The repos climbing GitHub this month tell you where agent engineering went: skills, not crews. Anthropic's public anthropics/skills repository, the community awesome-agent-skills collection with more than a thousand entries, NVIDIA's skills pack for robotics and CUDA workflows, and community libraries carrying 380-plus production skills all share one shape. Each skill is a folder with a SKILL.md that a coding agent loads when the task matches. No orchestrator personas. No twelve-box crew diagram. The market voted, and it voted for reusable competence over rented attention.
Short answer: a skill is a written procedure with a trigger, inputs, forbidden actions and a verification step. One strong agent plus a shelf of skills beats a crew of specialists on cost, on consistency and on debuggability, for the reasons the Nature Machine Intelligence coordination study quantified in July. This piece gives you the decision table, the skill anatomy we ship, a daily ritual loop and the cost cap that keeps it honest.
Skill versus agent: the decision table
| You want | Write a skill when | Spin up an agent when |
|---|
| Consistent output format | Always. Formats are rules, not personalities. | Never. |
| Domain procedure (deploy, publish, audit) | The steps are known and repeatable. | The steps require a different tool set or model than the main agent has. |
| Quality check | The check can be run as a script or checklist inside the same run. | The checker must not see the maker's reasoning (audit wall). |
| Parallel throughput | Units are few and sequential is fine. | Many independent, read-heavy units that do not need to agree. |
| Memory of project facts | Always. A lock file, read at start. | Never. Memory agents are the most expensive way to store a paragraph. |
The right column is short on purpose. It matches the three tests in the multi-agent tax: different capability, audit wall, real parallelism. Everything else is a skill.
Anatomy of a skill file
A skill that survives contact with a real project has five parts. Frontmatter with a name and a one-line description the agent matches against. A trigger section that says when to load it and, just as important, when not to. Inputs it expects. Forbidden actions. A verification block that produces evidence, not a feeling.
---
name: publish-blog-post
description: Add a post module, register it in the barrel, validate SEO fields, and verify the route renders. Use when asked to publish or add a blog post.
---
# Publish blog post
## Trigger
Use when the task creates or edits a post under src/data/blog-posts/.
Do NOT use for WordPress-resident posts (see wp-post-edit skill).
## Inputs
- slug, title, meta_description, content (HTML), category, tags
- published_at in YYYY-MM-DD
## Procedure
1. Create src/data/blog-posts/-.ts exporting a BlogPost[]
2. Import it in src/data/blog-posts.ts, add slug to POST_ORDER (top), spread into posts
3. Run the validator: node scripts/validate-post.mjs
## Forbidden
- Never edit next.config.ts redirects; use src/data/blog-redirects.ts
- Never set author to a personal name; author is the brand
- Never publish under 1,500 words in STANDARD mode
## Verification (must all pass)
- validator exits 0: meta 120–150 chars, no apostrophes, title 50–65 chars
- npx tsc --noEmit reports 0 errors
- curl -s localhost:3002/blogs/ | grep -c 'application/ld+json' >= 2
Notice what is not in it: no persona, no "you are an expert editor." The agent already has a model. What it lacks is the procedure, the guardrails and the definition of done. A skill supplies exactly those three things, and nothing else.
Where skills live and how they load
Claude Code reads project skills from .claude/skills/<name>/SKILL.md and personal skills from ~/.claude/skills. Codex, Gemini CLI and Cursor now read the same folder shape through the community libraries, which is why the big repositories advertise compatibility with eight or nine agents at once. Keep project skills in the repo so they version with the code, and keep the shelf small: a dozen skills that are each loaded weekly beat a hundred that nobody remembers. The /skill-doctor report covered in how teams manage skills files tells you which ones have drifted from the code they describe.
The project bible and the consistency lock
Skills are procedures. They still need facts. Two files carry them.
Project Bible (
CLAUDE.mdorAGENTS.md): what this project is, the hard rules that came from real incidents, the paths that matter, the commands that are the only sanctioned way to do a thing.Consistency Lock: the immutable vocabulary. Brand name, voice rules, banned phrases, output formats, colour tokens. Read at the start of every run and referenced by skills rather than duplicated inside them.
Between them, a single agent has more usable context than a crew where each specialist carries a partial copy of the same facts and they disagree at turn nine.
The daily ritual loop
A Creator OS is a loop, not a launch. Ours has four beats and each beat is a skill.
