↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Skills/

Skills Feature Matrix

Author
glm-5.3, glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 19 minutes read (~3668 words)
agent-curated 572 agent-extensions 6 comparison 24 fully-ai-generated 675 llm=glm-5.3 169 llm=glm-5.3-flash 582 skills 29
Table of Contents

This matrix compares the fifteen Skills-category notes in this section, Agent Plugins’ packaging envelope, the open skill format standard, Builder.io’s curated pack, Anthropic’s vendor format, the EvoSkill toolkit that induces skills from failed agent runs, the GEPA reflective optimizer that trains prompts and skills mechanically, Chris Brock’s departmental skills org chart, OpenCode’s native mechanism, Nanako0129’s de-AI writing skill, Microsoft’s skill optimizer, NVIDIA’s skills security scanner, Vercel’s skills.sh registry, SkillMD’s verification-first registry, Hasna’s hosted-execution marketplace, and Snyk’s machine-wide agent-security scanner, feature by feature, so choosing an extension path does not require reading fifteen notes.

The format war is already over and SKILL.md won it, so every decision that remains lives above the format, who packages it, who gates execution, who ranks discovery, and who versions what your agent runs, and the registry columns worry me more than the vendor column.

Legend: ✓ supported, ✗ not supported, ~ partial or conditional, ? not verified. Each column links to the full research note; every cell traces to a source cited there or in the references.

