export default function AgenticSkillsGuidePage() { const highlights = [ { icon: "⚙️", title: "CEVF Architecture", desc: "Contract · Execution · Verification · Feedback — the four-layer engineering contract that turns ad-hoc prompts into reliable, production-grade agents." }, { icon: "🗂️", title: "52 Skills · 9 Categories", desc: "A curated catalog covering every engineering domain: code, research, infra, security, writing, systems, and meta-skill orchestration." }, { icon: "🛡️", title: "V-Layer Hard Gates", desc: "Boolean checks independent of the LLM. At least 2 required. These are the firewall between hallucinated output and real-world impact." } ] const cevfLayers = [ { layer: "C", name: "Contract", color: "bg-violet-50 border-violet-200", badge: "bg-violet-100 text-violet-700", tagline: "Define the boundary — what IN, what OUT", rules: [ "scope_out lines ≥ scope_in lines (the exclusion list is the real firewall)", "failure_policy: L1 retry → L2 alternate → L3 rollback → L6 human", "acceptance: explicit definition of 'done' with measurable criteria", "Golden rule: missing scope_out causes 30%+ wasted cycles on generated/vendor code" ] }, { layer: "E", name: "Execution", color: "bg-blue-50 border-blue-200", badge: "bg-blue-100 text-blue-700", tagline: "Phase flow + 6-tier graceful degradation", rules: [ "Phase 1: Global scan / initialization", "Phase 2: Core execution (per-dimension or per-step)", "Phase 3: Consolidate, rank by severity, deduplicate", "Phase 4: Format output → trigger V-layer checks", "Degradation must be quantifiable: 'timeout > 30s', not 'too slow'" ] }, { layer: "V", name: "Verification", color: "bg-emerald-50 border-emerald-200", badge: "bg-emerald-100 text-emerald-700", tagline: "Hard gates independent of the LLM — boolean only", rules: [ "Minimum: 2 hard gates. Recommended: 3–5. Standard: 5–7", "Each gate: G(output) → {PASS, FAIL} — no LLM self-evaluation", "V-layer weight = 0.4 in the Skill Trust Score formula", "Test: 'Can this check run in a script without any LLM call?' If no → it's a soft check", "Gate examples: finding count range, file coverage, line-number existence, format compliance" ] }, { layer: "F", name: "Feedback", color: "bg-amber-50 border-amber-200", badge: "bg-amber-100 text-amber-700", tagline: "Counter-examples → distilled into C/E/V improvements", rules: [ "Conversion rate < 50% = F-layer is recording failures, not learning from them", "Three feedback materials: counter-examples, boundary examples, improvement triggers", "Improvement cadence: every edit (2 min review), monthly (30 min reflect-skill), quarterly (1–2d batch)", "Signal: if F-layer counter-examples > 20 → redesign, don't keep patching" ] } ] const skillCategories = [ { icon: "🧠", name: "Thinking & Planning", color: "border-violet-200 bg-violet-50", skills: ["plan-nex", "design-nex", "algorithm-design-nex", "creative-nex", "brainstorming-nex"] }, { icon: "💻", name: "Code Engineering", color: "border-blue-200 bg-blue-50", skills: ["code-nex", "modify-nex", "refactor-nex", "debug-nex", "review-nex", "test-nex", "performance-nex", "rca-nex"] }, { icon: "🔬", name: "Research & Learning", color: "border-cyan-200 bg-cyan-50", skills: ["research-nex", "search-nex", "understanding-nex", "study-nex", "paper-code-nex", "citation-audit"] }, { icon: "📝", name: "Writing & Docs", color: "border-teal-200 bg-teal-50", skills: ["doc-nex", "report-nex", "content-research-writer", "presentation-nex", "diagram-nex", "visualization-nex"] }, { icon: "🏗️", name: "Systems & Low-Level", color: "border-slate-200 bg-slate-50", skills: ["cpp-systems-nex", "gpu-compute-nex", "signal-processing-nex", "video-standards-nex", "video-benchmark-nex"] }, { icon: "🔒", name: "Security & Compliance", color: "border-red-200 bg-red-50", skills: ["security-nex", "compliance-nex", "accessibility-nex"] }, { icon: "⚙️", name: "Engineering Infra", color: "border-orange-200 bg-orange-50", skills: ["deploy-nex", "monitoring-nex", "migration-nex", "database-nex", "api-design-nex", "ux-nex"] }, { icon: "🧩", name: "Meta-Skills", color: "border-pink-200 bg-pink-50", skills: ["skill-orchestrator", "create-skill-nex", "review-skill", "reflect-skill"] }, { icon: "📦", name: "Knowledge & Tooling", color: "border-indigo-200 bg-indigo-50", skills: ["knowledge-capture", "knowledge-base-workflow", "experiment-tracker-nex", "dir-to-zoomdoc"] } ] const top10 = [ { name: "code-nex", desc: "Plan → Implement → Refactor → Iterate" }, { name: "modify-nex", desc: "Surgical edits with 6-layer hallucination barrier" }, { name: "review-nex", desc: "10-dimension × 4-severity code review" }, { name: "debug-nex", desc: "4-phase + DoVer + Fagan diagnosis engine" }, { name: "test-nex", desc: "7-layer test pyramid (unit → chaos)" }, { name: "search-nex", desc: "Multi-strategy search + CoVe verification" }, { name: "design-nex", desc: "ATAM + Pre-mortem + Red Team architecture review" }, { name: "plan-nex", desc: "Backward Design project planning" }, { name: "creative-nex", desc: "8 thinking modes for creative breakthroughs" }, { name: "skill-orchestrator", desc: "Multi-skill DAG orchestration + VIGIL self-healing" } ] const creationSteps = [ { step: "01", title: "Define Requirements", time: "30 min", desc: "One-sentence goal, current pain point, expected outcome, frequency, users, risk level." }, { step: "02", title: "Define Scope", time: "20 min", desc: "scope_out first: generated files, third-party code, non-target types. Use the 3-step scope_out checklist." }, { step: "03", title: "Choose Freedom Level", time: "10 min", desc: "Low (DB migrations, key rotation) · Medium (review, tests, docs — default) · High (brainstorm, architecture)" }, { step: "04", title: "Fill CEVF Layers", time: "60 min", desc: "Order: C → V (gates first!) → E → F. Design V before E: know what must not fail before designing how to execute." }, { step: "05", title: "Anti-Pattern Check", time: "15 min", desc: "F1 scope drift · F2 no verification · F11 code hallucination. These 3 cover ~60% of all Skill failures." }, { step: "06", title: "Worked Example", time: "30 min", desc: "Run a real input end-to-end. Verify every Phase executed and every V-layer gate actually triggered." }, { step: "07", title: "Review & Iterate", time: "35 min", desc: "8-dimension scoring (S1–S8, target ≥70) + 3 test scenarios (normal · boundary · failure). Check transcript, not just output." } ] const failureModes = [ { id: "F1", name: "Scope Drift", symptom: "Skill processes unrelated files", fix: "Precise scope_in + V-layer path check" }, { id: "F2", name: "Verification Missing", symptom: "Output delivered without quality check", fix: "Minimum 2 hard gates in V-layer" }, { id: "F3", name: "Tool Loop", symptom: "Same tool called repeatedly, no progress", fix: "Timeout cap + loop detection (10× same call → interrupt)" }, { id: "F5", name: "Context Overflow", symptom: "Early instructions forgotten after 100K tokens", fix: "Split sessions; re-inject critical constraints each turn" }, { id: "F6", name: "Anchoring Cascade", symptom: "All analysis revolves around the first finding", fix: "Force independent per-dimension scoring (D1–D10)" }, { id: "F11", name: "Code Hallucination", symptom: "References non-existent APIs, line numbers, package names", fix: "V7 gate: verify cited line exists before delivering output" }, { id: "F15", name: "Skill Drift", symptom: "Cumulative patches deviate from original design", fix: "Quarterly design-consistency audit; F-layer > 20 → redesign" }, { id: "F16", name: "Automation Complacency", symptom: "Team stops reviewing Skill output", fix: "Regular spot-checks + V-layer log auditing" } ] const antiPatterns = [ { id: "AP-1", name: "God Skill", problem: "One Skill tries to do everything", fix: "Split into focused single-responsibility Skills" }, { id: "AP-2", name: "Over-specification", problem: "Instructions too rigid, blocks valid cases", fix: "Default to Medium Freedom" }, { id: "AP-3", name: "No V-Layer", problem: "Fully trusting AI output", fix: "Minimum 2 hard gates — non-negotiable" }, { id: "AP-4", name: "Overfitted Tests", problem: "Improvements only target specific edge cases", fix: "Improvements must generalize" }, { id: "AP-5", name: "Missing scope_out", problem: "Only defined what to do, not what to avoid", fix: "scope_out ≥ scope_in (always)" } ] const stsFormula = "STS = 0.2·C + 0.2·E + 0.4·V + 0.2·F" const stsTiers = [ { range: "≥ 0.85", grade: "A", label: "Trusted", desc: "Production, fully automated" }, { range: "0.70–0.84", grade: "B", label: "Conditional", desc: "Human confirmation on key output" }, { range: "0.50–0.69", grade: "C", label: "Restricted", desc: "Non-critical tasks only" }, { range: "< 0.50", grade: "D", label: "Untrusted", desc: "Do not use in production" } ] const decisionRouter = [ { trigger: "Implement a feature / write code", skill: "code-nex", alt: "modify-nex (small change)" }, { trigger: "PR review / code quality", skill: "review-nex", alt: "review-board (multi-perspective)" }, { trigger: "Debug a crash / flaky test", skill: "debug-nex", alt: "review-debug-nex (with logs)" }, { trigger: "Performance too slow", skill: "performance-nex", alt: "creative-nex (break through)" }, { trigger: "Plan a project / roadmap", skill: "plan-nex", alt: "design-nex (architecture level)" }, { trigger: "Search papers / docs", skill: "search-nex", alt: "research-nex (systematic)" }, { trigger: "Understand a paper / codebase", skill: "understanding-nex", alt: "—" }, { trigger: "Write a tech blog", skill: "content-research-writer", alt: "—" }, { trigger: "Draw architecture diagram", skill: "diagram-nex", alt: "visualization-nex (data charts)" }, { trigger: "Create a new Skill", skill: "create-skill-nex", alt: "—" }, { trigger: "Don't know which Skill to use", skill: "skill-orchestrator", alt: "—" } ] return (

