A Practical Guide to Building Portable Agent Skills
- Source
- x.com
- Date
Agent skills let you capture repeatable instructions so you and your agents can reuse them. A practical guide on agent skills to keep past instructions, institutional knowledge, and other context 👇 https://t.co/G2A4KdnTOo
Context
An is a folder centered on a SKILL.md file that tells an when and how to do a repeatable task. Anthropic published the format as an open standard in December 2025, and by October more than 70 agent platforms, including GitHub Copilot and Cursor, had integrated it (our coverage). A guide on GitHub's blog by Andrea Griffiths, based on a GitHub Universe 2026 session she gave with Shishir Tewari of Procore, argues that the same instructions carry over between agents most easily when project details and environment setup are kept out of them. Its starting point is a limit: a skill can describe how to query a database, but it does not grant access to that database. The tools, credentials and permissions have to come from the environment the skill runs in.
The guide's main example, Data Forge, splits a data-pipeline workflow into several skills that share instead of each rediscovering it. One, forge-init, inspects the project and writes what it finds, such as source tables and naming conventions, into a manifest file, and the other skills read that manifest instead of hardcoding a project's names, database or conventions. Another, forge-dbx, is the single route the skills use to reach the data platform, choosing a Databricks or DuckDB adapter based on the manifest. The authors' test for portability is whether another team could use the skills with its own manifest without rewriting them. Because supporting several agent clients can mean keeping several copies in sync, the example keeps one SKILL.md per skill as the source and uses a packaging script to generate the plugin and agent files that Claude Code, Cursor and GitHub Copilot read.
The guide is equally clear about what instructions should not be trusted to do. A skill can tell an agent which standards to follow and when to ask for approval, but a check that must block a merge belongs in the repository's CI checks and rules. Its second example, Proof Agent, rests on the rule that self-verification is not verification: a separate verifier reviews a code change without seeing the coding agent's own account of it and must cite evidence in the changed files for a PASS, FAIL or PARTIAL verdict. The GitHub Action behind it reviews the diff without running code or tests, and in the demo its findings were advisory. The authors also say skills suit bounded, repeatable work such as a release or a migration, and that work needing shared state, an audit trail or permissions enforced outside the model is better built as an application. Since clients differ in models, tools and permissions, they recommend checking after installation that each client picks the skill for the intended task, loads its references, can use its tools and returns the evidence the skill asks for.
- agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
Checking sign-in…
Loading comments…




