Smorster

9 Claude Code Plugins to Build 10x Faster

Austin Marchese · 14m 25s · Watch on YouTube · 21 sources

Decision Card

Effort: Each plugin is a minutes-long install (/plugin marketplace add ... + /plugin install ...); adopting the whole stack is an evening of setup, but Exa, Firecrawl, Morph, and Higgsfield each require accounts, API keys, and paid tiers to be useful.

Honest take: The video is sponsored by Higgsfield (disclosed at 04:42), plugin #7 is the creator’s own product, and the Morph numbers quoted (“warp grep 50% cheaper, 28% faster”) exceed Morph’s own published figures (15% cost reduction, 28% faster) — so treat the “10x faster” framing as marketing, not measurement, especially since the presenter admits the Morph savings are “theoretical statistics” he can’t test directly.

Concrete next steps:

  • Caveman — adopt (5 min, free): install from github.com/JuliusBrussee/caveman; note it only cuts output tokens and adds ~1–1.5k input tokens per turn. Skip if you already keep responses terse via CLAUDE.md instructions.
  • Exa + Firecrawl — try (30 min, API keys needed): worth it if you do research-heavy sessions; docs at exa.ai and firecrawl.dev. Skip if your Claude use is pure coding with little web research.
  • Compound Engineering — adopt (10 min, free): /plugin marketplace add EveryInc/compound-engineering-plugin; repo at github.com/EveryInc/compound-engineering-plugin. Skip if you already run a plan→build→review→retro loop with your own skills.
  • Higgsfield MCP — skip unless you produce marketing/media assets; it’s the sponsor segment and image-to-video requires a ~$49/mo paid plan.
  • Anthropic official plugins (skill-creator, frontend-design, security-guidance) — adopt (10 min, free, first-party maintained): github.com/anthropics/claude-plugins-official. security-guidance ships enabled by default in recent Claude Code — check before installing redundantly.
  • OpenAI Codex plugin — try (15 min, needs ChatGPT account): github.com/openai/codex-plugin-cc; /codex:rescue offloads work to your ChatGPT quota when you hit Claude limits. Skip if you have no OpenAI subscription.
  • BuildPartner.ai — skip: creator’s own product, “trained on experts” claims are unverifiable, and generic expert-framework advice is available free.
  • Morph — try cautiously (30 min, paid): real product (docs.morphllm.com/guides/claude-code) but benchmark the cost delta yourself; the video’s own demo is a single anecdote. Skip if your sessions are small — fast-apply shines on large files and scattered edits.
  • CodeBurn — adopt (2 min, free, local-only): npx codeburn; github.com/getagentseal/codeburn. No reason to skip — it reads local logs, uploads nothing.

TL;DR

Austin Marchese walks through nine Claude Code plugins spanning output compression (Caveman), research (Exa + Firecrawl), workflow discipline (Compound Engineering), media generation (Higgsfield, the sponsor), Anthropic’s official plugins, multi-model fallback (OpenAI’s Codex plugin), his own BuildPartner.ai, edit/search acceleration (Morph), and token-cost observability (CodeBurn). The recurring thesis is that most token spend is mechanical rather than reasoning, so compressing output, improving input quality, and offloading mechanical work saves both money and time.

Key Points

  • Caveman condenses every Claude response into terse “caveman speak,” cutting fluff and output-token spend while keeping technical substance 00:01
  • Claude’s native web search is keyword-matched and tends to surface SEO-optimized content; Exa’s semantic search finds resources by meaning instead 01:37
  • Firecrawl extracts page content (including JavaScript-rendered and embedded resources) and strips headers, footers, and buttons before it reaches Claude — the stack is Claude for basic fetch, Exa to discover, Firecrawl to extract 02:20
  • Compound Engineering encodes a five-step loop — plan, work, review, compound (persist lessons learned), repeat — which the presenter credits for a $481k vibe-coded product launch 03:20
  • Higgsfield’s MCP is the sponsored segment: media generation inside Claude Code that inherits project context (brand guidelines, audience) instead of needing re-explanation 04:42
  • Anthropic’s official plugins recommended: skill-creator, legal, frontend-design (e.g. generate six landing-page variants), and security-guidance for pre-launch audits 06:19
  • OpenAI ships a Codex plugin for Claude Code; the argument for multi-model is second opinions plus hedging against the end of VC-subsidized pricing 07:53
  • Morph accelerates the three mechanical costs of Claude sessions — file edits (fast apply), codebase search (warp grep), and context compaction — with the claim that most token spend is mechanical, not reasoning 11:00
  • The presenter’s CodeBurn dashboard shows ~$1,800 of token consumption on a $200/mo Claude Max subscription, his evidence that current pricing is subsidized 13:24
  • CodeBurn also generates actionable token-saving suggestions you paste back into Claude Code to apply 13:43

