Skills for Real Growth and Agent Operations
Agent skills I use for SEO, growth, and agent-operations work that needs evidence, execution, verification, and memory — not one-shot audits or slide-deck recommendations.
Real growth work is hard. One-shot prompts produce generic checklists. Framework-heavy workflows try to own the process, but they often bury the business context and make it hard to tell what actually shipped. These skills stay small, composable, and adaptable. They work with any model. Hack around with them. Make them your own.
Quickstart (30-second setup)
- Run the skills.sh installer:
npx skills@latest add JorgeMenaDev/skills
-
Pick the skills you want and which coding agents to install them on.
-
Open a session in a target repo and invoke the skill that matches the work — for example, start SEO work with seo-growth-workspace.
Most people begin with seo-growth-workspace. For unattended multi-phase shipping, add shiploop. When the question is where a plan should live, use work-tracking. To publish through Andy MCP, add andy-post.
crew
Use your configured T3 Code providers in one workflow. Crew opens visible worker threads, gives each a bounded brief and working directory, and brings its report back for verification. It includes the dispatcher and all helper scripts; authentication stays with your own T3 installation.
npx skills@latest add JorgeMenaDev/skills --skill crew
See setup and compatibility before your first run.
counsel
A two-vendor adversarial review protocol for hard-to-reverse architecture or design decisions. Two flagship reviewers from different vendors attack a file-backed proposal in parallel; the chair synthesizes, revises, and re-convenes for up to three rounds. A one-vendor run is a second opinion, not counsel. Runtime launch adapters stay consumer-local (e.g. Claude Code .claude/skills/counsel/).
npx skills@latest add JorgeMenaDev/skills --skill counsel
source-to-system
A system-first review loop for external articles, videos, repositories, prompts, and tactics. It captures the material claims in temporary scratch space, compares them with the user's existing systems, selects one owned improvement or no change, and stops for approval before durable workspace or external-system mutation.
npx skills@latest add JorgeMenaDev/skills --skill source-to-system
painted-music-video
A music video painted entirely in code: an original song with vocals (Google Lyria 3 Pro), watercolour animation in p5.brush, word-by-word karaoke, and chapters painted by parallel subagents from one storyboard, rendered to MP4 in headless Chrome. Ships the engine as templates and two worked examples (a 93 s intro film and a 62 s sequel made by a fresh agent from the skill alone). User-invoked: run /painted-music-video. Method inspired by PDoomVideo.
npx skills@latest add JorgeMenaDev/skills --skill painted-music-video
launch-video
A 15–20 s keynote-style launch film for one product feature, rendered entirely in code: one HTML file whose seek(t) draws any frame, cut to a royalty-free song's beat grid (fitted to the kick drums, so the payoff lands on the drop), real Mixkit sound effects placed by measured peak and mixed to -14 LUFS, and a Playwright + ffmpeg render at 60 fps with subframe motion blur. Ships the engine, the audio tools and a worked example (Andy Publicaciones). Method from zero (@twoclipping)'s article "How I Make $10K Launch Videos for $0".
npx skills@latest add JorgeMenaDev/skills --skill launch-video
scroll-film-landing
A landing page whose hero is a film the visitor scrubs with the scroll: AI stills of one continuous place, or the owner's real photos reframed to 16:9, joined by MiniMax H3 first/last-frame clips through the Higgsfield API, with real HTML captions and cards at each stop so text stays sharp. Ships the scrub engine (Astro + Lenis), the Higgsfield scripts (hidden key setup, free estimate before every paid call, spend log), an encoder that keeps poster and video colours identical, and a Playwright QA gate that measures the seams. Method after midudev's Arko landing.
npx skills@latest add JorgeMenaDev/skills --skill scroll-film-landing
youtube-thumbnails
YouTube thumbnails and channel banners painted by an image model (Codex image generation on a ChatGPT plan) and grounded in the real brand: the model paints one moment, the real logo is composited on top, and every candidate is judged at the 168×94 size phone viewers see. Ships the evidence behind each rule, prompt skeletons, a finish-and-export script and a review-sheet script.
npx skills@latest add JorgeMenaDev/skills --skill youtube-thumbnails
andy-post
Draft and publish social posts through Andy MCP. The operator gives a raw idea. The agent returns 2-3 drafts. After a pick, it publishes or schedules. Andy MCP is an active build: file skill or MCP friction on this repo.
npx skills@latest add JorgeMenaDev/skills --skill andy-post
grok-deep-research
An isolated deep-research loop over native xAI Grok Build CLI. It defaults to native web search, can expose only Firecrawl MCP when requested, audits evidence gaps in one temporary session, and emits a consistently structured cited report plus provider-usage metadata.
npx skills@latest add JorgeMenaDev/skills --skill grok-deep-research
Why These Skills Exist
I built these skills to fix failure modes I keep seeing in AI growth and agent-ops work.
#1: The Agent Gave Generic SEO Advice
"No-one knows exactly what they want."
David Thomas & Andrew Hunt, The Pragmatic Programmer
The problem. The internet already has enough SEO checklists. The useful work is deciding what matters for this business, this repo, this market, and this moment. Most agents skip that and optimize pages without understanding the business.
The fix is seo-growth-workspace — pull live Search Console, result and demand evidence, pick one to three sized bets a week, ship them, and stop the lanes that do not earn clicks.
- Before: "Improve meta descriptions across the blog."
- After: A
.seo/bets.md bet: the query and page behind it, its search demand, a live result check, a size scenario, a check date and a kill rule.
#2: The Next Session Forgot Everything
The problem. SEO compounds only when the next pass can trust the previous one. Without durable notes, agents re-audit the same surfaces, forget blockers, and repeat work that already failed.
