Skill Level Sources
External sources grounding the L1–L7 skill maturity framework defined in writing-skills.md. The framework is not our invention — it adapts a community framing, is validated against Anthropic's official guidance, and echoes the shape of long-established human-competence and process-maturity models. This page records where each idea comes from, what each source is good for, and — just as important — which frequently-suggested sources we assessed and deliberately excluded.
All links verified 2026-07-26.
1. Direct ancestry — agent skills and their levels
The sources the L1–L7 framework is actually built from.
| Source | What it is | Grounds |
|---|---|---|
| Simon Scrapes, "Every Level of Claude Code Skills in 27 mins" (YouTube, 19 Mar 2026) | The direct ancestor of our 7-level framing | The L1–L7 ladder itself |
| Anthropic, "Skill authoring best practices" | Official authoring guidance: conciseness, degrees of freedom, descriptions/triggering, progressive disclosure, "build evaluations first" | L1, L2, L3, L5 |
| Agent Skills open standard — published by Anthropic (18 Dec 2025), aligned with the Agentic AI Foundation (Linux Foundation; long-term governance not yet formally settled); adopted by Microsoft, OpenAI, Atlassian, Figma, Cursor, GitHub | The specification of what a skill is (SKILL.md format, frontmatter, progressive disclosure tiers, skills-ref validate) | L1 |
| Barry Zhang & Mahesh Murag (Anthropic), "Don't Build Agents, Build Skills Instead" (AI Engineer, late 2025) | The thesis: agents lack expertise; skills are composable procedural knowledge loaded at runtime | The overall premise |
| Anthropic, "Improving skill-creator: Test, measure, and refine Agent Skills" | skill-creator's Create/Eval/Improve/Benchmark modes; with-skill vs without-skill A/B benchmarking; pass-rate/latency/token metrics | L5 (see skill-evals.md) |
| MindStudio, "How to Build a Learnings Loop for Claude Code Skills That Self-Improve" (19 Mar 2026) | The persistent learnings.md feedback mechanism | L6 |
| Community implementations: claude-reflect-system, claude-meta, turbo, self-learning-skills | Working learnings-loop variants (correction capture, confidence tracking, golden-path persistence) | L6 |
Key fact: there is no official Anthropic concept of "skill levels" — neither the Claude Code docs nor the Agent Skills standard define maturity tiers. Anthropic provides the machinery (eval tooling, hooks, dynamic context injection); the leveling is a community/Conduction construct.
2. Human skill-level frameworks — the ladder's ancestry
Established frameworks that grade people. They legitimize the "numbered levels of skill" shape and give us shared vocabulary with clients, but none of them grades an artifact the way L1–L7 does.
| Source | Levels | Mapping onto L1–L7 |
|---|---|---|
| SFIA 9 (SFIA Foundation, Oct 2024) — Skills Framework for the Information Age | 7 levels of responsibility: 1 Follow → 7 Set strategy | Same ladder shape and count, but measures human autonomy/influence. SFIA 9 explicitly covers AI skills and even offers guidance on using its levels to decide which tasks to assign to AI |
| Dreyfus model of skill acquisition — Dreyfus & Dreyfus, UC Berkeley ORC-80-2 (1980) | 5 stages, canonically Novice → Advanced Beginner → Competent → Proficient → Expert (the 1980 paper uses novice → competence → proficiency → expertise → mastery) | Rule-following → intuition in a learner ≈ L1 (follow anatomy) → L6 (self-improvement), for human cognition |
| European e-Competence Framework (e-CF), EN 16234-1:2019 (CEN TC 428) | 41 competences × 5 proficiency levels e-1…e-5, aligned to EQF 3–8 | Closest human analogue to "one skill, graded levels" — a proficiency-per-competence matrix |
| KWIV — Kwaliteitsraamwerk Informatievoorziening 4.0 (Rijksoverheid, Dec 2023) | ~61 kwaliteitenprofielen for government IV/IT roles, explicitly built on e-CF | The Dutch-government hook: the Rijksoverheid already runs its IT workforce on e-CF levels — a skill ladder that name-checks e-CF/KWIV speaks our clients' language |
| Bloom's taxonomy (revised) — Anderson & Krathwohl (2001), A Taxonomy for Learning, Teaching, and Assessing | 6 cognitive levels: Remember → … → Evaluate → Create, plus a knowledge dimension | Evaluate → Create maps onto L5 → L6/L7; "metacognitive knowledge" is a good frame for the L6 learnings loop |
3. Maturity models — levels applied to capability, not people
The conceptually closest ancestry: maturity of a process or capability, which is what L1–L7 grades for a skill artifact + its improvement process.
