Something would catch my eye. A good deal, maybe. Or just a title that sounded right in the moment. A little dopamine hit, "ooh, this looks good, let me get it," and onto the shelf it went. Most of them stayed exactly there. Unread. Just... collected.
Really, I was just building my shelf-esteem.
Now I recognise I'm doing the same thing with AI skills.
The AI Skills Hoarding Cycle
If you're anywhere near AI content right now, you know the pattern.
Someone showing off a workflow, and the comments are just "link please," a keyword "XYZ," over and over.
Someone drops the repo. You click through, you install it onto Claude or ChatGPT or whatever you're running, and it takes maybe 2 minutes.
You now have a new skill. It’s like Neo in the Matrix being able to install new skills - that's the easy part.
So why does it still feel like a mess?
Because 2 minutes later, I've got 15 spanking brand-new skills installed into my account, and I genuinely cannot tell someone what they do.
Skills are like the eggs in Alien, they wait until something triggers them
I've installed them the same way I've been buying books, on impulse. In the moment.
And I move straight on to the next video without actually using those skills.
And it compounds.
New skills get recommended every single day. Half the time I click through excited about some "new" find, only to realise I already installed it weeks ago and completely forgot.
Same shelf-esteem, just showing up as skills instead of books.
You're probably doing this too right now, saving links and installing skills without ever making sense of what you've got or how to use it.
What I'm actually doing about it
I'm turning this into a growing record, a running post of exactly what skills I've installed, on my computer and on the web versions, and how each one is actually meant to be used.
A working list I can come back to.
Because honestly, I may never remember why I installed half of these in the first place. But if I've got a proper overview, I can look at it and actually decide, keep this, update that one now that the repo's changed, delete this because I clearly never needed it.
I don't want another shelf. I want skills I actually open and use.
I'll be updating this as I go. If you're sitting on the same pile of unused skills, you already know exactly what I'm talking about.
What's Actually On The Shelf
So here's the shelf, opened up properly.
Below is the real list: every skill I've got running in Claude and Claude Code, on web and desktop, no more guessing what's actually installed. Each entry gets a date I added it, where I got it from, why I bothered in the first place, whether it clashed with something I already had, whether it's been improved or folded into anything else since, and whether it's something I actually built myself instead of grabbed off someone's repo.
Last updated: 31 August 2026.
▸
The doc-handling pack
Documents
✓
✓
Jul 28, 2026
From: docx, pdf, pptx, xlsx, skill-creator, webapp-testing, frontend-design, vercel-react-best-practices, web-design-guidelines. Genuinely not sure if I chose these or they came bundled with Claude Code by default. Why: All nine grabbed on the same day. No idea. Never consciously decided on any one of these nine. Clash: Not that I've noticed, but I've barely touched them. Improved: No.
From: A free community lesson pack of ready-made Claude Code plugins, worked through one at a time. Why: Builds a paste-ready prompt for a builder-and-critic loop, useful when I want two passes: one to build, one to tear it apart against a standard I've actually named. Clash: No. Improved: Not yet, still exactly as installed.
Installed
▸
doctor-plus
Coding & Dev
✓
✓
Aug 8, 2026
From: Same lesson pack as gauntlet-loop. Why: An extended version of Claude Code's own /doctor health check, audited against Anthropic's own context-engineering guidance. Clash: No. Improved: No.
From: Same lesson pack, built by Peter Yang. Why: To catch the AI writing tells I can't always hear myself, on my own drafts before I publish them. Clash: Yes, with asher-voice below. Its defaults want to flatten fragments, ellipses, question-headers and callback endings as generic AI tells. Those are just how I've written since 2013. Improved: Resolved, not removed. asher-voice now explicitly overrides it on those four patterns, so the two run together instead of fighting over my own sentences.
From: Same lesson pack again. Why: An action-first output style. Only turns on if I explicitly type the command for it, off by default, so it doesn't quietly change how everything else behaves. Clash: No, precisely because it stays off unless I ask for it. Improved: No.
Installed
▸
fabsol + model-router
Coding & Dev
✓
✓
Aug 8, 2026
From: Set up together as one task, routes Claude Code sessions through a local proxy to other models. Why: Wanted the option to run sessions on a different model through my own proxy instead of only ever using Claude directly. Clash: No. Improved: Half-working. The free-tier options I tried routing through turned out unusable, one wouldn't authenticate properly, another kept hitting rate limits fast. Still sitting there, not fully sorted.
Modified
▸
The Cloudflare pack
Coding & Dev
✓
✓
Aug 8, 2026
From: agents-sdk, cloudflare, cloudflare-email-service, cloudflare-one, cloudflare-one-migrations, durable-objects, sandbox-migrate-to-next, sandbox-next, sandbox-stable, turnstile-spin, web-perf, workers-best-practices, wrangler. From Cloudflare's own skill catalogue. Why: Thirteen skills grabbed in one sitting, same evening as the lesson pack above. Honestly can't remember deciding to grab all thirteen. Wanted a couple of them, the rest came along for the ride. Clash: Not that I've noticed. Improved: No.