The fix is a predictable workspace. seo-growth-workspace keeps .seo/: context, owner decisions, bets, a research log, reports and a short run log in files the next agent can find. work-tracking applies the same idea when the question is broader: repo markdown, GitHub Issues, Linear, or memory.
#3: "Done" Meant a Doc, Not Live Proof
"Always take small, deliberate steps. The rate of feedback is your speed limit."
David Thomas & Andrew Hunt, The Pragmatic Programmer
The problem. A content plan is not done because the calendar looks nice. A technical fix is not done because code changed. A backlink is not live because it was submitted.
The fix is verification built into the loop: live URLs, rendered metadata, sitemap state, Search Console evidence, analytics proof, UI checks, or public links. Every action pushes toward proof, not paperwork.
#4: Five Lanes Open, Nothing Shipped
The problem. Growth agents love opening five lanes at once — audit, content, schema, backlinks, reporting — and finishing none of them. Unattended runs die when context resets. Incidents get "fixed" without a durable record.
The fix is one current focus ticket, clear done criteria, and skills for the heavy lifts:
seo-growth-workspace: one to three sized bets a week, each with a check date and a kill rule.shiploop — multi-phase shipping through GitHub issue ledgers, gated PRs, and worker adapters when the run must survive context loss.
Summary
Growth fundamentals matter more than ever: context before advice, memory between sessions, proof before "done," and one focus at a time. These skills condense that into repeatable operating loops for SEO work, plan tracking, autonomous shipping, and production incidents.
Reference
Invocation is declared in each skill. Crew is user-invoked: run /crew explicitly.
Growth
- seo-growth-workspace — Durable SEO operating workspace: bootstrap
.seo/, capture business context, audit evidence, prioritize backlog, implement one action, verify live, log handoff. Technical SEO, Search Console, schema, local SEO, content ops, internal links, backlinks, pSEO, conversion paths, monthly reporting.
- andy-post — Draft and publish X, LinkedIn, and Instagram posts through Andy MCP: raw idea → 2-3 drafts → operator picks → publish or schedule.
- posthog-growth-workspace — Durable product-data growth workspace on live PostHog data: bootstrap
.growth/, funnels/activation/retention, experiment and campaign registries, session-replay mining, HogQL cookbook + query runner, monthly growth reviews. Sibling of seo-growth-workspace — SEO stays there; install doctrine stays with your stack.
Agent Operations
- crew — Dispatch visible T3 Code workers across your configured providers, with verified model options, local reports and recoverable hand-backs. Includes the complete dispatcher.
- source-to-system — Turn external material into one evidence-backed, owned system improvement or an explicit no-change decision.
- grok-deep-research — Run isolated, bounded Grok research with native or Firecrawl search and produce a stable cited report plus an iteration ledger.
- t3code-local-model — Run a local open-weight model (for example Qwen2.5-Coder 7B) as a T3 Code provider through Pi and Ollama: a doctor script that checks whether a model fits the machine's memory now and each link of the chain, six checked steps, model sizing and troubleshooting.
Software Development
- mobile-monorepo-ios — Build, debug, and release Expo/React Native, native iOS, native Android, or bounded hybrid mobile work through explicit architecture, release-system, device, and distribution proof contracts.
- design-system-keeper — Extract, encode, and enforce a repo's real design system as an agent-readable canon (v0 Design Systems 2.0 shape): sourced rules, a primitive index, verify-against-source builds, and a drift backlog instead of memory-built lookalikes.
- cursor-subagent — Cursor sidecar delegation for explicit Cursor/Grok exploration, review, or isolated implementation.
- using-git-worktrees — Storage-aware development lifecycle: source-only task workspaces, bounded hydration, reusable local-main runtimes, isolated local state, automatic dehydrate, and reviewed retirement.
- work-tracking — Decide where multi-step work should live: repo markdown, GitHub Issues, Linear, memory, or a mix. When to promote tasks to issues and how future agents find current state.
- sync-github-fork — Confirm a GitHub fork's parent, measure divergence, and integrate upstream changes into its published default branch without rewriting fork history.
- file-pr — Open a pull request that ships safely: a title in the repo's convention, problem-first description with before/after evidence, a ship checklist driven by what the diff touches (env, migrations, flags, crons) with smoke and rollback, then review and worktree cleanup.
Productivity
- product-feedback-report — Create resumable founder-led feedback reports for web and mobile products, with conversational capture, screenshot/video evidence, automatic prioritization, and an always-current PDF.
- visual-plan — Turn ordinary text plans into rich interactive visual plans with diagrams, file maps, annotated code, open questions, and UI/prototype review when useful. Agent-Native visual planning, portable across hosted, self-hosted, and local-files modes.
DevOps
- shiploop — Turn a plan, fix, or feature into an unattended shipping run: GitHub issue ledger, dependency-gated phases, gated PRs, review evidence, optional worker adapters.
fixloop — deprecated 2026-07-05 (history): incident debugging now = Sentry evidence + STACK.md handles (Matias profile) + the diagnosing-bugs skill (github.com/mattpocock/skills) + per-repo AGENTS.md debugging notes.
Documentation and templates
See docs for reusable engineering guidance, including an AGENTS.md template you can adapt for your repository, and prompts you can reuse.
Maintainer Notes
cursor-subagent source lives in skills/software-development/cursor-subagent/. Update it there first, bump version: in SKILL.md, commit and push JorgeMenaDev/skills, then update consumers such as Matias with:
cd ~/.hermes/profiles/matias
npx skills@latest update cursor-subagent -p -y
Public page: https://www.skills.sh/jorgemenadev/skills/cursor-subagent
