Eighteen open Claude skills are worth a serious look, and plenty of them quietly solve the problem sitting right next to them. Installing all eighteen isn't a stack. It's a pile.
So: four categories, a map of where each one's skills genuinely overlap, and three honest drafts of the stack that survives. The right cut depends on what you actually do, not on which repo has more stars.
How to read it: pick Draft A / B / C in any section. Survivors light up, cuts grey out, and the synergy stack rebuilds beside them.
Every skill here wants the same thing: an agent that behaves like a senior engineer. They differ in weight — how much process they impose — and in what state they keep on disk.
Two overlap clusters, so two decisions: methodology and memory.
- 1 Methodology clusterWhich methodology — and how much process do you want imposed?
- 2 Memory clusterOne memory layer, or two?
- Methodology cluster — One goal at three weights: heavy framework, mid command-set, one-file nudge. Pick a weight.
- Memory cluster — planning-with-files stores task state; memU stores knowledge. Whether that's a dedup at all is Draft C's argument.
- No real overlap — Vercel's agent-skills complements any methodology; understand-anything is comprehension, not workflow. Both survive every draft.
Cards show the current draft's verdict. Switch drafts below to re-cut them.
Superpowers
KeptFull SDLC methodology: brainstorm → spec → plan → subagent TDD → review → merge. Runs for hours unattended.
agent-skills (Addy Osmani)
Cut24 skills behind 8 slash commands (/spec, /plan, /build, /review, /ship…). Invoked deliberately, not automatically.
andrej-karpathy-skills
CutOne CLAUDE.md, four principles, aimed at Karpathy's list of LLM coding failure modes. Zero overhead.
agent-skills (Vercel Labs)
KeptFramework-specific quality: cost/perf audit, 40+ React rules, 100+ a11y/UX rules.
planning-with-files
KeptTask plan, findings, and progress kept on disk. Survives /clear, context loss, and crashes.
memU
CutFile-based memory of facts, preferences, and learned patterns across sessions. Still under heavy construction.
understand-anything
KeptTurns a codebase or docs set into a searchable knowledge graph. An onboarding tool.
- Superpowersspine
- planning-with-filesdurable state
- Vercel agent-skillsReact/Next quality gate
- understand-anythingonboarding new repos
One methodology, one memory, two specialists.
-
Draft A Maximalist spine recommended
If an agent gets hours of autonomy, one framework that owns the lifecycle beats stitching small skills together. Superpowers already contains brainstorming, planning, TDD and review, so Addy's commands and the karpathy nudge fold into it. planning-with-files stays as the crash-recovery layer; memU is dropped as unstable and redundant.
Cut: Addy's agent-skills, karpathy-skills, memU
-
Draft B Lean & composable
Superpowers is heavy and opinionated — a lot to trust unattended. Lighter and fully under your control: karpathy-skills sets behavior, Addy's agent-skills supplies the workflow. One is a personality, the other a process, so they aren't redundant. Memory stays minimal.
Cut: Superpowers, memU
-
Draft C Keep both memories
The memory dedup isn't one. planning-with-files answers “how far am I?”; memU answers “what do I know about this user and workspace?” Cutting either loses an axis. Keep both, pair with whichever methodology you picked — but treat memU as experimental, not load-bearing.
Cut: nothing on the memory axis (methodology still resolved via A or B)
Draft A for heavy autonomous work, Draft B to keep your hands on the wheel. On memory, side with A/B for now: planning-with-files stays, memU waits until it stabilizes. Vercel's agent-skills and understand-anything survive every draft.
These all exist because AI interfaces have a smell — purple gradients, dead-center everything, no motion. The question is whether any two fix the same smell.
Two decisions: which generator, and whether the deck/video pair is redundant.
- 1 UI generation & qualityWhich generator — or chain both?
- 2 Visual output — a medium splitDeck, video, or both?
- UI generation & quality — Different layers: pro-max designs the system, taste-skill polishes the output.
- Visual output — a medium split — An HTML deck and a rendered MP4 are different mediums, not duplicates. This one only looks like a dedup.
- The auditor — no overlap — Generates nothing; checks UI code against 100+ a11y/UX/perf rules. The free keep in all three drafts.