Notable Quotes

“From the creator of Claude Code, Boris Cherny, the most important thing you can do is get the plan right.” 04:28

“if you’re just one Anthropic-released plugin away from being absolutely cooked, maybe rethink what you’re building.” 06:47

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

Verified Claims

  1. Claim: The Caveman plugin condenses Claude’s output into terse caveman-style responses, saving tokens. 00:01 Sources: JuliusBrussee/caveman on GitHub Verdict: Confirmed — the repo claims ~65% average output-token reduction, though only output tokens shrink and the skill itself adds ~1–1.5k input tokens per turn.

  2. Claim: Exa uses semantic (meaning-based) search rather than keyword matching. 01:37 Sources: Exa blog — semantic search technology Verdict: Confirmed — Exa indexes the web with neural embeddings; its auto mode also blends in keyword search for exact-match queries.

  3. Claim: Firecrawl handles JavaScript-rendered pages and strips navigation/footer/boilerplate to return clean content. 02:20 Sources: Firecrawl, firecrawl/firecrawl on GitHub Verdict: Confirmed — executes JavaScript before extraction and returns LLM-ready markdown with menus, ads, and banners stripped.

  4. Claim: Compound Engineering is an installable plugin implementing the plan → work → review → compound loop. 03:20 Sources: EveryInc/compound-engineering-plugin on GitHub, Every — Compound Engineering Verdict: Confirmed — built by Kieran Klaassen and Every, shipping /ce-plan, /ce-work, /ce-code-review, /ce-compound plus 37 skills and 51 agents.

  5. Claim: Boris Cherny is the creator of Claude Code and stresses that planning matters most. 04:28 Sources: Boris Cherny — LinkedIn (Creator & Head of Claude Code), Pragmatic Engineer — Building Claude Code with Boris Cherny Verdict: Confirmed — Cherny is creator and head of Claude Code at Anthropic.

  6. Claim: Anthropic publishes official plugins including skill-creator, frontend-design, security-guidance, and a legal plugin. 06:19 Sources: anthropics/claude-plugins-official on GitHub, Claude Code docs — discover plugins Verdict: Confirmed for skill-creator, frontend-design, and security-guidance (first-party, with security-guidance enabled by default in recent versions); the “legal” plugin did not appear in the sources checked — inconclusive on that one item.

  7. Claim: OpenAI ships a Codex plugin for Claude Code, with /codex:rescue as a key command. 07:53 Sources: openai/codex-plugin-cc on GitHub, OpenAI community — Introducing Codex Plugin for Claude Code Verdict: Confirmed — /codex:rescue delegates tasks to Codex, billed to your ChatGPT account rather than your Anthropic subscription.

  8. Claim: Morph’s fast apply is “up to eight times faster and 90% cheaper” and warp grep is “50% cheaper, 28% faster.” 11:21 Sources: Morph docs — Claude Code guide, morphllm/morph-claude-code-plugin on GitHub Verdict: Disputed — Morph’s published figures for WarpGrep plus compaction are 15% cost reduction, 28% faster, and +6% accuracy on SWE-Bench Pro; the video’s “50% cheaper” exceeds the vendor’s own claim, and the presenter concedes his numbers are theoretical.

  9. Claim: Claude Max costs $200/month while the presenter’s actual token consumption is closer to $1,800. 08:58 Sources: Claude pricing, Claude Help — What is the Max plan? Verdict: Confirmed for the $200/mo Max 20x tier; the $1,800 figure is his personal CodeBurn readout — plausible for heavy users but not independently verifiable.

  10. Claim: CodeBurn shows what you spend tokens on and suggests ways to reduce cost. 12:55 Sources: getagentseal/codeburn on GitHub, codeburn.app Verdict: Confirmed — free, open-source, local-only TUI/dashboard that parses session logs across 31 tools and prices every call by token type.

Tools, Papers & Standards Mentioned

Follow-up Questions

  1. What is the measured end-to-end cost delta (input + output + cache tokens) of running Caveman-style output compression, given that the skill adds ~1–1.5k input tokens per turn while only shrinking output?
  2. How does Morph’s fast-apply accuracy compare to Claude Code’s native str_replace editing on real refactors — does the 98%-accuracy merge model introduce silent edit errors that offset the speed gain?
  3. If per-token pricing is genuinely subsidized (the $200 subscription vs. ~$1,800 consumption claim), what usage patterns or plugin stacks would remain economical under true-cost API pricing?

Sources