Skills Became Infrastructure

In May I published a first library, twenty named capabilities, and a warning to stop breeding agent armies. Both still hold. What changed over the summer is not the thesis. It is the operating system. A skill is no longer a clever markdown file you keep next to Claude. It is a portable folder that Cursor, Codex, Copilot, Claude, and a long tail of runtimes all load. Procedural knowledge left the prompt and became infrastructure.

If you are still collecting "our best prompts" in Notion, you are a version behind the trench. If you dumped eighty skills into every repo and called it a platform, you are a version ahead of governance and a version behind how context actually works. This is the adult version.

A folder is the unit. A prompt is folklore.

The spec is almost insultingly small. A skill is a directory. It must contain SKILL.md with YAML frontmatter: name and a description that says what it does and when to use it. Optional: scripts/, references/, assets/. That is the whole trick.^[1]

Anthropic's engineering note is the part most CoEs skip: progressive disclosure is the design. Name and description load at startup. The body loads when relevant. Linked files load on demand. Context bundled into a skill is effectively unbounded if you split it. A 2,000-line SKILL.md is not thoroughness. It is an attack on the context window the rest of the session needs to think.^[2]

Practical rules the pioneers actually use:

• Keep the body under ~500 lines. Point at chapters. Do not paste the encyclopedia.
• Scripts are run, not read. Only the output spends tokens. If Claude is reading your Python to "understand" it, you authored the skill backwards.
• The model is already smart. Only add what it does not have: your exception, your forbidden phrase, your cited allowable, your golden lot.
• Degrees of freedom match fragility. Taste gets heuristics. Welds, signoffs, and regulators get scripts and tight tool allow-lists.
• The description is the trigger. A vague description is a silent skill. You will blame the model. It was your metadata.

Cursor's extras matter for operators: paths to scope a skill to files that match; compatibility with .agents/skills, .cursor/skills, .claude/skills, .codex/skills. Pick a canonical tree. I would pick .agents/skills and treat the rest as aliases, not three libraries.^[3]

Author from failure, not from a workshop

The authoring loop that works is the one Anthropic describes and that every serious operator already feels: run the agent on the real task, watch where it struggles, write the skill to close that gap, watch the trajectory, ask the model to capture the mistake into the folder.^[4] You do not invent a library in a kickoff. You harvest one from loops that already hurt.

That is why the May catalog is a starting menu, not a constitution. Teams that printed all twenty as empty shells got a graveyard. Teams that wrote three skills from last week's actual failures got a floor.

Distribution is product, not a CoE

Two working patterns, copied from how Claude Code's own team ships:

• Small team, few repos: commit skills next to the work. Review them like code. Roll them back like code.
• At scale: every skill in every repo taxes metadata at startup. Use a sandbox, let people opt in, promote to an internal marketplace when pull is real. Anthropic does not have a central committee deciding which skills exist. They promote what already got used. Remember that sentence when someone proposes a Skill Review Board.

Role-based bundles beat the 200-skill pack. A brand operator should not load firmware bring-up. A process engineer should not load your Reddit GTM skill. Metadata is a public good. Treat it like one.

Enterprise vetting is not optional. Scripts execute with environment access. MCP references extend reach. Network calls exfiltrate. Hardcoded keys are a gift to the incident channel. Audit the folder like you audit production software, sandbox the scripts, ban secrets in files, pin versions, checksum, rollback.^[5] Governance belongs in the path, the same way I want FDE to carry it rather than wait on it.

A first library that survives the bench

Software skills you already know. Hardware is the same object with gravity. Write these as folders with scoring scripts, not as chat recipes:

• Datasheet → pinout check, reads the schematic, fails on a mismatch, does not invent a pin.
• HDL module discipline, parameterized, tested, house style. Failures fail.
• FPGA / bare-metal bring-up, bitstream, JTAG, the board that will not boot.
• Firmware boot chain, measured boot, the boring path that actually starts.
• DfM against toolpath, draft, bend radius, residual stress. Geometry that cannot be cut is not a design.
• STEP import → audit, watertight, cited materials, no hallucinated allowables.
• FEA ladder, analytic check → beam → continuum. Each tier must agree with the one below or you are decorating.
• PLC / HMI test harness, the machine that runs, not the render.
• Twin vs live telemetry, firmware regressions and signal-integrity faults, fail loud.
• Signoff precheck, DRC/LVS, qualification packet, the document a regulator will read.

If a skill cannot fail CI, it is a style guide. Style guides do not keep people out of the incident review.

The failure modes (updated for summer)

Prompt museum. Notion, Confluence, a Slack canvas. Not files. Not versioned. Not loadable when the SKU dies.

The 2,000-line skill. You were afraid the model would miss a nuance. You spent the nuance budget on the table of contents.

Install our pack. Two hundred skills, one context window. Descriptions get truncated. Triggers go silent. You will say "skills don't work."

Vendor-only skills. Beautiful in one chat product. Invisible to the runtime on the plant network. Replay the kill switch in slow motion.

Uncited materials. The model is helpful. The alloy does not exist.

Skill Review Board. You reconstituted the product-org monolith and called it enablement.

Stand it up this month

Canonical directory. Three skills harvested from real failures, one office, one bench if you have a bench. Evals on those three. A sandbox for unofficial fourths. A named adult who can say yes to a PR in hours. No catalog of the future. No workshop that produces empty shells.

The library in May was a menu. This is the kitchen. If you only ship menus, you will keep eating out of the vendor's kitchen, and you already know how that Friday ends.

Skills became infrastructure when they became folders every runtime could load. Progressive disclosure, evals-first, organic promotion, portable paths. Write from failure. Split the file. Run the script. Cite the material. Leave the CoE out of the trigger.

Related: Build Skills, Not Agent Armies, The Skill Library, and Evals When the World Can Break.