Cards show the current draft's verdict. Switch drafts below to re-cut them.
ui-ux-pro-max
KeptReads your requirements and outputs a tailored design system — patterns, sections, style, WCAG target — across platforms.
taste-skill
Kept“Anti-slop” polish: layout, typography, motion, spacing. Also ships image-gen for reference boards.
frontend-slides
KeptContent into single-file HTML decks (or PPTX to web). Pick a style from generated previews.
hyperframes
KeptRenders HTML/CSS/animation into deterministic MP4 video. 20 skills, launch clips to explainers.
Vercel web-design-guidelines
KeptNot a generator — an auditor. 100+ a11y/UX/perf rules. Ships inside Vercel Labs' agent-skills.
- ui-ux-pro-maxscaffold the system
- taste-skillpolish the components
- Vercel web-design-guidelinesaudit the result
- frontend-slidesdecks
- hyperframesvideo
A pipeline, not a choice — macro → micro → audit.
-
Draft A One generator + the auditor
One strong generator plus an objective checker covers most projects. Keep ui-ux-pro-max as the broader tool and let Vercel's auditor catch the accessibility gaps generators miss. taste-skill goes as overlapping polish; hyperframes goes unless you actually ship video.
Cut: taste-skill, hyperframes
-
Draft B Taste over templates
AI UIs fail at craft — spacing, hierarchy, motion — not at lacking a “design system,” and a generated system can itself be generic. So keep taste-skill, cut pro-max, keep the auditor. Video is a real differentiator, so hyperframes stays.
Cut: ui-ux-pro-max
-
Draft C Both generators, different layers recommended
They don't actually overlap, so chain them instead of choosing: pro-max scaffolds the system → taste-skill polishes the components → Vercel audits the result. Decks and video are just different mediums. Nothing here is redundant; the only cost is more skills installed.
Cut: nothing (choose deck vs. video by need)
Draft C is the most defensible: pro-max and taste-skill are macro vs. micro, and pipelining beats picking. Trimming hard? Draft B. Either way the deck/video pair stays split by medium — that's two tools, not a dedup — and the Vercel auditor is the free keep.
The sharpest clash on the list: one skill hides that AI wrote your text, another catches and fixes the same signals. They can't both be right for one job, so this dedup is a values call.
Two decisions: which research pipeline, and the loaded one — mask vs. improve.
- 1 Research pipelinesWhich research pipeline — or chain them?
- 2 Writing style — the real dedupMask the AI signals, or repair them? They can't coexist.
- Research pipelines — They overlap only in the middle — lit review, citations, writing. scientific-agent-skills owns the computation; ARS owns the manuscript.
- Writing style — the real dedup — Same target, opposite intent: humanizer masks machine-sounding prose, ARS repairs it. Keeping both is incoherent.
- Comprehension helpers — no overlap — context-engineering compresses source piles into briefs; understand-anything maps them into a searchable graph.
Cards show the current draft's verdict. Switch drafts below to re-cut them.
academic-research-skills (ARS)
KeptResearch → publication: reference hunting, citation-hallucination auditing, consistency gates, and a writing check that flags machine-sounding prose to fix.
scientific-agent-skills
Conditional148 domain skills — genomics, chem, imaging, physics, ML — plus a scientific-communication suite. A computational research assistant.
humanizer
CutStrips the signals of AI-generated writing so text reads as human. One portable Markdown file.
context-engineering
CutCurating and compressing large context — useful for digesting source piles into briefs before you write.
understand-anything
KeptPoint it at a docs set and get a searchable knowledge graph. Good for literature comprehension.
- ARSwrite + audit
- understand-anythingdigest the literature
- scientific-agent-skillsonly if computational
humanizer is cut on principle — it fights ARS's own quality checks.
-
Draft A Serious academic writing recommended
For real papers, integrity is the whole game: ARS audits citations for hallucination and gates output on consistency. humanizer is cut on principle — masking AI authorship is the opposite of academic integrity, and it fights ARS's own checks. Keep scientific-agent-skills only if your work is computational.
Cut: humanizer (always); scientific-agent-skills (unless STEM)
-
Draft B Non-academic / content writing
For blog posts and marketing, ARS's integrity machinery is overkill. humanizer becomes the honest keep — it makes drafts read naturally and does nothing else. Pair it with context-engineering and understand-anything. No academic pretense, so the mask-vs-improve clash never arises.
