§1 — Your first question
The question that proves your skill is working.
Once you have installed your skill in an AI, the first thing to do is ask a question that a generic AI could not answer well from general knowledge alone. This is the proof-of-work moment — the point where you see the difference between an AI answering from Wikipedia and an AI answering from your specific source.
The question does not need to be complex. It needs to be specific to your source. Ask about a framework that is yours, a term you defined, a decision rule from your methodology, or a claim that only appears in your document. If the AI answers with precision and cites the relevant section, your skill is loaded and working. If it gives a generic answer, check that you invoked the skill correctly before the question — see the install guide for the exact invocation step per platform.
Five prompt patterns that consistently reveal whether a skill is grounded correctly:
"What does [source title] say about [topic] — with page references?"
Use this when your source has a distinctive position on something. The answer should include a citation — a page number, chapter reference, or direct quote from your document. If there is no citation, the skill is not grounded yet. Check your invocation.
"Rewrite this in our voice and list which rules you applied."
Paste a piece of copy or writing. For a brand voice skill, this is the defining test. The AI should rewrite it using your specific vocabulary, register, and patterns — and tell you which rules from your guide it applied and where those rules appear in the source. Generic rewrites that improve the prose but apply no specific rules mean the skill is not loaded.
"Qualify this lead the way we do it."
For a sales playbook or methodology skill. Paste a brief description of a prospect or situation and ask the AI to run it through your framework. The response should reference your specific qualification criteria, ICP definition, or scoring method by name — not generic frameworks like BANT from general training. Specific output = skill is working.
"[Ask the same question in two separate sessions and compare the answers]"
Ask the same question twice — once today, once tomorrow. If your skill is loaded correctly, both answers should be grounded in the same passages from your source and reach the same conclusions. One of the most valuable properties of a skill is the same answer every time. If the answers diverge significantly, there may be an issue with how the skill is being invoked.
"What does my source NOT cover about [topic]?"
A well-built skill knows its own edges. Ask the AI what your source does not address on a given topic. The answer should be honest and specific — pointing to real gaps in your document rather than giving a generic "this topic is complex" response. This prompt is also useful for identifying what to add to your next skill version.
Most people describe the same experience: they ask a question they already know the answer to — one specific to their own work — and the AI answers with precision they did not expect. It cites the page, uses their vocabulary, and applies their framework exactly. That is the moment the skill is real. It is usually the third or fourth question of the first session.
§2 — How loading works
What actually happens when the AI loads your skill.
Understanding the mechanism helps you use the skill well — and diagnose it when something seems off.
A skill is a folder of plain text files in the SKILL.md format. When you invoke it — by
typing /your-skill-name in Claude Code, or by starting a conversation in a
Claude Project, or by opening your Custom GPT in ChatGPT — the AI reads the
SKILL.md index file first. This file is short. It contains the primary
frameworks, the key vocabulary, the use-when triggers, and a navigation map pointing to
the rest of the pack.
When you ask a question, the AI uses the index to decide which chapter from the
chapters/ folder is most relevant to that question. It then opens only that
chapter — not the entire skill — before generating its answer. This is why answers stay
fast even for large skills: the AI is not loading everything at once, it is loading the
right section on demand.
A skill does not modify the AI model. It does not fine-tune the model on your data and it does not train any AI on your content. The model itself is unchanged. What changes is the context the model reads before responding to your question. Your source file is deleted after conversion — typically within minutes, never longer than 7 days. The skill files that remain are documents the model reads as context, the same way it would read a file you pasted into a chat window — except they are structured, indexed, and loaded automatically every time you invoke the skill.
This architecture has two important practical implications. First, you need to invoke the skill at the start of each session — the AI does not carry context between sessions automatically unless you are using a persistent setup like a Claude Project or a saved Custom GPT. Second, the skill only knows what your source contained. If your document did not cover a topic, the skill will tell you so honestly rather than making something up — which is the correct behaviour.
The chapters file contains your knowledge broken into topic clusters. The glossary.md file locks in your specific vocabulary — so the AI uses your definitions of key terms instead of generic ones from general training. The cheatsheet.md file holds quick-access rules and decision tables that the AI reaches for in high-frequency use cases. The patterns.md file contains the recurring structures and moves that characterise how you actually work.
Together these files give the AI a coherent, structured briefing that it can navigate precisely. This is the difference between "add your docs" tools that dump everything into a context window and hope, and a skill that the AI knows how to use.