{/* ── Hero ── */}

Engineering Guide · Context Engineering

Agentic Skills
Best Practices

A field guide for engineers on designing, deploying, and continuously improving production-grade AI agent skills — grounded in the CEVF framework, 52-skill catalog, and real failure-mode data.

52 Skills 9 Categories CEVF Framework 16 Failure Modes 200+ papers synthesized

{/* ── Highlights ── */}

{highlights.map((h, i) => (

{h.icon}

{h.title}

{h.desc}

))}

{/* ── Core Concept: What is a Skill? ── */}

What is a Skill?

A Skill is a structured workflow that an AI agent executes to complete a well-defined engineering task. It sits between raw LLM inference and production automation: it defines scope, phases, quality gates, and degradation paths.

Skills are the unit of context engineering — not prompts, not tools alone, but a reusable, testable, improvable contract between the engineer and the agent.

Decision Rule

Multi-step? No → Tool
Multi-step + no quality gate → Light Skill
Multi-step + quality gate → Full CEVF Skill

Entity Hierarchy

{[ { label: "Agent", sub: "Autonomy · Tool calls · Long-term memory", color: "bg-violet-100 border-violet-200" }, { label: "Skill", sub: "Scope · Phase flow · Verification gates · Degradation", color: "bg-blue-100 border-blue-200" }, { label: "Tool", sub: "Single API call · Read file · Search API", color: "bg-emerald-100 border-emerald-200" }, { label: "Sub-Agent", sub: "Limited scope · Depth-bounded · Parent-supervised", color: "bg-amber-100 border-amber-200" } ].map((e, i) => (

{e.label} {e.sub}

))}

{/* ── CEVF Layers ── */}

CEVF Framework

The four-layer contract that turns a Skill from "feels right" to "provably safe." Design order: C → V → E → F (gates first, then flow).

{cevfLayers.map((l) => (

{l.layer}

{l.name}

{l.tagline}

· {r}

))}

Skill Trust Score (STS)

{stsFormula}

{stsTiers.map((t) => (

{t.grade} {t.label} {t.range} {t.desc}

))}

V-layer weight (0.4) is highest — verification is the primary trust signal.

