The Claude Code add-ons with the strongest evidence in our corpus are free patterns, not packages: Boris Cherny's plan-first CLAUDE.md, a primer.md handoff file, and hooks that enforce your rules. Among installs, keep Anthropic's frontend-design and skill-creator, Impeccable for design, Caveman with Codeburn for token cost, and find-skills for discovery. Skip GSD, whose repo is archived and which used about six times the tokens of plain Claude Code in a head-to-head test (Comment “agent” to get my Claude code guides).
Last verified 2026-10-22 · GitHub status checked through the GitHub API on that date
We went through the 68 saved posts in the DejaViewed corpus that cover Claude Code skills, plugins, CLAUDE.md files, hooks and memory add-ons (19 S-tier, 25 A-tier, 21 B-tier, 3 C-tier by our ratings). Every item below is something at least one of those posts actually covers. The verdict weighs how strong that evidence is, whether the repo is still maintained, and what it costs you in tokens or setup. The result is 13 keep, 19 try and 3 skip.
Skills are folders with a SKILL.md file that Claude loads when a task matches [1]. Plugins bundle skills, hooks, commands and MCP servers and install from a marketplace with /plugin[2]. Hooks are shell commands that run on events such as before a tool call [3].
Status comes from the GitHub API on 22 October 2026 [4]. Maintained means a commit in the last 90 days, slow updates means 90 to 365 days, stale means older, and archived means the owner archived the repo. Unverified means no repo could be matched to the entry. Install commands are copied from the entry or the repo README, never written by us.
Skill that generates stills for a site, turns approved ones into video clips with…
Caveman (JuliusBrussee/caveman)
KeepMaintainedskill
Skill that makes Claude answer in terse 'caveman' style at light/full/ultra levels to cut output tokens without changing code or reasoning.
Why keep: Two S-tier and two B-tier entries; savings apply to response text only, not total usage. Very actively maintained, Apache-2.0. One entry notes skills and context hygiene save more.
Anthropic's official skill that makes Claude decide on a visual direction before writing UI code, to avoid generic-looking pages.
Why keep: Named in three A-tier entries as the baseline design skill, paired with CLAUDE.md design rules and screenshot self-review. Official Anthropic repo, actively updated.
/graphify skill that turns a folder of code, docs and PDFs into a queryable knowledge graph the agent reads instead of grepping files.
Why keep: S-tier walkthrough plus B-tier reel; the 70x and 71.5x token-saving numbers are the creators' unverified claims. Very actively maintained, Apache-2.0.
Microsoft's Playwright command line that coding agents use for browser automation and screenshots instead of the Playwright MCP server.
Why keep: S-tier and A-tier roundups both prefer the CLI over the MCP for agents; design entries rely on screenshots for self-review. Official, maintained.
Anthropic's skill for writing new skills and testing, benchmarking and evaluating existing ones.
Why keep: Called the most powerful item in an S-tier roundup; a B-tier entry uses it to cut token use by moving context into skills. Official, maintained.
Skill that removes signs of AI-generated writing, based on Wikipedia's 'Signs of AI writing' list.
Why keep: A-tier entry names the repo and shows extending it with research-paper tells and enforcing it via a CLAUDE.md rule or hook. Actively maintained, MIT.
Plugin that records what the agent does each session, compresses it, and injects relevant context into later sessions.
Why try: S-tier list plus a B-tier reel (repo inferred from features). '95% fewer tokens' claim is unmeasured in the entries. Maintained, Apache-2.0; runs a background worker.
Codex plugin for Claude Code (openai/codex-plugin-cc)
TrySlow updatesplugin
OpenAI's plugin that lets Claude Code call Codex to review code or take delegated tasks.
Why try: A-tier and B-tier entries use it for second-opinion audits and to split work across two subscriptions. Official repo; updates slowed (last commit July 2026). Needs an OpenAI account.
Skill pack that makes Claude Code brainstorm, write a plan, use test-driven development and request code review before calling work done.
Why try: Two S-tier entries recommend it, but the only head-to-head in the corpus (Comment “agent” to get my Claude code guides) found plain Claude Code as good: about 250K vs 200K tokens and three times slower for the same landing page. Very actively maintained. Try it on a multi-step build, not small tasks.
UI UX Pro Max (nextlevelbuilder/ui-ux-pro-max-skill)
TryMaintainedskill
Design-knowledge skill with UI styles, color palettes and font pairings that generates a design system before building interfaces.
Why try: One S-tier list mention plus a B-tier 'websites in one line' reel with no before/after. Very actively maintained, MIT. Overlaps with Impeccable; pick one design skill.
