Write the instruction once.
Everything the assistant does is a job you will want done again. A Skill is that job written down — the steps, the rules, which of your data it must check first, and the shape the answer comes back in — saved and re-run with one tap. Every other assistant resets to a blank box between uses. Your library is where these live.
“Which of your data it must check first” is the other half of this page. Tools are what a Skill is allowed to reach before it answers — your messages, calendar, reminders, contacts, files and knowledge base, or a web search — and you grant them per Skill, in the same editor as the instruction.
What are Skills
A Skill holds the instruction, how it should sound, what it is allowed to go and look at, and the documents it should read first. Every entry in your library is one of two kinds.
Always-on instructions added to every reply — tone, style, and approach. A system prompt plus tone sliders, and nothing else to configure.
For example: A voice for the family group, one for work, one for the client you write to every week.
Loaded on demand when relevant — a task the AI runs with its own tools and knowledge base, for work instructions alone can’t do.
For example: Propose a time from your calendar. Research a trip. Check whose birthday is coming up.
Which one to write: if you want to change how a reply sounds, write a Prompt. If you want the AI to go and find something out first, write a Skill.
And your own documents: the Knowledge Base
Attach documents and URLs to a Skill — up to 10 documents and 20 URLs — and it references them when it works: your price list, your policies, your brand guidelines, your team wiki. It answers the question out of what you already wrote down instead of out of memory. Everything is processed and stored on your device, searched with hybrid semantic + keyword retrieval, and cited in replies. Adding even one source switches the knowledge-base tool on for you. If that Skill drafts with Claude, only the retrieved passages travel with the prompt — never the whole document, and never the library.
Create Skills
Four routes, and all four end in the same editor — write one from a blank form, import one somebody else published, open a SKILL.md you already have, or let a shortcut build it. Tap a row to see how it goes.
Skill library → + → Create New Skill
A name, a description, and the instructions in plain English — then the tone sliders, the tools it may reach, and the documents it should know about. Nothing else is required: an entry with a name and a paragraph of instruction already works.
The description is doing more work than it looks like it is. On a Skill it is the trigger — the AI reads descriptions to decide which entry to load — so it should name the situation out loud rather than the outcome.
Step 1 of 4: Open your Skill library
Everything is per entry: the instructions, five tone sliders, the sampling mode, the tools it may reach, its documents, and which engine drafts with it — Apple Intelligence or Claude, with Extended Thinking optional on Claude. So the one you write for work and the one you write for family can sound nothing alike.
Not sure what to put in the box? Writing one well at the bottom of this page has a fill-in-the-blank template for each kind.
Skill library → + → Browse Marketplace
Skills are SKILL.md files — the open Agent Skills format, the same one Claude Code and the Claude apps read. Anything written for one runs here.
The built-in marketplace browses SkillsMP’s index of open-source Skills published on GitHub and imports one in a tap. It is a public index rather than a curated store, and much of what is in it was written for coding agents — so treat an import as a starting point you edit, not a finished assistant for your chats. Importing copies only the Skill definition to your device; the marketplace never sees a message.
Sales, marketing, finance, and project management.
Writing, design, documents, and content creation.
Knowledge bases, technical docs, and educational material.
LLM prompts, machine learning, and data analysis.
Wellness, writing, philosophy, arts, and culinary.
Academic, scientific computing, and lab tools.
Step 1 of 3: Browse the Skills Marketplace
Skills Marketplace
Discover open-source agent skills for Claude Code, Codex, ChatGPT, and any tool that uses SKILL.md.
Skill library → + → Import Skill
Pick a SKILL.md out of Files or iCloud Drive and it opens in the editor as a new Skill — the name, the description and the instructions read straight out of the file’s front matter. It is the same open format the marketplace serves, so a Skill written for Claude Code, sent to you by a colleague, or downloaded from anywhere at all comes in the same way.
A Skill that ships with reference material imports as a .zip of its folder instead, and the supporting files inside it land in that Skill’s Knowledge Base on the way in — nothing to attach by hand afterwards.