Installed
▸
recall
Memory & Data
✓
✓
Aug 9, 2026
From: Built it myself. Why: Technically a slash command plus a script, not a Skill in the strict sense, but it does the same job. Needed a fast way to look things up across my own memory files and project docs without reading everything by hand every time. Clash: No. Improved: No, still the original version.
Custom
▸
asher-voice
Writing
✓
✓
Aug 12, 2026
From: Nowhere. Built it myself, from 21 of my own posts going back to 2009, plus direct feedback on drafts it got wrong the first time. Why: Every generic AI voice was cutting my stories down to a highlight reel instead of the actual story. Needed something trained on how I actually write. Clash: Yes, with no-ai-slop, see above. Improved: Two full iterations, each benchmarked with-skill against without-skill on the same test posts, before I trusted it enough to actually use it.
From: A marketplace built around Andrej Karpathy's own observations on common LLM coding mistakes. Why: Wanted a standing check against overcomplicating things, unnecessary abstractions, assumptions I hadn't stated out loud. Clash: No. Improved: Its entire content got copied straight into my global CLAUDE.md as an always-on rule instead of staying a Skill I'd have to remember to invoke. Right now it isn't technically "installed" as a Skill at all, it got absorbed into something more permanent.
From: A commercial-grade design toolkit whose main skill is literally called "design", same name as one I already had running. Installing the whole thing would've silently overridden that. Why: Only brought in the reference half, styles, colour palettes, font pairings, accessibility rules, searchable through a local script with no network calls. Left out the half that generates actual logo and icon images, since that needs a paid Gemini API key I haven't set up. Clash: Yes, the name collision above. Avoided it by only taking half, under its own name. Improved: Yes. Fixed a few hardcoded script paths left over from how the original expected to be installed, and dropped a test suite that had no reason to ship with it.
Why: One of a batch of 5 design skills grabbed the same day. This one's the anti-slop discipline, three dials for variance, motion and density, a hard ban on em dashes, a 70-item pre-flight checklist before anything ships. Pairs with no-ai-slop, but for interfaces instead of writing. Clash: None. Straight copy. Improved: No.
From: Nowhere to copy from, technically. The source is a curated list of links to other people's DESIGN.md files, Stripe, Linear, Apple, Notion, Tesla, 68 of them, not a skill file itself. Why: Same batch of 5. Written from scratch as a small pointer skill so pulling a real brand's colour and type system as a starting point is one lookup instead of 68 browser tabs. Clash: None. Improved: No, unchanged since I wrote it.
Why: Same batch. Nothing else on this list covered motion or animation specifically, so it filled a real gap. Reads a brief in one of two modes, build or audit, and weighs three different designers' philosophies depending on what the project actually calls for. Clash: None. Improved: No, straight copy.
Why: Same batch, originally called "One Skill to Rule Them All." Watches a session for patterns worth turning into a new skill or an improvement to an existing one, and logs them. Basically automates the kind of review this whole post does by hand. Clash: None noticed. Improved: No, but it wants a CLAUDE.md instruction or a session-start hook to trigger reliably on its own, and that part's not set up yet.
From: Same RoboNuggets lesson pack as the others above. Why: Blocks dangerous commands (git reset --hard, recursive deletes, DROP TABLE) before they actually run, with an allowlist and a one-time escape hatch for when I genuinely mean it. Clash: No. Improved: No, still exactly as installed.
From: Installed through Claude Code's own official plugin marketplace, one command, not hand-copied. First tried in a temporary remote sandbox session, which didn't stick since that environment doesn't persist. This time it's on my actual machine. Why: Wanted a second memory system running alongside the one I already use day to day (the recall entry below), not replacing it. Clash: Checked for it directly and confirmed it left the existing recall system alone. Improved: Actually running now, the background worker is live and confirmed reachable, unlike the earlier attempt.
A Chrome extension that auto-extracts a site's visual design (typography, colours, spacing, components) while you browse it, then generates a DESIGN.md or SKILL.md file from it -- meant to be handed straight to Claude Code or similar tools as a design-system blueprint. Genuinely useful-sounding, and closer to design-references above than I first guessed -- one pulls a curated library of existing brands' DESIGN.md files, this one would generate my own from any site I point it at. Never actually installed it.
An open-source AI gateway -- one local endpoint instead of juggling separate API keys and dashboards for 351 different providers, with auto-fallback across subscription/API-key/cheap/free tiers and prompt compression on top. Basically the same problem fabsol + model-router already solves for me, just a different implementation. That's almost certainly why I never followed through on installing this one -- I'd already built my own version of the same idea.
From: A context-compression tool, shrinks tool outputs, logs and conversation history before they reach the model. Installed as its own command-line tool, not copied, since there's nothing to copy, just a skill file written to document how to use it. Why: Real potential to cut token usage on long sessions. Had been sitting undecided for a while, not rejected, just never revisited. Clash: Checked carefully against model-router and fabsol, which also route sessions through a local proxy. Different ports, and model-router already keeps itself off the main session's environment by its own rule, so normal use of each doesn't collide. Documented the one thing to avoid, stacking both on the same command. Improved: N/A, just installed. Not wired up as an always-on wrap, opt in per session.
