Smorster

9 Claude Code Plugins to Build 10x Faster

Austin Marchese · 14m 25s · Watch on YouTube · 26 sources

Decision Card

Effort: Per-plugin install via Claude Code’s /plugin marketplace — most take ~2 minutes each (e.g. /plugin marketplace add openai/codex-plugin-cc then /plugin install); budget an afternoon to install, wire up MCP connectors (Higgsfield, Firecrawl, Exa, Morph need accounts/keys), and decide which 2–3 you actually keep.

Honest take: This is an affiliate-driven roundup — every plugin is “linked in the description,” one (Higgsfield) is a paid sponsor and another (buildpartner.ai) is the creator’s own product, so the framing leans promotional rather than evaluative. The “10x faster” and per-plugin savings figures are explicitly the creator’s own “theoretical statistics” he admits he “can’t test directly,” and even the Caveman plugin’s own README pegs real-world savings at 4–10%, not the “75%” the marketing implies.

Concrete next steps (per item — adopt | try | skip):

  • Caveman — try: free, reversible, genuinely cuts output verbosity. Install from github.com/juliusbrussee/caveman. Skip if you rely on Claude’s explanations to learn.
  • Exa + Firecrawl — adopt if you do research-heavy work: semantic discovery + clean scraping is a real, well-understood combo (exa.ai, firecrawl.dev). Skip if Claude’s native web search already covers your lookups.
  • Compound Engineering — adopt: the plan→work→review→compound loop is the most substantive idea here; it’s free and open-source (github.com/EveryInc/compound-engineering-plugin). Skip if you don’t iterate on the same codebase repeatedly.
  • Higgsfield — try only if you generate marketing media; it’s the paid sponsor, costs money, and is unrelated to “building software faster.” Skip if you don’t need image/video assets.
  • Anthropic official plugins (skill-creator, frontend-design, security-guidance, legal) — adopt: first-party, maintained, free (claude.com/plugins). Skip the legal one for anything real (“not legal advice,” per the video).
  • OpenAI Codex plugin — adopt for a second-opinion reviewer (github.com/openai/codex-plugin-cc). Skip if you don’t have/want an OpenAI subscription.
  • buildpartner.ai — skip unless the “expert advice” framing resonates; it’s the creator’s own paid product and the weakest-evidenced claim.
  • Morph — try if mechanical edit/search latency annoys you (morphllm.com); it costs money and the speedups are model-claimed. Skip if you’re on a flat Max plan and not latency-bound.
  • CodeBurn — adopt: free, open-source token observability, useful regardless of the rest (github.com/getagentseal/codeburn). Skip if you never hit usage limits.

TL;DR

Austin Marchese walks through nine Claude Code plugins — spanning output compression, web research, media generation, multi-model review, and token observability — that he claims help you “build 10x faster” and spend less. The genuinely useful, free picks are Compound Engineering, Anthropic’s official plugins, and CodeBurn; the rest are situational, paid, or promotional (one sponsor, one self-owned product).

Key Points

  • Caveman compresses Claude’s output by forcing a terse “caveman” speaking style, cutting tokens and “AI slop” 00:01
  • Exa adds semantic search (matches meaning, not keywords) where Claude’s native keyword search returns SEO-optimized noise 01:35
  • Firecrawl pulls clean content from pages — handling JavaScript and stripping headers/footers/buttons before it reaches Claude 02:18
  • Compound Engineering encodes a five-step loop — plan, work, review, compound (learn), repeat 03:43
  • The creator cites Boris Cherny (“creator of Claude Code”) that “the most important thing you can do is get the plan right” 04:27
  • Higgsfield (the video’s sponsor) adds an MCP server so Claude can generate images and video with project context 04:46
  • Anthropic’s official plugins — skill-creator, legal, frontend-design, security-guidance — are favored because they’re maintained by the model’s maker 06:19
  • OpenAI’s Codex plugin lets you run a different model inside Claude Code (/codex:rescue), illustrating multi-model independence 07:53
  • He argues current model pricing is VC-subsidized — paying $200/mo for Claude Max while consuming ~$1,800 in tokens 08:45
  • Morph claims to speed mechanical work: fast-apply edits “up to 8x faster, 90% cheaper,” plus WarpGrep search and session compaction 11:21
  • CodeBurn surfaces where tokens go and suggests copy-paste fixes to cut spend 12:55

Notable Quotes

“Once I installed this, I realized how much time I was wasting reading AI slop.” 00:01

“If you’re just one Anthropic-released plugin away from being absolutely cooked, maybe rethink what you’re building.” 06:45

“A lot of your token spend is on mechanical work, not actually reasoning. So, Morph cuts the mechanical waste.” 12:28

Verified Claims

Boris Cherny is the creator of Claude Code. 04:27

Exa uses semantic/neural search while Claude’s native search is keyword-matched. 01:35

Firecrawl handles JavaScript and returns clean LLM-ready markdown, stripping page noise. 02:18

Higgsfield shipped an official MCP server letting Claude generate images/video. 04:46

OpenAI released an official Codex plugin for Claude Code with a /codex:rescue-style delegation command. 07:53

Morph’s fast-apply edits code dramatically faster (he says “8x faster, 90% cheaper”). 11:21

  • Morph — Fast Apply Model (cites 10,500 tok/s, 98% accuracy); Morph MCP
  • Verdict: Inconclusive — Morph’s own marketing cites high throughput and “35% faster edits,” but the specific “8x / 90% cheaper” framing and the $3.50→$0.80 example are the creator’s own admittedly untested “theoretical statistics.”

Caveman cuts a large share of output tokens. 00:01

  • github.com/juliusbrussee/caveman (markets “~65–75% token cut”)
  • Verdict: Disputed — the plugin’s own README/coverage notes real-session savings are closer to 4–10%, not the headline 65–75%.

Claude Max ($200/mo) is heavily subsidized vs. real token cost (he cites ~$1,800). 08:45

Tools, Papers & Standards Mentioned

Follow-up Questions

  1. In a controlled benchmark on a real repo, how much wall-clock time and token spend does Morph (fast-apply + WarpGrep) actually save versus native Claude Code edits — and does it recoup its own subscription cost?
  2. Several of these plugins idle as MCP servers; given that “five idle plugins can burn 55,000 tokens before you type a word,” what’s the net token economics of installing all nine versus the savings they each claim?
  3. How durable is the multi-model “model independence” thesis — if VC subsidies end and prices rise, do plugins like the Codex bridge meaningfully lower lock-in, or just shift the dependency?

Sources