Skip to content

Choosing a skills approach

Observation date: October 27, 2026. See State of the Skills Market for the taxonomy this guide applies, and Choosing Your Approach for the technique-level decision guide this page complements.

Skills-related offerings serve different purposes. Identify the gap you need to address before choosing a library, routing mode, methodology, marketplace, or internal catalog.

What are you choosing?

Category Example Choose it when... Trade-off
Composable skill library mattpocock/skills You want a handful of independently selectable practices (TDD, code review, debugging) that you mix, edit, or ignore You retain full control, but nothing enforces sequence, coverage, or quality gates
Lifecycle library addyosmani/agent-skills You want skills organized around development stages (define → plan → build → verify → review → ship) without a mandatory gate between them Stage-mapped commands give more structure than a composable library, but each command can still be invoked out of order
Orchestrating mode pstack/poteto-mode You want a dispatcher that reads a request and automatically routes to the right skill or playbook in a library you already use Adds request-classification logic on top of a library — it has no function of its own if the underlying skills aren't installed
Skill-driven methodology Superpowers You want prescribed TDD, debugging, and review practices More process overhead and no team coordination; instruction compliance is not guaranteed
Native runtime Claude Code, Codex CLI, GitHub Copilot plugin marketplaces You want the simplest path: subscribe to a managed, auto-updating skill bundle inside the agent you already use Convenient, but you're trusting the publisher's next push over a reviewed, pinned copy
Custom internal catalog A team- or org-maintained internal skills repository You need house conventions, proprietary domain knowledge, or compliance-specific instructions that no public library provides You own all the maintenance, evaluation, and governance work a public library's maintainer would otherwise do

You can combine these options. Spec Kit can install skills through extensions; Squad agents can use Superpowers-style skills; orchestrating modes depend on a library; internal catalogs can adapt public skills. Start with the option that meets the immediate need.

Decision criteria

Check these criteria before adoption. See Distribution and Governance for details.

Criterion Question to ask Why it matters
Task shape Is expertise missing, or do you need a record of decisions? Skills supply reusable expertise; a methodology can add decision records and workflow structure.
Workflow ownership Do you want to select practices yourself or delegate sequencing? Composable libraries leave selection to you. Lifecycle libraries organize practices by stage; methodologies prescribe more of the sequence.
Update channel Does it auto-update through a native plugin marketplace, or do you pull updates manually via a CLI? Auto-update is convenient but hands you upstream changes unreviewed. A vendored copy is reviewable but only as current as your last update run.
Portability Does the runtime parse SKILL.md and execute the hooks, scripts, or context: fork features the skill needs? Successful installation does not establish support for optional runtime features.
Provenance Is there signing, attestation, or lockfile-style pinning — or just a repository URL and a star count? As of this observation date, nothing examined across the skills layer provides signing or attestation. Pin a revision if reproducibility matters.
Evaluation Was quality independently checked or reported by the maintainer? Treat "N/N runs passed" as a maintainer report, not third-party proof.
Dependency completeness Does a per-skill install pull everything that skill needs, or only the named directory? Partial installs can silently drop shared dependencies — a documented failure mode in at least one lifecycle library's issue tracker.
Compatibility Is the skill's current version the same across every agent and install path you use it from? The same library can be current in one runtime's marketplace and stale in a vendored copy, with no cross-channel version signal today.
Policy and control Who decided whether this auto-updates, and does that match your change-control policy? Auto-update-by-default versus manual-update-by-default is a policy choice — make it explicitly, especially in regulated or audited environments.
Team coordination Does the choice need to work the same way for every person on a team using different tools? A runtime-locked methodology (Squad, HVE are both Copilot-only) creates friction on a mixed-tool team; a portable specification or skills layer avoids it.
Risk Could unexpected behavior cause a bad suggestion, unwanted script execution, or a leaked secret? Skills can bundle scripts and tool permissions. Registry presence or an audit badge does not replace source review, permission limits, and pinning.

Practical guidance by need

If you want... Choose Trade-off
Project- or change-level artifacts driving implementation A specification framework — Spec Kit, OpenSpec, GSD, or BMAD Heavier upfront spec-writing investment; pick brownfield-first OpenSpec vs. project-level Spec Kit/GSD vs. full-team-simulation BMAD based on scope
Team-scale coordination or enterprise governance A multi-agent/runtime approach — Squad or HVE Copilot lock-in for both; skills become one ingredient inside a larger, heavier runtime
Automatic upstream updates, accepting less local review A native plugin marketplace subscription (Claude Code, Copilot, or Codex) Convenient, but upstream changes arrive without a local review step
Reviewable, reproducible files you control A vendored copy through skills.sh or the skills CLI, preferably pinned You own updates, conflicts, and local modifications
House conventions or proprietary domain knowledge no public library covers A custom internal catalog, optionally forked from a public library Full control and fit, but you own all maintenance, evaluation, and governance that a public maintainer would otherwise carry

Before combining skills with a methodology, check for duplicate instructions, conflicting workflows, unsupported runtime features, and unreviewed scripts.

The decision rule

Apply the selection rule from Choosing Your Approach:

  1. Start with the model and its native runtime for a bounded task with independent checks.
  2. Add skills for expertise the model does not reliably apply.
  3. Add a methodology for prescribed practices, decision records, or coordination.
  4. Strengthen controls as autonomy, duration, cost, or risk increases.