You need to be skillsmaxxing (10x your Claude/Codex)

Remy argues that agent skills are the single most important concept for saving time with AI on repeatable processes, defining them as SOPs packaged as markdown files that Claude reads to execute a process consistently. Without skills, Claude forgets user preferences between sessions and requires re-explanation each time, which Remy identifies as the root cause of slop output from AI. He estimates a single well-built skill can save two hours a week, and stacking multiple skills compounds that benefit significantly.

Remy follows a thin agents, thick skills philosophy, keeping agent instruction files lean while embedding all context and step-by-step instructions inside the skill files themselves so any harness can execute them without additional guidance. Skill files contain a name and description at the top acting as a spine label, with full instructions inside that the agent reads when needed. He has built skills covering proposal generation that one-shots a branded proposal, a Notion formatting skill enforcing consistent white space and blue highlights, a brand voice skill for email and landing page copy, and an email formatting skill for Resend covering layout, buttons, signature, and logo. He structures repeatable multi-step workflows as skill chains rather than monolithic skills when sub-processes will sometimes need to run independently, citing a YouTube publish workflow that triggers title, thumbnail, and description skills separately or in sequence.

Remy has added a self-improvement loop to the bottom of all skill markdown files that prompts Claude to review failures, user corrections, and discoveries after every skill run, proposing a change only when significant enough and requiring user approval before the skill is updated. He notes this addition is relatively new and may not yet be present in every skill file.

The current state of skills sharing is described as analogous to the Microsoft Word era before Google Docs, with no central source of truth. Sharing skills via dot skill or zip files on Slack or email creates version control problems because changes do not sync and teams end up with multiple divergent copies. Google Drive, Dropbox, and Obsidian each require workarounds because Claude only reads from its own dot claude slash skills folder, and Obsidian additionally requires every team member to install it.

The recommended method for sharing skills across a team is storing them in a GitHub repository and exposing them as a plugin, where a plugin in Claude Code is defined as a bundle of skills, MCPs, and hooks, though any combination or subset is valid. Skills in the repository are organized by department so team members install only the plugins relevant to their role. Installing a plugin requires only entering the GitHub repository URL in the slash plugin marketplace menu, and auto update must be enabled so that skill changes made by one person propagate automatically to everyone else. The plugin system has been confirmed to work in both Claude Code and Codex, and GitHub provides version control so teams can revert to older skill versions if something breaks. A practical example of the shared system working is that a team member named Harold encountered an image upload error in Resend, found a workaround, updated the email formatting skill, and the fix propagated automatically to all team members.

Rami lost his dot Claude folder containing 150 skills representing over 500 hours of work with no backups approximately two months ago, which he cites as a direct argument for cloud-based storage. Skills built by employees on their own machines can also be lost to the company if those employees leave, making a centralized GitHub organization repository a solution for ownership retention. On Claude enterprise accounts, organization-level plugins can be configured by a technical founder or CTO and baked into all team member accounts, removing the need for non-technical employees to use the terminal.

Rami has built a web app UI on top of the team skills GitHub repository so every skill displayed pulls live from that repository, a skills assistant that allows querying the private library in natural language, and a usage tracking feature added that day that hooks into Claude and tallies every skill use broken down by team member. Usage data could identify skills unused after six months and remove them to prevent the library from becoming bloated. Rami maintains a separate personal repository for individual skills like inbox triage and morning brief that are too user-specific to share. Both hosts predict that teams effectively managing and distributing skills will have an unfair advantage, and that skills will eventually represent significant enterprise value in the same way SOPs train employees to perform economically valuable tasks.