Library
Search
Search the workspace and open a destination.
Library
Everything from the workshop, written to be useful afterwards rather than during. It's ordered as a path, so start at the top and each note assumes the one before it. Nothing here takes longer than eight minutes.
Start here
Read these in order. By the end you'll have a working skill installed and a realistic sense of what this is and isn't.
- I am just a text file
ruben.substack.com · Ruben Hassid on turning your voice, taste and opinions into a markdown file, and the interview method for building one.
- How to build a skill4m
Turn something you do the same way every time into a skill your whole team can use.
- Adding a skill to your workspace3m
Download a skill from the Skills page and install it, in about a minute.
- Which model should I use?4m
Start with Sonnet. The test for when a task has earned something more expensive, and why the reflex to reach for the top tier is costly.
- Your first 30 days3m
What to actually do, week by week, so this doesn't become the workshop you enjoyed and forgot.
Making the output good
A first draft from any model is roughly 70% of the way there. These are the moves that close the gap, and the ones worth building into every skill you write.
- Make it ask you questions first3m
One line that fixes more bad output than any amount of prompt rewriting.
- Reducing hallucinations with self-verification2m
A prompting technique for catching a model's own factual errors before they reach a client.
- A second reader for your output2m
Why having one model check another's work catches more than asking the same one twice.
- Writing that doesn't sound like AI3m
The specific tells people react to, and how to keep them out of everything you produce.
- Artifacts, when a list isn't enough2m
One word that turns a wall of bullet points into something you can actually look at.
Skills to agents
A skill runs when you ask. An agent runs on a schedule, calls tools, and hands off to other skills. This is the step between the two.
- What actually makes it an agent3m
A definition that survives contact with the hype, and the three things that separate an agent from a prompt.
- Chaining skills together2m
Why five small skills beat one big one, and how to make them hand off.
- Turning a skill into an agent3m
Scheduling, what changes when nobody's at the keyboard, and the one step people skip.
- Connectors, and what Claude can actually reach3m
What a connector is, why half of all broken skills are a connector problem, and how to find out in ten seconds.
- Sharing a skill with someone else2m
How to hand a skill to a colleague, the file-extension gotcha, and keeping track of versions.
Craft and context
For the half of the room already building. Context is the scarce resource, and most disappointing output is a context problem wearing a prompt problem's clothes.
- Context rot, and why long chats get worse3m
The single most useful thing to understand about why good output degrades.
- Global instructions vs skills2m
Where to put a rule so it fires when you want it and stays out of the way when you don't.
- Skill limits and official docs2m
The hard limits on a SKILL.md file — name, description, and body length — plus where to find Anthropic's own guidance.
Rolling out to a team
The leadership track. What this costs, who owns the output, and how skill libraries fail once more than one person depends on them.
- What this actually costs3m
How tokens work, where the money really goes, and which levers are worth pulling.
- Rolling out skills across a team3m
The path from one person's skill to a company standard, and the three ways it goes wrong.
- Who's accountable for the output3m
The short answer is you. The useful part is what that means in practice.
- Where teams get stuck3m
An honest map of what's hard after the easy wins, and what separates the pilots that work from the 95% that don't.
Playbooks by role
Each playbook names the specific time sink raised in the room, then gives three skills worth building against it, in the order they compound.
- Playbook: business development3m
RFPs, competitive positioning, and pipeline reporting. Three skills, in the order they compound.
- Playbook: producers and account leads2m
Turning conversations into schedules, and staying on top of what you committed to.
- Playbook: finance3m
Variance analysis, formatting that never varies, and the limit worth knowing before you start.
- Playbook: legal3m
Getting the queue down without ever putting judgment on autopilot.
- Playbook: creative and trend research3m
Staying current without reading everything, and building trend work that holds up.
Reference
Definitions and the questions that came up most, in the words they were asked in.
- Glossary2m
Plain-English definitions for the words that get used as though everyone already knows them.
- Questions from the room4m
The things people actually asked in the sessions, answered properly rather than in thirty seconds with a microphone.
- Prompt library
The 13 lines demoed live, with a copy button on each. Paste them into a chat or build them into skills you rerun.
Resources
Off-site, free, and better than anything we’d write ourselves. Every link was opened and checked. If one breaks, tell us and we’ll fix it.
Courses
Free, and structured better than anything we could write. Take them in this order if you're starting from zero.
- Anthropic Academy21 courses
Anthropic · The official catalogue. Start with Claude 101 and Introduction to Claude Cowork, then AI Fluency: Framework & Foundations if you want the non-technical grounding.
- Introduction to agent skills6 modules
Anthropic Academy · The direct sequel to the session you sat in. Six modules from writing your first SKILL.md through sharing skills across an enterprise.
- Agent Skills with Anthropic2h 19m
DeepLearning.AI · Ten video lessons taught by Anthropic's head of technical education. Slower and more thorough than the Academy version, and worth it if you learn by watching.
- Introduction to subagents
Anthropic Academy · The technique behind having one model check another's work. Search the catalogue for it, then pair it with our note on using a second reader.
Official documentation
When you need the exact rule rather than someone's summary of it.
- Agent Skills overview
Anthropic · What a skill is, how Claude decides to load one, and where skills live across Claude apps, Claude Code, and the API.
- Skill authoring best practices
Anthropic · The one to read after you've written three skills and started wondering why one of them never triggers. Description writing is most of the answer.
- Prompting best practices
Anthropic · The living reference for clarity, examples, XML structuring, role prompting, and prompt chaining. Everything a good skill body does, explained.
- Agent Skills specification
Open standard · Skills started at Anthropic and were released as an open format. Cursor, VS Code, GitHub Copilot, Gemini CLI and Codex all read the same SKILL.md, which is why what you learned here travels.
- Model Context Protocol
Open standard · The plumbing behind every connector. Read it when you want to know what Claude can and can't reach in Microsoft 365, Salesforce, or Figma.
Two essays worth more than most courses
Written by the people who build the models. Both are short, and both will change how you write your next skill.
- Building Effective AI Agents
Anthropic Engineering · Argues for simple composable patterns over frameworks, and draws the line between a workflow (a fixed path) and an agent (one that decides).
- Effective context engineering for AI agents
Anthropic Engineering · Why what you put in the window matters more than how you word it. This is the source material behind everything we've written on context rot.
The research behind the techniques
Two papers that come up whenever someone asks whether any of this actually works.
- Chain-of-Verification Reduces Hallucination in LLMs
Meta AI, 2023 · The four-step method behind our verification note. Skim the abstract and Figure 1; the rest is for the curious.
- The GenAI Divide: State of AI in Business
MIT Project NANDA · The study behind the '95% of pilots return nothing' figure. Read it for the reason why: the failures were integration and learning gaps, not model quality.
Working examples to steal from
Reading someone else's skill teaches faster than any explanation of skills.
- anthropics/skills
GitHub · Anthropic's own open skill library, plus a blank template and the format spec. The document skills (PowerPoint, Word, Excel) are the closest thing to what most of you need.
- Prompt engineering interactive tutorial
Anthropic · Nine chapters of before-and-after prompts you edit yourself. There's a Google Sheets version if a notebook isn't your thing.
Stuck?
Talk to the GPC team.