{/* ── Skill Catalog ── */}

Skill Catalog — 9 Categories

52 skills covering the full engineering lifecycle. Choose by intent, not by tool name.

{skillCategories.map((cat, i) => (

{cat.icon}

{cat.name}

{cat.skills.map((s, j) => ( {s} ))}

))}

{/* ── Top 10 ── */}

High-Frequency Top 10

{top10.map((s, i) => (

{String(i + 1).padStart(2, "0")}

{s.name}

{s.desc}

))}

{/* ── Decision Router ── */}

Quick Routing Table

Match your intent to the right Skill in under 10 seconds.

{decisionRouter.map((r, i) => ( ))}

Your Intent	Primary Skill	Alternative
{r.trigger}	{r.skill}	{r.alt}

{/* ── Creation Guide ── */}

7-Step Creation Guide

Total ~3.5 hours for a production-grade Skill. First time? Use fast mode (~1 hour): C-layer + 3 V-gates + 1 worked example only.

{creationSteps.map((s, i) => (

{s.step}

{s.title}

{s.time}

{s.desc}

))}

Three Mindset Shifts

{[ { from: "Feature-complete", to: "Minimum viable", note: "First Skill goal: works without big mistakes, not perfect" }, { from: "Writing a manual", to: "Writing a trigger", note: "SKILL.md's reader is the Agent, not a human. 'What to do when' > 'why'" }, { from: "Write once right", to: "Execute → Feedback → Iterate", note: "Step 7 review is the highest ROI action in Skill creation" } ].map((t, i) => (

{t.from} → {t.to}

{t.note}

))}

{/* ── Failure Modes ── */}

Failure Mode Dictionary

F1 + F2 + F11 cover ~60% of all Skill failures. Always check these three first.

{failureModes.map((f) => (

{f.id}

{f.name}

{f.symptom}

Fix: {f.fix}

))}

5 Fatal Anti-Patterns

{antiPatterns.map((ap) => (

{ap.id}

{ap.name}

{ap.problem}

✓ {ap.fix}

))}

{/* ── Minimum Checklist ── */}

Minimum Viable Skill Checklist

Before Shipping

☐ {item}

Production Deployment Guards

☐ {item}

{/* ── Context Engineering Layers ── */}

Context Engineering — 4-Layer Model

Skills live in Layer 1. Understanding all four layers explains where Skill knowledge flows and how it persists.

{[ { layer: "Layer 1", name: "Working Context", color: "bg-violet-50 border-violet-200", badge: "bg-violet-100 text-violet-700", desc: "Immediate input per call: System Prompt + Skill instructions + current task. Must be present every call." }, { layer: "Layer 2", name: "Session", color: "bg-blue-50 border-blue-200", badge: "bg-blue-100 text-blue-700", desc: "Persistent log of current conversation: history, intermediate results, decision chain. Can be compressed." }, { layer: "Layer 3", name: "Memory", color: "bg-emerald-50 border-emerald-200", badge: "bg-emerald-100 text-emerald-700", desc: "Long-term knowledge retrieved on demand: team conventions, historical decisions, lessons learned." }, { layer: "Layer 4", name: "Artifacts", color: "bg-amber-50 border-amber-200", badge: "bg-amber-100 text-amber-700", desc: "Large data accessed by reference only: code files, documents, datasets. Never loaded into context directly." } ].map((l, i) => (

{l.layer}

{l.name}

{l.desc}

))}

{/* ── Key Insight ── */}

"AI accelerates orientation and context."

Judgment, synthesis, and insight remain human responsibilities.

{[ { emoji: "🔺", title: "Highest ROI", body: "V-layer hard gates > trigger words > scope_out > worked examples" }, { emoji: "⚖️", title: "Impossible Triangle", body: "Precision × Generality × Simplicity — pick two. Compose via skill-orchestrator for the third." }, { emoji: "🔄", title: "Improvement Cadence", body: "Every edit: 2-min review. Monthly: reflect-skill. Quarterly: batch redesign." } ].map((c, i) => (

{c.emoji}

{c.title}

{c.body}

))}

{/* ── Footer ── */}

) }