| Source | Levels | Mapping onto L1–L7 |
|---|---|---|
| CMMI V3.0 (ISACA, Apr 2023) | 1 Initial → 2 Managed → 3 Defined → 4 Quantitatively Managed → 5 Optimizing | The canonical "maturity level" ancestry. Elegant parallel: CMMI L4 (quantitative measurement) ≈ our L5 (evals); CMMI L5 (optimizing) ≈ our L6 (learnings loops). Our L7 extends beyond CMMI's scope |
| Microsoft, Agentic AI adoption maturity model | Level 100 (Initial) → agent-first optimized | Grades an organisation's adoption, not a skill artifact — a complementary axis |
| Gartner, "Agentic AI Maturity Roadmap" (Aug 2025; paywalled, reseller summary) | 5 levels, task automation → expert autonomous systems | Secondary citation only (paywalled); same organisational axis as Microsoft's |
4. Measurement methodology — what L5 borrows from
Model benchmarks do not measure skill maturity (they grade models), but their methodology is exactly what skill evals borrow. Cite these for the technique, not the leaderboard.
| Source | Technique we borrow |
|---|---|
| SWE-bench (Jimenez et al., ICLR 2024) + SWE-bench Verified | Task-based, end-to-end verifiable agentic evaluation — the template shape for skill evals |
| τ-bench (Yao et al., Sierra) + τ²-bench | Tool-agent-user interaction evals; the pass^k reliability metric (run the same case k times — consistency is part of quality) |
| Terminal-Bench (Stanford + Laude Institute) | Containerized task + programmatic verifier suite per case |
| Berkeley Function-Calling Leaderboard (UC Berkeley Gorilla) | Tool-call correctness grading |
| LiveBench (Abacus.AI + NYU et al.) | Contamination-free rolling refresh — keep eval cases fresh so skills don't overfit to them |
| MathArena (ETH Zurich SRI Lab + INSAIT) | Evaluate on genuinely unseen problems (fresh competitions) with objective ground truth |
| LMArena / Chatbot Arena (ex-LMSYS) | Pairwise-preference + Elo/Bradley-Terry for comparing two skill variants when no objective ground truth exists. Caveats: now commercial; known critiques (vote gaming, style bias, "Leaderboard Illusion") |
| Zheng et al. 2023, "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena" (NeurIPS 2023) | LLM-as-judge — and its documented biases (position, verbosity, self-preference). Required reading before grading skill output with a model |
| OpenAI Evals | Open-source eval framework/registry; useful reference implementation |
| Prompt Engineering Guide (DAIR.AI) + Anthropic Cookbook | Authoring patterns (L3) and practical eval notebooks incl. LLM-as-judge caveats (L5) |
5. Assessed and excluded — do not re-add without new evidence
These come up in every "top LLM resources" list (a 20-item list of them was vetted on 2026-07-26). They were excluded deliberately:
Category error — human upskilling, not agent-skill maturity. Hugging Face LLM Course, DataCamp, DeepLearning.AI courses, Karpathy's videos/nanoGPT, Cohere LLM University, Sebastian Raschka's books, mlabonne/llm-course, KDnuggets roadmaps, Awesome-LLM link dumps, Towards Data Science / Ahead of AI newsletters. All fine for humans learning about LLMs; none says anything about the maturity of a skill artifact. (If we ever write a personal-development learning path, that's a different document.)
Category error — model capability, not skill maturity. Artificial Analysis Intelligence Index, GPQA Diamond, MMLU-Pro, Humanity's Last Exam. They rank models; a skill's maturity is independent of which model runs it.
Dead. Hugging Face Open LLM Leaderboard — officially retired March 2025 (v1 and v2), archive-only. Useful only as a historical lesson in benchmark saturation.
Unverifiable provenance. benchmarklist.com and llm-benchs.com — anonymous aggregators with no identifiable operator, no methodology page, no independent coverage. Opacity is the disqualifier.
Single-firm content marketing. benchmarkingagents.com (Digital Signet) — decent editorial roundup, but it aggregates the primary sources cited in §4; cite those directly. ai-radar.aoe.com (AOE GmbH) — an honest ThoughtWorks-style tech radar for AI tooling, but one agency's opinion; at most a tooling-discovery pointer.
