plugins/development/skills/agent-design/SKILL.md
Knowledge base for designing AI agents — persona crafting, system prompt patterns, multi-agent orchestration, memory design, evaluation. Use when creating or improving an autonomous AI agent persona, writing agent templates, or debugging agent behavior. NOT a workflow — this is reference material to inform decisions.
npx skillsauth add petrogurcak/skills agent-designInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
Reference material for designing production AI agents. Use alongside development:create-persona (scaffolding) and agent template writing.
Announce: "Nacitam agent-design knowledge base pro informovane rozhodnuti."
development:create-personadevelopment:development-workflowSystem prompt = casting, ne programování. Model nehraje instrukce — hraje postavu, kterou si odvodil z kontextu.
Klíčové principy:
Anti-patterns:
Zdroj: https://www.anthropic.com/research/persona-selection-model
275 archetypů v geometrickém persona prostoru. Hlavní osa = jak moc je persona "assistant-like."
Zdroj: https://www.anthropic.com/research/assistant-axis
Mechanistická interpretabilita — konkrétní aktivační směry v modelu odpovídají traits (sycophancy, hallucination tendency, agreeableness).
Zdroj: https://www.anthropic.com/research/persona-vectors
Reálný operační spec pro agentic behavior (CC0 licence).
requires_human rules a safety sekcí v agent templatesZdroj: https://www.anthropic.com/constitution
Production-tested vzory pro agent architektury:
Anti-patterns:
Zdroj: https://www.anthropic.com/research/building-effective-agents
Velké constitutional dokumenty způsobují truncation a snižují compliance.
Zdroj: https://arxiv.org/pdf/2602.02584
Thought → Action → Observation — základ většiny production agentů.
Zdroj: https://arxiv.org/abs/2210.03629
Dva fundamentální primitiva:
Kdy co:
Zdroj: https://cookbook.openai.com/examples/orchestrating_agents
| Topologie | Kdy použít | Naše použití | |-----------|-----------|--------------| | Centralized (orchestrator → workers) | Jasná hierarchie, definované subtasky | orchestrator → persony | | Peer-to-peer | Agenti potřebují komunikovat přímo | agent-team-development teammates | | Sequential pipeline | Output A = input B | learning-advisor → proposal → apply |
"80% effort into task design, 20% into agent design."
Tři povinné fieldy pro funkčního agenta: role, goal, backstory. Backstory není fluff — ovlivňuje jak agent interpretuje edge cases.
Zdroj: https://docs.crewai.com/en/concepts/agents
| Typ | LLM implementace | Naše použití |
|-----|-------------------|-------------|
| Sensory | Raw input (current request) | Template input variables |
| Short-term | Context window | Konverzace v rámci jednoho běhu |
| Long-term | External storage + retrieval | memory/ soubory, agent memory v projektech |
Většina agent systémů (včetně našich) má jen sémantickou paměť. Episodická chybí — agent neví co se stalo minule v tomto konkrétním projektu.
Praktický závěr: Agent memory files by měly mít sekci pro trendy/historii (episodická) vedle faktů/pravidel (sémantická).
Zdroj: https://arxiv.org/abs/2502.06975
Paměť jako knowledge graph, ne flat store. Každá memory unit je:
Implikace: Naše memory/ soubory by mohly benefitovat z cross-referencí mezi agenty (analytics memory odkazuje na copywriting findings).
Zdroj: https://arxiv.org/abs/2502.12110
Zdroj: https://www.zenml.io/blog/what-1200-production-deployments-reveal-about-llmops-in-2025
Zdroj: https://cognition.ai/blog/devin-annual-performance-review-2025
Agent generuje verbální feedback na vlastní failed trajektorie a ukládá ho jako episodickou paměť pro další pokus.
Zdroj: https://arxiv.org/abs/2303.11366
Tři vlastnosti, jejichž kombinace = katastrofální zranitelnost:
Každý náš agent co crawluje web (copywriting-reviewer, seo-monitor) má 1+2. Proto máme Hostile Content Defense sekce v templates.
Zdroj: https://simonw.substack.com/p/the-lethal-trifecta-for-ai-agents
Při psaní nového agent template zkontroluj:
development
Builds a pre-launch social proof strategy through structured beta programs using D'Souza Brain Audit interviews. Use when launching new products/services and need compelling testimonials, planning a beta cohort, designing interview questions to harvest objection-busting social proof, improving video testimonials for landing pages, or designing case studies with metrics. Trigger phrases include "beta tester program for testimonials", "pre-launch social proof", "Brain Audit testimonial framework", "case study harvest", "reverse testimonial", "video testimonial mechanics", "social proof landing page", "sběr referencí", "beta tester program", "testimonial pro landing page", "social proof před launchem", "rozhovor s klientem", "case study sběr", "reference před spuštěním". NOT for ongoing case study production (use growth-hacking case-study approach), offer design (use offer-creation), or conversion optimization (use ux-optimization).
development
Use when planning a product launch and the product type is unclear or could be either generic (SaaS/app/physical) or info-product. Routes between marketing:launch-strategy (generic launches) and marketing:info-product-launch (courses, memberships, ebooks, cohorts, communities). Trigger phrases - "launch", "spuštění", "go-to-market", "product launch", "release strategy", "uvedení na trh", "launch plan", "spuštění produktu", "launch sequence", "launch strategy". Do NOT trigger when product type is already clear (use specific skill directly).
testing
Specialized 8-week launch cadence for info-products — online courses, cohort programs, memberships, communities, ebooks, masterminds. Combines Jeff Walker's Product Launch Formula (Seed/Internal/JV variants, PLC sequence, open-cart day-by-day) with Stu McLaren's membership mechanics (closed cart, Success Path) and Hormozi Grand Slam Offer stacking. Use when planning "launch online kurzu", "info-product launch", "PLF launch", "course launch", "membership launch", "cohort launch", "ebook launch", "open cart close cart", "8-week launch of online course", "beta cohort to launch sequence", "spuštění kurzu", "launch členské sekce", "open cart strategie". Differentiates from marketing:launch-strategy (generic SaaS/app launches) — info-product-specific. NOT for SaaS launches, physical products, or services.
development
Use when releasing an Expo/React Native mobile app to App Store and Google Play - covers eas submit, ASC "Submit for Review", Play promote Internal→Production, OTA update, and decoding common silent failures (Apple agreement expiry, missing English locale, Background Location declaration, web bundle failure on react-native-maps).