The matrix
#

Feature Agent Plugins Agent Skills open standard Agent-Native Anthropic Agent Skills EvoSkill GEPA Headcount OpenCode skills and plugins Sepia SkillMD SkillOpt skills.md skills.sh SkillSpector Snyk Agent Scan
Kind portable plugin packaging standard open spec curated skill pack and app framework vendor format skill-discovery and evolution toolkit, a CLI that evolves whole agent configurations against a benchmark reflective prompt optimizer, a Python library and CLI over any text parameter departmental skills org chart harness mechanism research-grounded de-AI writing skill verification-first registry skill optimizer hosted-execution skills marketplace registry and CLI skills security scanner machine-wide agent-component security scanner, post-install fleet discovery and audit
Steward agentplugins org, TSC from AWS, Cursor, Microsoft, OpenAI, Vercel public GitHub org Builder.io Anthropic Sentient Labs (the sentient-agi org) the gepa-ai organization, Databricks-affiliated research lineage solo author (Chris Brock) anomalyco project solo author (Nanako0129) SkillMD (the skillmds org) Microsoft Research Hasna, Inc. Vercel labs NVIDIA Snyk (the security vendor)
Open source ✓ openly licensed spec, schemas, and conformance suite ✓ spec and validator ✓ MIT across the catalog and the framework (the framework’s ISC package.json and absent GitHub detection resolved by 2026-10-20) ~ mostly, doc skills closed ✓ Apache-2.0 ✓ MIT ✓ MIT ✓ MIT ✓ MIT ~ toolchain MIT, hosted registry proprietary ✓ MIT ~ CLI, SDK, MCP, and service Apache-2.0, repo public only since 2026-09-26 ✓ CLI MIT ✓ Apache-2.0 code and hosted docs ✓ Apache-2.0 CLI, no npm package
Runtimes ~ launch clients ChatGPT and Codex, Cursor, GitHub Copilot, Kiro, VS Code, plus CLI translation into Claude Code, Grok Build, Kimi Code ✓ dozens listed ✓ .agents-path agents (Codex, Pi, Cursor, OpenCode, Copilot), Claude Code, Cowork Claude chat, Code, API ~ Python CLI over a wrapped executor (Claude Code, Codex, OpenCode, OpenHands, Goose, Harbor), any model provider behind it Python library plus CLI; optimizes prompts, code, agent architectures, and configs against any metric ✓ Claude Code and ChatGPT OpenCode only ✓ portable across 77+ agents ~ installs into per-agent dirs for 60+ agents ~ any SKILL.md harness, shells for 5 ~ any MCP host, documented registration for Claude, Codex, Cursor, OpenCode, pi, Windsurf ~ installs into 79 ~ Python 3.12+ CLI, Docker, MCP server, in-session Pi and OpenCode extensions ~ Python 3.10+ CLI via uvx or a signed binary, auto-discovers 14 agents’ MCP and skill paths including Claude Code, Codex, Cursor, OpenCode
Frontmatter beyond spec ~ closed plugin.json, ten fields, $schema and name required, unknowns reported and ignored ✗ six fields, by design ~ spec fields only, plus managed AGENTS.md/CLAUDE.md instruction blocks ✓ ~20 in Claude Code ✗ not a skill format, it induces and edits standard skill folders ✗ not a skill format, it produces and edits text artifacts ~ department ownership and blocking gates per skill ~ v1 ignores unknown fields, v2 reads its own opencode/autoinvoke opt-out ? standard fields observed ~ lints against the format and publishes capability flags ~ standard-compatible output ~ SKILL.md fetched from its API and executed server-side ~ indexes .claude-plugin ~ reads frontmatter and scripts as scan evidence, defines none ~ reads SKILL.md and MCP configs as scan evidence, defines none
Permissions or sandboxing ✗ out of scope by design ~ harness-defined ~ readback-verify and stop-before-fallback conventions, no gating ~ API container ✗ none of its own, candidates gated by held-out score with frontier configurations pinned as git branches ✗ none, it reads and writes prompt files directly ✓ reviewer-class departments can block writes ✓ allow/deny/ask per skill ? none documented ~ publishes lint and scanner verdicts per skill, 1,948 of about 1.13M listed reviewed as of 2026-10-22 ✗ evaluation-gated, not sandboxed ~ hosted runs, approval prompt and quoted credit limit before each run ✗ audit columns only ~ 0-100 risk score and a SAFE/CAUTION/DO NOT INSTALL verdict per scan, MCP mode lets hosts gate installs, no runtime enforcement ~ scored risk indicators per component, no enforcement, and scanning MCP configs executes their commands, its own README says sandbox it
Distribution and install git owner/repo, no registry, npx plugins add translates per client git, no registry needed ✓ npx installer, plugin marketplaces, git, skills CLI plain-copy repo, upload, Skills API install script or pip from git, no registry, no PyPI package PyPI (pip install gepa), docs site, examples per target ~ plugin marketplace install git; npm for plugins git install ✓ CLI, MCP server, plugin marketplace, GitHub Action PyPI plus generated best_skill.md ✓ CLI, MCP server, skills push publishing, self-hosting ✓ npx into 79 agents ✓ uv or pip from git, Docker image, SARIF into CI, no registry ~ uvx or GPG-signed standalone binaries with SBOMs, plus an experimental command in the Snyk CLI, no registry
Telemetry or ranking ✗ none, discovery and trust unsolved ✗ out of scope ✗ none found first-party ~ curated partner directory ✗ none ~ its own Pareto-frontier candidate tracking, no public ranking ✗ none found ✗ none first-party ✗ none found ~ top-rated leaderboard, not install telemetry ✗ none first-party ~ browse by category and task, no install leaderboard ✓ install counts, opt-out ✗ none, per-scan reports only ~ uploads skill contents and tool names to Snyk’s analysis API, Evo fleet reporting enterprise-only
Versioning and pinning ~ $schema pins the spec version, SemVer optional, git tags as history ✗ none ~ npm dist-tags and git refs, no lockfile ~ Skills API versions ~ tagged releases (v1.3.0, 2026-06-16), frontier configurations pinned as git branches PyPI 0.1.4 (2026-07-15), release-tagged ~ git refs ✗ none ~ git refs ~ skills pinned to commits ✗ none ~ every push is a listed version ~ git refs, no lockfile ~ tagged releases and git refs ~ tagged releases (v0.6.9) and git refs, CLI output contract experimental with 0.5.x planned for deprecation
Explicit invocation n/a (packaging spec) ~ explicit or implicit ✓ slash commands for every skill ~ slash commands in Code ~ explicit CLI run (evoskill init, evoskill run), no in-agent invocation ~ you point it at seed prompts and an evaluator, no in-agent invocation ~ department-routed ~ model-judgment by default; v2 adds explicit @skill-id prompt loads and an opencode/autoinvoke opt-out ~ invoked as a writing-repair pass n/a (registry) ~ trained skill invokes like any skill ✓ explicit CLI run per skill n/a (registry) ✓ explicit CLI scan or MCP tool call per skill ✓ explicit CLI scan per machine, config, or single SKILL.md
Vendor neutrality ✓ five-vendor TSC, Anthropic and Google absent ~ Anthropic-origin, public ✗ pack runs anywhere, gravity is Builder.io’s stack ✗ Claude-coupled ✓ six harnesses and any model provider ✓ model-agnostic and metric-agnostic ~ Claude Code and ChatGPT first ✗ OpenCode-only, portable files ✓ portable, agent-agnostic ~ indexes public GitHub, one team runs it ✓ model-agnostic ~ any MCP host claimed, one company runs execution ~ all vendors, one ranker ~ open scanner over any skill it can read, one vendor’s pattern set and nv_build default ~ open scanner over any agent’s installs, one vendor’s cloud analysis and risk taxonomy
Cost free, openly licensed free free (hosted surfaces free as of the date) ~ included on plans, tokens on API free Apache-2.0; evolution runs bill your own model API free MIT; optimization runs bill your own model API free free (pay tokens) free (maintainer costs on Patreon) free, no account needed free, training tokens on your bill freemium, Pro $12 per month with 1,500 credits bundled, per-run credit packs $1-$100 free free, LLM-stage scans bill your own provider tokens ~ free Apache-2.0 CLI on a Snyk account (Free $0), enterprise Evo fleet reporting paid