§3 — Daily use per platform
How the skill lives in each AI you use it with.
Each platform handles the skill slightly differently. Here is what day-to-day use looks like across the four supported platforms. For first-time install steps, see the install guide.
Your skill lives as a Custom GPT in your ChatGPT account. Once built, you access it from Explore GPTs > My GPTs, or by pinning it to your sidebar for faster access. Each conversation you start within the Custom GPT begins with the skill already loaded — you do not need to re-invoke it each time.
To pin for fast access: from the ChatGPT sidebar, click the three dots next to the GPT name and select Pin. It will appear near the top of your sidebar on every subsequent visit.
Knowledge file limit: Custom GPTs accept up to 20 files in the knowledge section. A standard skill pack contains 5 files and well within this limit. If you add additional reference documents to the GPT knowledge section alongside the skill, they will all be available as context.
Requires ChatGPT Plus, Team, or Enterprise for Custom GPT creation. On free accounts you can upload the skill files to a ChatGPT Project instead — the context persists within the project but without the system prompt from instructions.txt, so citation behaviour may be less precise.
In Claude Code (the desktop app), your skill lives in your ~/.claude/skills/ directory on macOS and Linux, or C:\Users\[you]\.claude\skills\ on Windows. Claude Code auto-discovers all skills in that folder at startup.
To invoke a skill, type /your-skill-slug at the start of a new
conversation. For example, if your skill is named brand-voice-acme,
you type /brand-voice-acme. Claude loads the SKILL.md index and
confirms the skill is active, then you continue with your question.
The slash-command is the invocation ritual. If you start a session without it, Claude does not load your skill — it answers from general knowledge. Make typing the slash-command the first thing you do when opening a session where you need your skill.
You can have multiple skills installed simultaneously. Claude Code shows all
available skills when you type / and waits for autocomplete. If you
have many skills, name them clearly so autocomplete is unambiguous.
Claude.ai Projects give you a persistent knowledge base that Claude references across every conversation you start within the project. Upload all files from your skill pack to the project's knowledge section, and Claude will use them as context automatically — you do not need to type a slash-command.
To access your projects: at claude.ai, click Projects in the left sidebar. Each project you create can have its own set of knowledge files. A common setup is one project per skill, named after the skill topic.
Context window behaviour: Claude.ai loads the project knowledge into the conversation context. For very large skill packs, Claude may not read every file in detail on every turn — it retrieves the most relevant sections. Start sessions with a specific question rather than a vague opener to help Claude retrieve the right chapter immediately.
Claude.ai Projects are available on all Claude.ai plans including free. For heavier professional use — especially if you need to use the skill in a terminal or editor workflow — Claude Code is more precise because the slash-command invocation is explicit.
Gemini Gems are Google's version of Custom GPTs — a named, persistent AI assistant with your knowledge files and system instructions attached. Your skill lives as a Gem at gemini.google.com > Gems. Once created, you access it from the Gems section of your sidebar.
Requires Gemini Advanced ($20/mo on a Google One AI Premium plan). Without Gemini Advanced, you can still upload the skill files to a standard Gemini conversation, but the context does not persist between sessions.
Daily use: open the Gem and start a conversation — the skill is active immediately. Gemini loads the knowledge files you uploaded during setup as its context. Ask your question directly; there is no invocation command.
Updating a Gem: when you build a new version of your skill and want to update the Gem, go back to the Gem editor, delete the old skill files from the knowledge section, and upload the new files from the updated pack. Save the Gem. The old knowledge is replaced immediately.
§4 — Updating your skill
Your method evolves. Your skill should too.
A skill is a snapshot of your source at the time it was built. If your source evolves — a revised methodology, a new edition of the book, an updated brand guide, expanded research notes — the skill will not automatically reflect those changes. You need to build a new version.
The process is the same as the original build: upload the updated source, go through the confirm step, and wait for delivery. The new skill is independent of the old one — your previous skill stays in your account until you explicitly delete it.
When to rebuild versus when to edit
Skills are plain text .md files. If you notice a small issue — a term defined slightly wrong, a rule in the cheatsheet that does not quite match your current practice, a chapter that missed a nuance — you can edit the skill files directly in any text editor. This is faster than rebuilding and costs no credits.
