loopkit vs other skill packs

Written in July 2026 after mining the top three loadout repos. This document is honest — where another pack beats loopkit, it says so. Use it to choose the tool for your loop, not to score points.

TL;DR

• loopkit — you want mini-skills + a working harness, no methodology commitment, drop-in.
• obra/superpowers — you want Prime Radiant's spec-first + TDD philosophy end-to-end, and you code in Claude Code.
• mattpocock/skills — you want Matt's specific engineering opinions (grill-with-docs, triage) and are OK adopting his flow.
• addyosmani/agent-skills — you want a 6-phase lifecycle (/spec → /plan → /build → /test → /review → /ship) with slash-command muscle memory.

They all share the same base: YAML-frontmatter skills that load on trigger. The differences are shape and floor.

Feature matrix

Feature	loopkit	superpowers	mattpocock	osmani
Skills load on trigger (SKILL.md YAML)	yes	yes	yes	yes
Ships a harness (.claude/settings.json, hooks, verifier subagent)	yes	partial	no	no
Loop runner (run.sh Plan→Act→Verify)	yes	no	no	no
Adversarial verifier subagent	yes	via TDD	via code-review	via /review
Methodology imposed	none	Prime Radiant	GSD-adjacent	6-phase lifecycle
Slash-command shortcuts	no	some	many	8 lifecycle cmds
Reference checklists (on-demand)	yes	no	some	yes (7)
Compatible with non-Claude agents	yes	Claude-first	multi	multi
Comparison doc against peers	yes (this)	no	no	yes
Skill count	33	~40	~19	~24
Reads as "personal loadout"	yes	no	yes	yes

Where each pack wins

loopkit

• Only pack that bundles the harness (settings.json + verifier subagent + loop runner) as part of the install. Everyone else assumes you already have a working .claude/.
• No methodology tax. You keep your workflow; skills specialize the agent, they don't own it.
• Adversarial-verify skill is unique — an explicit list of the 11 shortcuts agents take to fake "done".

obra/superpowers

• Deepest TDD enforcement in the ecosystem. Prime Radiant methodology is opinionated and coherent.
• The 10-agent integration story is real: plugs into far more than Claude Code out of the box.

mattpocock/skills

• grill-with-docs and grilling are unique — a genuine interviewing-your-domain interaction pattern nobody else has.
• Cleanest User-invoked vs Model-invoked separation. Sharpest skill hierarchy.
• Backed by newsletter + audience. If Matt updates it, it stays updated.

addyosmani/agent-skills

• Best lifecycle diagram. The /spec → /plan → /build → /test → /review → /ship mental model is memorable.
• 7 reference checklists (Definition of Done, Security, Perf, A11y, Observability, Testing Patterns, Orchestration) — the deepest supplementary layer.
• Anti-rationalization tables inside every skill. Concrete rebuttals to "I'll add tests later."

Where each pack loses

loopkit

• No slash-command shortcuts. If you want /build auto UX, look at Osmani.
• Solo maintainer. If continuity matters more than freshness, Anthropic's canonical skills are the safer bet.

obra/superpowers

• Prime Radiant is a commitment. If you don't want the whole methodology, the skills are less useful in isolation.
• Assumes you're using the full stack — the pieces are less drop-in than they look.

mattpocock/skills

• Deeply Matt-flavored. His triage state machine, his PRD synthesis. Great if you buy his taste, friction if you don't.
• Newsletter integration is genuine but adds a signup surface some teams won't want.

addyosmani/agent-skills

• Heaviest floor. The 6-phase lifecycle is powerful but adds ceremony to small changes.
• Slash-commands are agent-specific — reduces cross-agent portability.

Cross-pack composability

Skills are markdown files with YAML headers. Nothing stops you from mixing:

npx skills add Archive228/loopkit          # base + harness
npx skills add mattpocock/skills/grilling  # Matt's interview loop
npx skills add addyosmani/agent-skills/checklists  # reference layer

The trigger system means loading extras is nearly free — they only fire when their frontmatter matches. Mixing is the expected mode.