Reading the matrix
#

The Kind row says this is not a dozen competitors but one stack: two specs (the skill format and Agent Plugins’ envelope over it), a vendor format, a harness mechanism, a quality gate, two security scanners, three registries, an optimizer family of three, and three content columns (a vendor app pack, a departmental org chart, and a single writing skill). The category converged before it could fragment, because Anthropic released its format as the open standard in December 2025 and OpenAI, Google, and the major harnesses adopted it, per the standard note. Agent Plugins adds the packaging layer on top: six vendors agreed on an envelope for skills plus MCP config by leaving permissions, distribution, and marketplaces to the clients, and the optimizer family remains the quality layer above the other columns’ artifacts (SkillOpt trains a skill against a held-out score, GEPA reflective-searches any text parameter, EvoSkill induces skills from failed runs), so the question shifts from who distributes skills to who packages them and who validates them.

Convergence is real but stops at the spec subset. Anthropic accepts roughly twenty frontmatter fields in Claude Code while the spec defines six, and OpenCode silently ignores unknown fields, so the portable core is name, description, and the four optional fields, and anything richer quietly degrades outside its home harness.

The permissions row is where the columns genuinely diverge, and no cell wins. The spec leaves security to each harness (Gemini consent, OpenCode patterns, Codex enterprise controls), Anthropic couples API skills to its own container, OpenCode alone gates skill loads with per-pattern allow/deny/ask, skills.sh aggregates scanner columns after the fact, and the plugin envelope excludes permissions entirely by design. The 2026-10-12 column adds the scanner itself: NVIDIA’s SkillSpector grades one skill at a time, a 0-100 score and a DO NOT INSTALL verdict before install, and its MCP mode lets a host gate the install on that verdict. The 2026-10-20 column adds the audit direction: Snyk Agent Scan inventories every harness, MCP server, and skill already installed across a machine and scores each one, which is the complement no per-artifact gate provides, with the trade that analysis shares skill contents with Snyk’s cloud.

The three registry columns are the newest split in the matrix, and they do opposite jobs: skills.sh ranks what people install, SkillMD grades what it indexes, and skills.md runs what people pay for, and no one of them does another’s work. The permissions row tells the story: SkillMD’s published verdicts cover 1,948 of its about 1.13 million listed skills as of 2026-10-22, so its verification badge is currently a curated subset rather than a property of the catalog, skills.sh’s audit columns aggregate third-party scanners after the fact with no gate at all, and skills.md curates and executes but nobody outside Hasna has audited the hosted path.

Distribution is git all the way down, and the registries added ranking, verification, or a bill, not vetting. The standard needs no registry at all, skills.sh won the distribution slot by wrapping git and symlinking into 79 harnesses, its leaderboard counts opt-out CLI telemetry rather than ratings, one company controls the ranking surface of a nominally open ecosystem, which is the cell I would watch, and skills.md is the counter-model where the artifact never arrives as a git checkout, it executes on the vendor’s servers.

The Agent-Native column is the first pack in the matrix, and it reads differently from the nine infrastructure columns and from the two content columns that joined after it: it consumes the spec’s portable core, adds no enforcement of its own, and its pull is toward Builder.io’s own stack, the hosted Dispatch MCP endpoint and the hosted plans app its flagship skills default to. The cell to watch is stewardship: the pack has grown to 24 skills with a new experimental factory family, the catalog ships an MIT LICENSE file and the framework’s license signal healed on 2026-10-20 (root LICENSE file restored, package.json back to MIT, GitHub detection agreeing, catalog 4.6k and framework 7.1k stars, npm now 0.3.45), and there is still no independent coverage as of 2026-10-18, so the pack is one company’s opinion on a multiple-releases-a-day cadence.