Attribution traps (errors found in circulating lists — don't reproduce them): LiveBench is Abacus.AI + NYU et al., not ETH Zurich (MathArena is the ETH one — they are unrelated projects); the Agent Skills standard is published by Anthropic and aligned with the Agentic AI Foundation, which does not (yet) formally own it.
6. User AI-fluency frameworks — the other ladder
Skill maturity (L1–L7) grades the artifact. A separate, adjacent question is how proficient the human using the AI is. These frameworks grade users/practitioners, and matter for product decisions (progressive UI, guided modes, tutorials) in AI products like hermiq. Vetted 2026-07-26.
The five citable ones, ranked:
| Source | Shape | Why cite it |
|---|---|---|
| Anthropic AI Fluency: Framework & Foundations — Anthropic Academy course co-developed with Rick Dakan (Ringling College) & Joseph Feller (University College Cork); CC BY-NC-SA; also on Coursera and aifluencyframework.org | The 4Ds: Delegation, Description, Discernment, Diligence — competency dimensions, not levels | The anchor. For a Claude-based product, Anthropic's own academically co-authored framework is the coherent choice; use the 4Ds as the dimensions each level is measured on |
| EU AI Act Article 4 — AI literacy | Legal obligation, applies since 2 Feb 2025; national enforcement + penalties from 3 Aug 2026. Providers and deployers must ensure staff AI literacy calibrated to role, knowledge, and context; documentation expected (EU AI Office Living Repository) | The compliance driver, not a framework. Proficiency-aware features (guided modes, tracked tutorial completion, role-calibrated UI) map directly onto a deployer's Article 4 documentation duty — acutely relevant to Dutch government clients |
| UNESCO AI competency frameworks (Sept 2024) | Teachers: 15 competencies × 5 dimensions × 3 progression levels (acquire/deepen/create); students: 12 competencies × 4 dimensions × 3 levels (understand/apply/create) | UN-level authority for the shape: competencies × 3 progression levels is exactly the pattern for progressive UI tiers |
| OECD/EC AILit Framework (final June 2026, with Code.org; feeds PISA 2029) | 4 domains: Engage with AI, Create with AI, Manage AI, Shape AI | EU-institutional reinforcement; the domains map well onto consumer → builder → administrator → governance personas. Caveat: scoped to primary/secondary education |
| Microsoft — Agentic AI adoption maturity model (Levels 100–500, organisational) + Work Trend Index user segmentation (skeptics → novices → explorers → power users) | Org maturity ladder + the only large-N empirical user segmentation | Concrete enterprise example; vendor-research caveat applies |
Honorable mention: Alex Ewerlöf, "AI Fluency Leveling" (Jan 2026) — 7 levels, Casual Consumer → Prompt Coder → Context Developer → AI Engineer → AI System Architect → AI Platformizer → AI Pioneer. The best practitioner-authored consumer→builder ladder; cite as "one practitioner formulation", not authority.
The common shape across every credible framework: 3–7 levels; the same recurring dimensions (prompting/description sophistication, delegation judgment, discernment/oversight, building/orchestration, governance/responsibility); and the universal arc consume → direct → verify → compose → govern. A 4–5 level user model scored against the Anthropic 4Ds, with UNESCO/OECD's understand → apply → create progression as the leveling logic and Article 4 as the compliance rationale, sits squarely inside all of them.
Assessed and excluded from this category (vetted 2026-07-26): Larridin's AI Proficiency Model (real a16z-backed startup, but SEO funnel content for their SaaS); "5 Layers of AI Engineering" (aibuilderclub.com — a system architecture layering, not human proficiency; paid-community marketing); "Levels of AI Engineering Proficiency" LinkedIn posts (anonymous-grade, no citations); the "12-level Prompt Engineering Mastery" progression (could not be located at all — treat as apocryphal); aimlinsights.com and Pertama Partners (content-farm / regional vendor marketing); KDnuggets roadmaps, StrataScratch, To Data & Beyond, outcomeschool (legitimate developer curricula or monetized roadmaps — wrong axis for user proficiency; the same goes for mlabonne/llm-course and the LLM Engineer's Handbook, both excellent for engineer upskilling). The InformationWeek "101–401" agentic learning model is a real op-ed by Dan Mitchell (Digital Wave Technology, Jul 2026) — citable as illustrative pedagogy only, attributed to Mitchell.
