Standard

Skills are system instruction files that tell your AI agent exactly how to work with xysq. Install a skill once and your agent knows how to authenticate, what to capture, when to recall, and how to use tags, without you explaining it every session.

​

Why skills matter

Modern coding agents (Claude Code, Cursor, Codex, Windsurf, Copilot CLI, Gemini CLI) all load instruction files at session start. Anthropic formalised this pattern as Agent Skills: portable, model-readable instruction packs that an agent loads on demand to acquire a specific capability. Without a skill, you re-explain context every session: who you are, what you’re working on, which tools you prefer. With a skill, the agent starts informed and stops asking. Less re-explanation means fewer tokens, which means faster, sharper output. Further reading from Anthropic:

• Introducing Agent Skills - the engineering rationale
• Skills overview - the official spec
• Equipping agents for the real world with Agent Skills - design principles

​

What the skill is

The xysq skill is a hand-crafted, version-controlled instruction pack. It covers the core protocol: session start, aggressive capture, surgical recall, source handling, and consent rules. It works from the moment you install it, with no memories required and nothing to generate. Every connected agent gets the same skill. xysq maintains it; you just download the current version and drop it in. On Claude Code, the standard skill is joined by a set of purpose-built recall skills (recap, decisions, actionables, blockers, prep) that tune recall for specific questions like “catch me up” or “what am I blocked on”.

​

Installing your skill

The download flow below is for agents that load a skill/rules file directly (Cursor, Windsurf, Codex, Claude Desktop, etc.). For Claude Code, prefer the plugin (see the note above).

​

Step 1 - Download from the dashboard

Go to app.xysq.ai and open the Dashboard. The Get Started panel has a Download Skill button. Click it and a dialog opens with download options for every supported platform.

​

Step 2 - Choose your platform

The Install guides page covers every supported agent: Claude Code, Claude Desktop, Claude.ai (web), Cursor, Windsurf, Codex CLI, Gemini CLI, Antigravity, and GitHub Copilot CLI. Jump to your agent from the page’s right-side nav. Only Claude products consume Anthropic’s Agent Skills format natively. For other agents, xysq emits the same content in the rules-file or system-prompt format that tool expects. The instructions are identical, only the wrapper differs.

​

Step 2b - Attaching the skill in Claude (web / desktop)

The video below shows how to paste your skill into Project Instructions in Claude.ai. The same steps apply to Claude Desktop.

​

Step 3 - Verify it’s active

On your first substantive message in a new session, the agent should call memory_recall(query=<your message>, budget="low") to pull relevant context before answering. That’s the skill working. authenticate() runs lazily, only on the first write (memory_retain, memory_delete), not at session start. memory_reflect is reserved for synthesis questions (“what do I prefer about X”, “summarise my stance on Y”), not warmup.

​

Keeping your skill current

The skill is maintained centrally. When xysq updates the template, re-download it from the Get Started panel and overwrite your installed file. Your agent picks up the new version at the start of its next session. There’s no per-user generation and nothing to regenerate.

​

The north star

Every skill is written around a single principle: fewer tokens, same or better output. Agents that use xysq effectively don’t need you to re-explain your context at the start of every session. They don’t ask what project you’re working on. They don’t forget your preferences. They start informed. The skill teaches your agent to capture aggressively, every decision, correction, preference, and fact, so that future sessions cost fewer tokens to reach the same quality of output. The memory you build today is tokens saved tomorrow.

​

What the skill instructs

Session start protocol (MCP platforms: Claude Code, Claude Desktop, Cursor, Windsurf, generic) On your first substantive message (not greetings):

1. Call memory_recall(query=<your message shaped as a lookup>, budget="low") - pulls the relevant memories for your message. The results become task-specific context before the agent answers.
2. Call authenticate() lazily, on the first write (memory_retain, memory_delete), not at session start.
3. Call memory_tags() just before the first memory_retain.

Capture rules Recall vs reflect: pick one, never both for the same question.

• memory_recall(query, budget="low") - default. Targeted lookup of facts and recent memories. Used on every substantive turn.
• memory_recall(query, budget="mid"/"high") - broader pulls, used sparingly.
• memory_reflect(query) - synthesis only. Reserved for questions that require reasoning across memory: “what do I prefer about X”, “summarise my stance on Y”, “compare my past decisions on Z”. Not for warmup.