Skill plus CLI that lets agents read and search Reddit, X, YouTube, GitHub and other sites without paid APIs.
Why try: A-tier demo: Claude fails to read a Reddit post and an X thread, then succeeds after install. Actively maintained, MIT. Relies on scraping backends that can break.
Motion-design skill that builds or audits animations using guidance distilled from Emil Kowalski, Jakub Krehel and Jhey Tompkins.
Why try: One A-tier entry ranks it among three must-have design skills. Repo updates have slowed (last commit May 2026). Niche: only matters if you ship animation.
Skill that runs a question past five subagent advisors with different thinking styles, has them peer-review anonymously, and returns a verdict.
Why try: B-tier entry (creator says he made the repo; credits Ole Lehmann and Karpathy's idea). Small repo, slow since April 2026, no license. Good antidote to agreeable answers.
Microsoft optimizer that rewrites a skill's instructions from run results and keeps only edits that beat a benchmark.
Why try: S-tier entry; claims of beating hand-written prompts and TextGrad come from the authors, not tested. Maintained, MIT. Needs an eval set to be useful.
Skill file listing banned filler phrases, throat-clearing openers and adverbs so prose stops reading as AI-written.
Why try: A-tier entry: removes most tells but can't add a point of view; last 20% still needs a human. Repo is slow (last commit March 2026). Overlaps Humanizer.
Spec-driven framework that interviews you, plans phases and executes them with verification on top of Claude Code.
Why skip: The original repo is archived and the project moved to open-gsd/gsd-core, which no entry covers. In the same head-to-head (Comment “agent” to get my Claude code guides) it used 1.3M tokens and 1h45 against 200K and about 20 minutes for plain Claude Code, for the same result.
Skill where Claude plans and GPT audits the plan, then roles swap so GPT builds and Claude audits.
Why skip: A-tier idea but the skill is comment-gated and no repo is named; several unrelated 'claudex' repos exist. Get the same cross-audit loop from the official Codex plugin.
Periodically audit every CLAUDE.md rule and skill and delete instructions newer models no…
Boris Cherny's CLAUDE.md workflow
KeepPatternpattern
Plan first, push hard work to subagents, log every correction to a lessons file, verify before marking done, run several sessions in parallel worktrees.
Why keep: Strongest evidence in the corpus: two S-tier and two B-tier entries tracing back to the Claude Code creator's own setup. Free, no install.
Claude Code hooks that run on events: block deletions into a staging folder, chime when input is needed, force a writing pass or audit on certain files.
Why keep: Three A-tier entries show concrete hooks; an S-tier report cites a stop hook that caught lazy behavior 173 times. Hooks enforce what CLAUDE.md can only ask.
Keep agent memory as a linked folder of markdown (often an Obsidian vault) structured per Karpathy's LLM-Wiki note, with session logs and per-project context files.
Why try: Four A-tier entries show working setups but most templates are comment-gated. Plain files, so it's cheap to try and portable across models.
Record a process by hand (e.g. Loom transcript), have Claude write it as a skill, patch it on failure, then schedule it; optionally a fixer skill improves the rest.
Why try: Two A-tier and three B-tier entries; one tests a skill against an n8n workflow and prefers the skill. Keep a human check on creative work.
Run LightRAG graph retrieval locally, then have Claude turn its server API endpoints into skills so it can query thousands of documents.
Why try: S-tier walkthrough with concrete setup (Docker, let Claude install it, wrap endpoints as skills). LightRAG is actively maintained, MIT. Only worth it past a few hundred docs.
Start with the free patterns: a short CLAUDE.md in Boris Cherny's style, a primer.md for session handoff, and one or two hooks for the rules you keep repeating. Then add one design skill (Impeccable or Anthropic's frontend-design, not both plus UI UX Pro Max), and Caveman with Codeburn once token cost matters.
Every skill loads text into context and every plugin can run code on your machine. Read a skill's files before you install it, and remove the ones you stop using. One S-tier post names old skills, idle MCP servers and a bloated CLAUDE.md as the usual places tokens leak (Claude — Burning tokens for no reason is a skill issue.).
What we could not verify
Token-saving claims such as claude-mem's "95% fewer tokens" and Graphify's 70x figure are the creators' own numbers. We have not measured them.
Claudex and the Scroll World skill are shared only with people who comment on the post, so we could not match them to a repo.
Twelve repos were matched by searching GitHub for the tool the entry names, then checking the repo description fits. Entries that name a repo directly are more certain.