TL;DRskills is the mechanism for packaging 'how I run this class of task' into a reusable unit: AGENTS.md records the project's rules, a skill records your playbook. The signal from that high-scoring HN thread is that OpenAI treats this as a long-term direction — which makes now a good time to start distilling your recurring workflows into skills.
Sources SOURCES
Distilled from public community discussions. This page only breaks them down and verifies — it never reposts the original; opinions belong to their authors.
Facts FACTS
Check these claims against the listed sources. For allowance mechanics, first-party account displays outrank community summaries.
- From December 2025, the skills mechanism rolled into ChatGPT and the Codex CLI (the official move behind the 587-pt HN discussion).
- In April 2026, OpenAI announced the plugin ecosystem had expanded to 90+ entries (skills, app integrations, and MCP combos); skills is one of the building blocks of that plugin stack.
Concepts CONCEPTS
Not everything said in a community thread is a fact — each concept carries its nature label. Read the label first.
A division of labor you can actually remember: AGENTS.md is 'the rules of this project' and travels with the repo; a skill is 'how I do this class of task' and travels with your account or team. One governs constraints, the other governs playbooks. The boundary test: would this instruction still be true in a different repo? Yes → skill. No → AGENTS.md.
A judgment the HN thread broadly endorsed: the third time you find yourself typing the same briefing for the same kind of task, it should become a skill. Three repeats is the codification signal. Below three you are guessing at the pattern; above it you have already lost time to retyping. And a skill nobody has invoked in a month is a deletion candidate.
Past a certain count, skills start generating selection noise: the agent's matching of skills to tasks is not always precise. A small set of high-frequency skills beats an exhaustive, cover-everything library. If matching degrades as the library grows, the cure is pruning by usage frequency, not writing better descriptions for unused entries.
Takeaways TAKEAWAYS
How skills relate to knowledge you already hold: procedural knowledge — test commands, review checklists, deploy steps — is prime skill material. Constraint knowledge — business rules, architectural no-gos — belongs in AGENTS.md. Mixing the two means every clone carries your personal habits, and every habit of yours becomes repo law. When each lives where it belongs, both stay short enough to actually get read.
A suggested on-ramp: do not write a skill yet. Collect repeated phrasings first. Spend one week logging the instructions you keep retyping, then in week two codify the single most frequent one. Skills built from real repetition get used; skills built from imagination do not. A week of honest logging also protects you from the classic failure mode: building skills for the workflow you wish you had instead of the one you actually run. Frequency beats elegance as the selection criterion — the ugly instruction you run daily is worth more than the beautiful one you run quarterly.
An ecosystem view: among the 90+ plugins announced in April 2026 there are plenty of official and third-party skills. For generic needs (code-style checks, say) look for an existing one before writing your own; reserve custom skills for proprietary flows. Before writing anything custom, search the catalog by the verb you keep typing — coverage of generic tasks is better than most people expect. The same search-first logic applies to revisions: before extending your custom skill, check whether a catalog update already closed the gap.
A cold read: the word 'quietly' in that HN title deserves a second look — the skills mechanism is a platform-level bet. People who distill early will carry a methodology into whatever tool comes next, not just a pile of config. Config rots when platforms change; a written-down method survives the migration.
Keep reading NEXT
Want today's verdict? Back to the live radar.