Plan (
plan-day): read the bible, read the queue, produce a ranked list with a token budget per item.Make (
make-<artefact>): one skill per artefact type. Blog post, product listing, video shot list, release metadata.Gate (
qa-gate): the one place an audit wall exists. Reads artefact and spec only, returns JSON with pass/fail per check and a reject class.Ship (
ship-<channel>): commit, deploy, purge, verify the live body, log the cost.
The gate is allowed to be a second agent because it passes the audit-wall test. Nothing else in the loop is.
Cost cap per ritual
Every skill declares a token ceiling in its frontmatter, and the plan skill sums them before the day starts. If the sum exceeds the day's budget, items get cut from the bottom of the ranked list, not squeezed. We size the ceilings with the token counter on the actual bible and lock files, because the fixed cost of reading context is the part people forget. A skill that costs 4,000 tokens to load and saves ten minutes is a bargain; a skill that costs 40,000 because it inlines a style guide the lock already holds is a leak.
Promotion path: skill to product
A skill that runs cleanly for a month is a product. The publish-blog-post skill above became the production rules in the Claude Code Production Pack, which ships the 26 agents, 14 skills and the rule files that run a live 4,000-product store. The bible-first pattern is packaged as the CLAUDE.md production rules, and if you are starting from nothing, the CLAUDE.md starter templates give you the bible and lock skeletons. For writing the gate prompt itself, our free Gauntlet Prompt Generator produces the criteria table the QA skill checks against.
What the trending repos get wrong
Volume. A library of a thousand skills is a search problem, and an agent that matches descriptions loosely will load the wrong one. Curate hard: pull in the five you need, rewrite the trigger sections for your project, delete the personas some authors sneak into the procedure, and add the verification block if it is missing. Most published skills stop at the procedure and never define done. A skill without verification is a suggestion.
Three skills to write this week
Do not start with a library. Start with the three procedures you repeat most and resent most. For a solo creator or a small product team these are almost always the same three.
ship-checklist. Trigger: any task that ends in a deploy or a publish. Inputs: the artefact path and the target environment. Procedure: the exact commands in the exact order, including the verification curl against the live URL. Forbidden: bypassing the pre-push hook, force-pushing, skipping the cache purge. Verification: the live body contains the new marker and the deploy run reports success. This skill alone removes the class of incident where the build passed and the page still served the old version.write-in-voice. Trigger: any public copy. Inputs: the audience and the artefact type. Procedure: read the Consistency Lock, draft, then run the banned-phrase scan and the reading-level check. Forbidden: the phrase list, personal names where the brand belongs, invented statistics. Verification: the scan returns zero hits and every number has a source line. Pair it with the editing workflow in writing with AI without the LLM tells.qa-gate. Trigger: invoked by the other two, never by a human directly. Inputs: the artefact and the spec only, with the transcript withheld. Procedure: run each criterion, record evidence, return JSON. Forbidden: rewriting the artefact, suggesting improvements, conversing. Verification: the JSON validates against the gate schema and every check has an evidence string.
Write these three, run them for a week, and count the turns your main agent no longer spends explaining itself. Then, and only then, pull a fourth from the public libraries and rewrite its trigger and verification blocks to match. A shelf built one proven skill at a time stays small enough that the agent loads the right one every time, which is the whole point.
Quick answers
What is the difference between an agent skill and an agent?
A skill is a written procedure with a trigger, inputs, forbidden actions and a verification step that a single agent loads on demand. An agent is a separate model instance with its own context and cost per turn.
Which skill repositories are worth starting from?
Anthropic's official skills repository for the reference shape, awesome-agent-skills for breadth, and vendor packs such as NVIDIA's for domain procedures. Curate down to a dozen and rewrite triggers for your project.
Do skills work outside Claude Code?
The SKILL.md folder convention is now read by Codex, Gemini CLI, Cursor and others through the community libraries. Keep skills in the repo so they version with the code.
When is a second agent justified?
When it needs a different tool set or model, when it must not see the maker's reasoning, or when many independent read-heavy units run in parallel. Otherwise write a skill.
How do I keep skill costs down?
Declare a token ceiling per skill, reference the lock file instead of inlining style guides, and sum ceilings in the plan step before the day starts.
Twelve skills, one agent, one gate. Every product mentioned is available at wowhow.cloud — pay once, ship forever.
Originally published at wowhow.cloud
Top comments (0)