Cut: ARS, scientific-agent-skills
-
Draft C STEM researcher, full pipeline
ARS and scientific-agent-skills look redundant but are sequential: one does the science, the other turns results into a defensible manuscript. Chain them. humanizer is still cut — doubly so for published science. The heaviest stack, and the only one covering bench to publication.
Cut: humanizer
Pick by output: C for computational science, A for non-STEM papers, B for content. The firm call is the style dedup — humanizer and ARS never coexist, and humanizer only survives in the explicitly non-academic draft.
The grab-bag: an efficiency hack, the installer the ecosystem runs on, a theory course, and two security libraries. Almost nothing here competes.
One real dedup — security, broad vs. deep. The rest are constants.
- 1 Security — broad vs. deepBroad framework-mapped coverage, or deep recon tradecraft?
- The constants — no overlap — An installer, a free efficiency win, and the theory layer. Each has its own lane and survives every draft.
- Security — broad vs. deep — cybersecurity-skills is the superset; claude-osint goes far deeper on recon alone. Both are authorized-use-only — the dedup is about coverage.
- Cross-refs — decided elsewhere — memU is settled by the memory dedup; hyperframes by the UI/UX output call. Both sit out this switcher.
Cards show the current draft's verdict. Switch drafts below to re-cut them.
add-skill (npx skills)
KeptThe CLI for the whole ecosystem — installs any skill into Claude Code, Codex, Cursor and 70+ agents from a GitHub shorthand.
caveman
KeptTerse “caveman-speak” answers — ~65% fewer output tokens, with code and errors kept byte-for-byte exact.
context-engineering
KeptA teaching collection: context degradation, compression, memory systems, tool design, evaluation. Learn why the other skills work.
cybersecurity-skills
Kept817 skills across 29 domains, each mapped to 6 frameworks (MITRE ATT&CK, NIST CSF, ATLAS…). Authorized use only.
claude-osint
Cut90+ recon modules, secret regexes, dorks, attack-path templates — deep external-recon tradecraft for authorized work.
memU
Decided elsewhereFile-based long-term memory. Decided by the memory dedup in Software engineering, so it sits out this switcher.
hyperframes
Decided elsewhereHTML→MP4 video. Decided by the output call in UI/UX, so it sits out this switcher.
- add-skillinstall everything
- cavemansave tokens
- context-engineeringtheory
- cybersecurity-skillsbroad, framework-mapped security
Breadth and compliance mapping beat niche depth for a generalist.
-
Draft A Breadth wins recommended
If security is one capability among many, one framework-mapped library beats a niche one. cybersecurity-skills covers defense, DFIR, offense and fraud, all tied to industry frameworks. claude-osint's recon strength already exists inside the 817, if shallower.
Cut: claude-osint
-
Draft B Depth wins
If you actually do recon — bug bounty, red team, authorized assessment — a library that's one skill deep per topic won't cut it. claude-osint's 90+ modules and attack-path templates will. Keep the specialist, drop the generalist as more than you'll use.
Cut: cybersecurity-skills
-
Draft C Both, layered
They sit at different layers, so keeping both is coherent: cybersecurity-skills is the map, claude-osint the deep dive for the recon phase. Scope with the broad library, switch to the specialist when depth matters. The only cost is a second install.
Cut: nothing on the security axis
add-skill, caveman and context-engineering survive every draft — an installer, a free efficiency win, a theory layer. The security call is about your use: A for a generalist, B if you do recon, C if security is a real focus.
Run all twelve drafts and a few skills never get cut. Install those first — nothing in this post threatens them.
Vercel Labs' agent-skills
The quality gate in engineering, the only auditor in UI/UX. Survives all six drafts across both.
understand-anything
Comprehension, not workflow, so it never competes with a methodology.
add-skill · caveman · context-engineering
An installer, a free ~65% token saving, and the theory layer. Constants everywhere.
The rest is four values calls, not feature comparisons. Methodology: one heavy framework, or a light pair you invoke deliberately. Memory: task state alone, or task state plus knowledge. Mask vs. improve: humanizer and ARS can't coexist, and for academic work that isn't close. Security: broad, or deep.
Everything else is installing more than you'll use.
Thanks for reading · back to home