Saving a useful post is easy. Turning it into a reliable capability for Claude Code is harder: the post may omit prerequisites, compress several decisions into one sentence, or make claims that need another source. XSkills handles that gap as a supervised compilation process rather than treating every bookmark as ready-to-run instructions.
What becomes a skill
A bookmark is a good candidate when it contains a repeatable workflow, names a concrete outcome, and provides enough evidence to reconstruct the steps safely. A product announcement, opinion, news item, or isolated tip is normally not enough. XSkills skips those instead of manufacturing a process the source did not support.
- A sequence another person could repeat.
- Inputs, constraints, or tools that can be identified.
- An outcome that can be checked rather than merely admired.
- Claims that can be traced to the post or supporting sources.
The bookmark-to-skill workflow
- Discover. XSkills reads newly discovered X bookmarks through read-only access. It cannot post, reply, or alter the account.
- Qualify. The bookmark is screened for a reusable AI or developer workflow. Low-signal saves stop here.
- Research. Relevant linked material is inspected and the evidence for each instruction is recorded. Missing requirements are called out instead of guessed.
- Compile and grade. The supported process becomes a Claude Code package with triggers, steps, guardrails, references, and evaluation queries.
- Review and install. You approve an exact checksum in Telegram or the dashboard. The macOS connector verifies every file before writing it to
~/.claude/skills.
What the Claude Code package contains
Claude Code skills are filesystem artifacts centered on a SKILL.md file. Claude discovers skill metadata and loads the full content when the request matches. The official Claude Code guide describes the same file-based model and user/project skill locations. Read the Claude Code skills documentation.
multi-agent-review/
├── SKILL.md
├── manifest.json
├── eval_queries.json
└── references/
├── source.md
└── research.mdSKILL.md defines when the skill should be used, the workflow, expected outputs, guardrails, and recovery behavior. The references keep the source and research attached. Evaluation queries record realistic prompts that should and should not activate the skill. The manifest binds the target harness, install root, and file hashes into one package.
A worked example
Imagine bookmarking a post that outlines a multi-agent review pattern: one planner, specialist reviewers, and a final judge using a shared rubric. The useful idea is not the post's engagement count or headline. It is the repeatable coordination pattern.
XSkills separates the pattern from unsupported performance claims, checks linked sources for the missing rubric and handoff details, and compiles only the instructions the evidence supports. The resulting skill can tell Claude Code when to use the pattern, what each role receives, how reviews are reconciled, and when the workflow should stop. You still decide whether that exact package belongs on your machine.
Why approval is tied to a checksum
Approval is not a reusable permission for a skill name. It applies to the exact package you reviewed. If any instruction or file changes, the checksum changes and the revision needs a new approval. This prevents a later edit from silently inheriting trust granted to an earlier version.
See the full XSkills security model, learn what makes an agent skill different from a saved prompt, review the setup documentation, or compare the available plans.