technical work, or when calibrating how much specification a model actually needs.

uygnoey Bundle

Evaluate a new model for an app-generation product by running it across different app types and measuring latency, cost, and build errors, plus stress builds that exercise unusual capabilities — then read the signals that matter for production, such as turns to completion, first-prompt completeness, and whether prompt changes break the cache. Use when a new model ships and someone asks whether to switch the generation engine, when designing an eval suite for a codegen or app-building product, or when an eval passes but production cost or latency regresses anyway.

uygnoey Bundle

Move a product or an internal platform from bolted-on AI features to an agent-first design where people and agents work on the same items. Use when AI features shipped, adoption looked fine for a month, and usage has flattened into occasional summarization; when deciding whether to embed agents into an existing workflow or rebuild the workflow around them; when agents live in a chat surface parallel to the real work and context must be pasted in by hand; when pilots stall before production because governance, permissions, and reliability were never designed in; or when you need concrete per-function agent jobs instead of one general assistant.

darker2016 Bundle

Makers 开发专家团主理人 Skill — 协调前端、后端、AI Agent 三位专业成员， 在 EdgeOne Makers 平台上完成 Web 全栈开发与部署任务。 触发词：全栈开发、搭建并部署、帮我做一个 Web 应用、EdgeOne Makers 项目、 在 EdgeOne 上开发、一站式交付。