Edit the files directly when: the change is small and contained to one section. Find the relevant file (SKILL.md for index updates, glossary.md for term corrections, cheatsheet.md for rule changes) and make the edit. Save and re-upload the edited file to your platform (replace the old version in your Custom GPT knowledge section, overwrite the file in your Claude skills folder, or re-upload to the Gemini Gem).
Rebuild from source when: the source has changed substantially — more than a few pages revised, a new chapter added, a framework meaningfully updated. The pipeline will re-derive the chapter structure, update the glossary entries, and re-run the QA pass across the new source. This produces a more coherent result than piecemeal edits across many files.
When you name a skill at the confirm step, append a version number: brand-voice-acme-v1, brand-voice-acme-v2. The SKILL.md file also has a version field in the frontmatter — fill it in with v1.0 on your first build and increment it on rebuilds. This makes it easy to track which version is installed in which platform, and to roll back if a new build does not perform as well as the previous one.
After you build a new version and install it, use the gap-check prompt (from §1) to verify that the new version covers anything you wanted the previous one to improve on. Keep a short note of questions where the previous version fell short — that list becomes your test suite for the next build.
§5 — Sharing
Private by default. Share deliberately.
Skills are private by default. They live in your account and are accessible only to you. No other user can see your skill files, access the content of your build, or find your skills through any search or discovery feature.
Shareable skill links are a Plus+ feature. On Plus and above, you can generate a read-only share link for a skill. The link gives the recipient access to download the skill pack — the same files you received after the build. They install it in their own AI and use it independently of your account.
What sharing does and does not do
Sharing gives the recipient the skill files — the packaged output of the build. It does not give them access to your original source document (which was deleted after conversion). It does not give them access to your account, your credit balance, or any other skills. Sharing a skill is equivalent to sending someone a .zip file with the SKILL.md pack inside — they get the structured knowledge, not the source.
If your skill was built from proprietary client work, internal methodology, or confidential source material, think carefully before sharing. The skill files contain a structured, readable representation of that knowledge. Share only skills built from material you have the rights to share, and only with people who should have access to that knowledge under your professional obligations. When in doubt, keep the skill private and install it only in your own AI environment.
The share link is revocable. You can deactivate it from your skill page at any time. Recipients who have already downloaded the files keep what they downloaded, but the link stops working for new downloads.
For team use — multiple people in an organisation using the same skill — the recommended approach on Starter is for each person to install their own copy from a shared download. Team-level skill management with access controls is a roadmap feature.
§6 — Five power habits
The habits that get the most out of a skill over months of use.
A skill that sits installed but rarely invoked does not help. The habits below are the difference between a skill that pays for itself every week and one that you forgot you built.
Start every relevant session by invoking the skill.
Make it reflexive. If you are opening a session to work on anything that touches the domain your skill covers — brand copy, sales calls, research notes, engineering decisions — invoke the skill before your first question. The marginal cost is one slash-command or one click. The benefit is that every answer in that session is grounded in your source. The habit stacks: sessions without the skill are visibly worse.
Ask for citations every time you make a decision.
Whenever you use an AI answer to make or support a decision, append "— cite the source" or "— which page or section does this come from?" to your question. This keeps you honest about whether the answer is grounded in your source or drawn from the AI's general training. Citations also make your work auditable — if a colleague challenges a recommendation, you can trace it back to the exact passage.
Keep a misses list — feed it into the next version.
Whenever your skill gives an answer that is vague, off-topic, or wrong, note it down. One sentence: the question you asked, and what was missing from the answer. After a few weeks you will have a list of gaps that tells you exactly what to add to your source before the next rebuild. A skill that improves with each version is exponentially more valuable than one that never changes.
Pair the skill with your own real data in prompts.
The skill provides the framework. You provide the current facts. The most powerful prompts combine both: "Based on the qualification framework in my skill, and given these three facts about this prospect [paste facts], what is your assessment?" The skill grounds the reasoning; your data makes it specific. This combination is what separates a skilled AI from a generic one.
Re-run the gap-check prompt monthly.
Once a month, ask your skill: "What does my source not cover about [topic I care about this month]?" The answer will change as your work evolves and your questions become more sophisticated. This monthly check-in is also a good moment to look at your misses list from habit 03 and decide whether it is time to rebuild. A skill that you keep current is a compounding asset.
None of these habits requires extra setup. They are questions and rituals you layer on top of using your AI the way you already do. The payoff is that your skill goes from something you tried once to something that changes how you work every day.