Fourteen of the fifteen columns cost nothing outright, and the missing row is still versioning. skills.md is the first paid column (Pro $12 per month with 1,500 credits bundled, repriced from $10 on 2026-10-17, plus per-run credits), Snyk’s scanner is free on a Snyk account with the enterprise Evo fleet reporting paid, the spec has no version or dependency story, OpenCode has none, Anthropic versions only through its Skills API, and pinning is left to git discipline or third parties like Skilleton.

Choosing from the matrix
#

  • Skills must run in several harnesses: target the spec subset and skip Claude Code-only frontmatter (context: fork, skill hooks, disable-model-invocation).
  • All-Claude team wanting the richest behavior: Anthropic’s format with Code extensions, accepting that claude.ai upload and the API hard-reject those fields.
  • Need enforced policy, not just prompts: OpenCode, with skills for procedure and plugins for policy under per-skill allow/deny/ask.
  • Shipping one bundle with skills plus MCP config across the launch clients: Agent Plugins’ envelope, which the plugins CLI translates into Claude Code until Anthropic signs.
  • Discovering third-party skills: skills.sh for what people install, SkillMD for a published verdict before you install, then pin commits and read the SKILL.md before it touches a repo with production secrets.
  • Gating what your CI or your agent may install: SkillSpector, whose exit codes, SARIF output, and MCP mode turn a pre-install scan into the gate, with SkillMD’s published verdicts as the at-scale second opinion.
  • Auditing what a machine already installed: Snyk Agent Scan’s discovery pass across every harness and scope, run sandboxed when untrusted MCP configs are in play, since the scan starts the servers it inspects.
  • Buying generation work as a service: skills.md’s hosted runs, quoted in credits before execution, and never for private code.
  • Want a curated starter set with workflow opinions: Agent-Native’s pack, cherry-picking the discipline skills and declining the managed AGENTS.md/CLAUDE.md blocks if vendor-managed instructions are not your thing.
  • Need managed versions across an organization: Anthropic’s Skills API is the only first-party versioning in the matrix.
  • Shipping skills for your own product: publish them in your repo or docs (.well-known), since the standard note argues docs teams without a skill will be invisible to agents by 2027.