From: A skill pack for making AI ad videos and images through the Arcads API. Why: Turned out to be six real skills, not one, once I looked properly. Brought all six in. Clash: None with anything already installed, but the skill file itself gave an instruction to keep pushing its own affiliate signup link at every opportunity. Kept the plain signup info, stripped that instruction out. Improved: Yes, restructured how the files reference each other to fit how skills are organised here, and fixed two script paths that assumed a different setup. Sitting on a branch, not live yet, since none of it does anything without an Arcads account and I haven't paid for one.
An open-source, local-first voice studio, voice cloning across several TTS engines, speech generation in 23 languages, dictation with a global hotkey, and voice output for AI agents, all processed locally rather than through a cloud service. Never actually installed it. Voice/audio work isn't something I'm doing yet, and unlike arcads-claude-code above, this one's a full desktop app rather than a skill, so there was nothing to actually bring in even if I wanted to.
From: Microsoft's real tool for converting files to Markdown. Doesn't come as a Claude skill on its own though, just a command-line program. Why: Second look after realising it covers things the doc-handling pack above doesn't, web pages, zip files, YouTube transcripts. Had to write the skill wrapper myself. Clash: None. Improved: No, unchanged since I wrote it. Sitting on the same kind of unmerged branch as arcads-claude-code above, still needs a look before it's actually live.
This one's actually a full open-source CRM, not a Claude skill at all. Might have been pasted in by mistake -- keeping it here as a Shelved entry rather than silently dropping it, since it wasn't clear whether it was meant to be here.
From: Anthropic's own official plugin marketplace. Got this one wrong before, thought it was just the marketplace mechanism itself rather than an actual skill. It's a real one: scans a codebase and recommends the top automations, hooks, skills, MCP servers, subagents, slash commands, worth adding. Why: Official, read-only, doesn't touch any files. Low-risk enough to just install and see what it suggests. Clash: None. Some overlap in spirit with task-observer above, that one watches live sessions for patterns, this one does a one-off codebase scan, different enough to keep both. Improved: No, just installed.
From: A 25-skill engineering lifecycle pack, spec, plan, build, test, review, ship, by Addy Osmani. Why: Real depth on things nothing else covers: enforced TDD, OWASP security hardening, ADRs, CI/CD patterns, observability. Installed all 25 on purpose, including 2 that overlap what's already here. Clash: Yes, two, kept anyway. Its web-performance-auditor duplicates the existing web-perf skill. Its interview-me is nearly the same idea as grill-me below, its own description even namechecks "grill me." Improved: Yes. Flattened the repo's own skills/ + references/ split to match how skills sit here, one shared references/ folder, paths rewritten.
From: 50 marketing skills, SEO, CRO, copywriting, ads, growth, sales/GTM, strategy, by Corey Haines. Why: Directly relevant to the coaching business. Installed the whole set rather than a curated slice. Clash: None with anything already installed. Improved: Yes, same fix as agent-skills, shared tools/ folder flattened to a sibling directory, paths rewritten.
From: One skill out of Matt Pocock's broader skills collection (skills/productivity/grill-me), not a separate repo. Why: Relentlessly interviews about a plan or decision until every branch is resolved. Straightforward, low-risk single skill. Clash: None. Improved: No, straight copy.
From: The skill half only, terse "caveman-speak" output mode to cut response tokens. The source repo also ships Caveman Proxy, a separate input-compression proxy under BSL-1.1, not MIT like the skill. Why: The install command first given for this one pointed at the wrong repo entirely (mattpocock/skills, which doesn't contain it). Corrected to the real source, JuliusBrussee/caveman, skill only. Clash: The proxy half would have been a third local token-compression proxy alongside model-router and headroom, so it was left out on purpose. The skill half has no such overlap. Improved: No changes to the skill file itself, just scoped which half of the repo to take.
From: An MCP server pitched as solving context/token bloat across sessions, sandboxed tool output, session continuity, code-execution-over-reading. Why not: Its README shows "used at Microsoft, Google, Meta, Amazon, IBM, NVIDIA..." company logos that all link nowhere, unsubstantiated social proof for a tool asking for hooks on every single lifecycle event plus an arbitrary-code-execution MCP tool plus a hosted dashboard it can phone home to. Also just redundant with claude-mem and headroom, already installed, for the same problem. Clash: Would have been a third tool in the same space as claude-mem and headroom. Improved: N/A, never installed.
From: NVIDIA's own security scanner for Claude Code/Codex/MCP skills, installed as a CLI tool, not copied, wrote a thin skill wrapper documenting it. Why: A real automated first pass for exactly the kind of manual safety review this whole table represents, catches unresolved external references, embedded directives, executable scripts. Clash: None. Improved: N/A, just installed. Tested it for real against two of today's other new entries before trusting it, not just taking its word for it.
Comments