7. Agent-platform maturity & autonomy (mid-2026 wave)
A third axis next to skill maturity (§1–5) and user fluency (§6): how mature the platform/organisation running agents is, and how much autonomy an agent is granted. Vetted 2026-07-27; four sources survive.
| Source | Shape | What we take from it |
|---|---|---|
| Pydantic, "A maturity model for applied generative AI" (Alex Che, 25 Jun 2026) | 5 levels (Experimenting → Adopted → Observed → Governed → Optimized) × 5 dimensions: visibility/observability, evaluation & quality, cost governance, access & identity, audit & incident response | The eval-coverage ratchet (ad-hoc → suites for critical agents → evals as CI gates → continuous online evals) and cost ladder (dashboards → alerts → server-side caps → per-run unit economics). Advice: fix "the gap currently bleeding", don't chase L5 everywhere |
| Addy Osmani, "Agentic Autonomy Levels" (2 Jul 2026) | 6 levels L0–L5: Assist → Supervised Action → Scoped Task Delegation → Goal-Driven Autonomy → Parallel Delegation → Managed-by-Exception Orchestration | Graduated per-agent autonomy; the per-run contract (goal, scope, non-goals, tool permissions, stopping condition, evidence, escalation, budget); risk-calibrated autonomy (risk × reversibility × available verification); KPIs: mean time between interventions, longest unattended run, auto-approval rate, token cost per accepted change |
| Arize, "How to write effective AI agent skills: 6 data-backed practices" (Laurie Voss, 24 Jul 2026) | Six practices backed by SkillsBench data | Human-curated beats self-generated (+18–25 pts vs −8–12); compact & procedural wins (comprehensive docs +0.7 pt only); route a minimal skill set (1–3 skills ≈ +18–19 pts; all-196 loaded = worse and +23% tokens); test per model×harness ("portable as files, not as behavior"); target gaps the base model can't fill; every skill change is a paired experiment against a frozen baseline. "The eval is the skill" |
| CMMI AI Maturity (CMMI AIM, ISACA, launched Jun 2026) | AI content across all 31 CMMI Practice Areas, 8 domains (Data, Development, People, Safety, Security, Services, Suppliers, Virtual collaboration) | Institutional weight for AI-capability maturity; People domain = workforce-readiness as a first-class axis. Caveat: model body is behind CMMI licensing — cite only the public structure, under the correct name "CMMI AIM" |
Assessed and excluded (2026-07-27): the widely-shared "Hakkoda AI maturity ladder (LLM velocity → build skills → chain skills → agents → agents-talk-to-agents)" could not be located on hakkoda.io as cited — their real publications say different things; treat the ladder as misattributed and back agent-to-agent rungs with Osmani L4–L5 instead. AgentPatterns.ai's 7-phase adoption model is thoughtful but authorless; HiBob's "AI Skills Framework" is a press release; The Neuron's 5-level stack (Projects → Prompting → Skills → Automations → Agents) is consumer up-skilling content — citable only as a footnote that even consumer guidance converges on skills → automations → agents; AgenticCareers' career ladder is job-board content marketing.
Convergent spine across the four citable sources: autonomy is granted per-run against verification capability (Osmani); quality is enforced by evals-as-gates (Pydantic + Arize); cost, access, and audit are server-side controls, not dashboards (Pydantic + CMMI).
How the levels map to their strongest sources
| Level | Strongest external grounding |
|---|---|
| L1 Anatomy | Agent Skills open standard; Anthropic best practices |
| L2 Triggering | Anthropic best practices (descriptions, progressive disclosure) |
| L3 Patterns | Anthropic best practices; Prompt Engineering Guide; Anthropic Cookbook |
| L4 Personalization | (inherently internal — no external source can supply your business context; e-CF's "competence in context" framing is the analogue) |
| L5 Measurement | Anthropic skill-creator blog; SWE-bench / τ-bench / Terminal-Bench / BFCL methodology; Zheng 2023 for LLM-as-judge caveats |
| L6 Self-Improvement | MindStudio learnings loop; community implementations; CMMI L5 "Optimizing" as ancestry |
| L7 AI Workforce | Zhang & Murag talk; Claude Code Agent Teams / Agent SDK docs; Microsoft agentic maturity model as the organisational mirror |