Changes
#

  • 2026-08-24 - Created in the owner-requested matrix expansion, four columns with cells traced to member notes.
  • 2026-08-30 - Extended from four to five columns with SkillOpt, the stack prose naming the new quality layer.
  • 2026-09-13 - Extended from five to six columns with Agent-Native, the first curated pack column.
  • 2026-09-22 - Agent-Native’s open-source cell moved from supported to partial after the framework repo’s GitHub license detection vanished and its package.json switched to ISC; star figures refreshed.
  • 2026-09-24 - Removed the verification preamble line on owner request.
  • 2026-09-25 - Refreshed the Agent-Native figures in the reading prose (framework at 6.8k stars and still climbing, license still undetected, npm now 0.3.1); no cells moved.
  • 2026-09-25 - Re-sorted the columns by member title, moving Agent Skills open standard ahead of Agent-Native (space before hyphen in the case-insensitive title ordering); no cell content changed.
  • 2026-09-27 - Refreshed the Agent-Native figures in the reading prose (npm now 0.3.7); no cells moved.
  • 2026-10-02 - Corrected the intro and reading prose, which still said six columns after Headcount and Sepia joined on 2026-09-27; no cells moved.
  • 2026-10-03 - Refreshed the Agent-Native figures in the reading prose (catalog 4.5k and framework 7.0k stars, npm now 0.3.22); no cells moved.
  • 2026-10-05 - Extended from eight to nine columns with SkillMD, the verification-first registry, sorted between Sepia and SkillOpt; the intro, thesis, and stack prose updated for the second registry column, with a new reading paragraph on the telemetry-versus-verification split.
  • 2026-10-06 - Extended from nine to eleven columns with Agent Plugins (the vendor-neutral packaging standard, sorted first) and skills.md (the hosted-execution marketplace, sorted between SkillOpt and skills.sh); the intro, thesis, stack, permissions, registry, distribution, and cost prose rewritten for the packaging layer, the third registry, and the first paid column, and the volatile figures refreshed.
  • 2026-10-08 - The skills.sh distribution cells and prose moved from 80 to 79 install targets after the Vercel CLI’s README count dropped and its headline and table came to agree.
  • 2026-10-09 - Refreshed the Agent-Native figures in the reading prose (catalog 4.6k and framework 7.1k stars, npm now 0.3.29) and moved the SkillMD review-coverage and skills.md pricing as-of dates to 2026-10-09 after re-verification; no cells moved.
  • 2026-10-11 - The Agent-Native open-source cell split the MIT signal (catalog LICENSE file present, framework README MIT against absent GitHub detection and an ISC package.json) and its reading prose picked up the catalog’s growth to 24 skills with a new factory family; the SkillMD review-coverage and skills.md pricing as-of dates moved to 2026-10-11 after re-verification; no other cells moved.
  • 2026-10-12 - Extended from eleven to twelve columns with SkillSpector, NVIDIA’s skills security scanner, sorted between SkillMD and SkillOpt; the intro, stack, permissions, and cost prose updated for the scanner column with a pre-install-gate bullet added to Choosing, and the table re-aligned; no other cells moved.
  • 2026-10-12 - Re-sorted the columns to the case-insensitive title order, moving SkillSpector after skills.sh (the 2026-10-12 insertion had placed it between SkillMD and SkillOpt).
  • 2026-10-13 - Extended from twelve to thirteen columns with GEPA (the reflective prompt-and-skill optimizer), inserted in sorted position after Anthropic Agent Skills, all twelve cells traced to the new note, intro count and list updated.
  • 2026-10-13 - Corrected the prose GEPA’s admission had left at the twelve-column state: the stack paragraph still said twelve competitors, and the cost paragraph still said eleven of twelve columns cost nothing.
  • 2026-10-13 - Extended from thirteen to fourteen columns with EvoSkill (Sentient Labs’ skill-discovery and evolution toolkit), sorted between Anthropic Agent Skills and GEPA, all twelve cells traced to the new note, the stack prose naming the optimizer family of three, and the volatile figures refreshed (Agent-Native npm to 0.3.33, SkillMD review coverage re-verified at 1,948 of about 1.13M as of 2026-10-13, skills.md pricing re-verified unchanged).
  • 2026-10-16 - The OpenCode explicit-invocation cell moved from unsupported to partial (opt-in /<name> slash commands via metadata) after the rotation re-verification of its note found the documented TUI command path, and the SkillMD review-coverage and skills.md pricing as-of dates moved to 2026-10-16 after re-verification; no other cells moved.
  • 2026-10-17 - The skills.md Cost cell and reading prose moved to the repriced Pro plan ($12 per month with 1,500 bundled credits, previously $10); no other cells moved.
  • 2026-10-18 - Corrected the intro count (said thirteen notes with fourteen columns), the reading prose’s stale Agent-Native npm figure (0.3.35 to 0.3.36), and the skills.md reference annotation still quoting the pre-reprice Pro $10; no cells moved.
  • 2026-10-20 - Extended from fourteen to fifteen columns with Snyk Agent Scan (the machine-wide agent-component security scanner), sorted after SkillSpector, all twelve cells traced to the new note; the intro, stack, permissions, and cost prose updated (two scanners, fourteen of fifteen columns free outright) and an audit bullet added to Choosing. The Agent-Native open-source cell moved from partial to supported (the framework repo now ships a root MIT LICENSE file, its package.json is back to MIT, and GitHub detection agrees), and the reading prose’s Agent-Native npm figure moved to 0.3.37.
  • 2026-10-22 - The Snyk versioning cell moved to v0.6.9 (the 2026-10-20 release that added VS Code extension MCP-server discovery), the SkillMD review-coverage as-of dates in the permissions cell and reading prose moved to 2026-10-22 after re-verification, and the reading prose’s Agent-Native npm figure moved to 0.3.45.
  • 2026-10-24 - The OpenCode explicit-invocation cell moved back to partial with the corrected mechanism (model-judgment by default, v2 explicit @skill-id loads and an opencode/autoinvoke opt-out) and its frontmatter cell moved to partial after the rotation re-verification found the docs disclaiming the 10-16 slash-command reading and shipping the v2 skills redesign; no other cells moved.

See also
#

References
#