This is also the way back out. Select entries in your library and Export writes the same two shapes — a bare SKILL.md, or a .zip when the Skill carries documents — with your sliders, tools and settings preserved in the front matter. Export from one iPhone and import on another and you get the Skill you had, not an approximation of it.
Shortcuts → AI Reply Assistant → Create Skill
Create Skill builds an entry without the editor: a name, the instructions behind it, and the three tone controls. Because it is a shortcut, the instructions can come from whatever ran before it — text you dictated, a note, a page you shared — so a Skill can be written by the chain rather than typed into a form. It is a Siri phrase too.
The action sets the name, the instructions, creativity, formality and response length. Tools and a Knowledge Base are not shortcut parameters — open the new Skill in the editor to add those. Its sibling, Find Skill, reads your library back to a shortcut on nine properties. Both are in the Shortcuts chapter.
AI Tools — what it checks before it answers
Tool-calling allows the model to fetch live, up-to-date information from your messages, calendar, reminders, contacts, files and knowledge base, or perform a web search, to provide contextual replies — along with action cards that let you draft a reply, create a calendar event or set a reminder in one tap. This works on both the on-device AI model and the Claude cloud model, image generation included — that one is drawn on your iPhone by Apple Intelligence whichever engine asked for it.
On Claude, the model can do more than generate text. Code execution gives it a secure sandbox to run real calculations in, and to analyse a spreadsheet or CSV you attach — working from the actual rows rather than an estimate. It can also produce a finished file: a workbook, a document, a PDF, a chart or an HTML page, downloaded to your device as a card you can preview and share.
You control which tools the model can access — per Skill, in the editor above, and per chat in the agent chat. That is a privacy control, and a token and context management one: granting a narrower set of tools reduces the reasoning required, conserves context, and produces faster, more accurate responses.
List of AI Tools
Give your Skill access to live information and your own documents.
Search across your contacts, messages, and conversations — with deep cross-platform visibility.
Check your calendar and to-do lists without leaving the chat.
Contextual awareness of where you are.
Browse your files, draw pictures, and turn what the assistant worked out into tappable cards.
Capabilities that exist only when the work is running on Claude — there is no on-device equivalent, so these do nothing on the Apple foundation model.
Web Search
What it does
Look things up on the web and read what comes back — current information, well past any model’s training cutoff, with the sources cited. The two engines reach the web by different roads: on device it searches DuckDuckGo, picks the promising results, fetches those pages concurrently and summarises each one; on Claude the same switch turns on Anthropic’s own hosted search and its page reader, which can open any URL you point it at rather than only the ones in your Knowledge Base.
The switch, and what it turns on
- Runs on
- On-device and Claude
- Where you switch it
- Skill editor and agent chat
- The 2 calls behind it
- web_search · web_fetch
Where it runs
Your messages are never sent — only what is being looked up. On device that query goes to DuckDuckGo; with Claude, the search and the page fetch both run on Anthropic’s servers.
Knowledge Base Search
What it does
Hybrid search across the documents, URLs and files you attach to a Skill — 80% semantic similarity blended with 20% keyword matching, then AI-reranked for relevance. The same switch also lets it open one of your saved links and read its live content, so a page you attached stays current instead of being frozen at the moment you added it.
The switch, and what it turns on
- Runs on
- On-device and Claude
- Where you switch it
- Skill editor and agent chat
- The 2 calls behind it
- search_knowledge_base · fetch_url_content
Where it runs
Documents are processed and stored on your device and never uploaded. If the Skill drafts with Claude, only the passages actually retrieved travel with the prompt — never the whole document, and never the library.
Switches itself on
Switches itself on — and locks on — as soon as a Skill’s knowledge base has anything in it, whether that is a document or a link.
Message Search
What it does
Search what was actually said, with 15+ filters — by sender, platform, date, type, starred, forwarded, media, read status, and more. Search one chat, several, or every chat at once (for questions like what you have not read yet), with results grouped by conversation, and a fall back to semantic search when nothing matches exactly. It also pulls the messages surrounding a hit, so a reply that only makes sense in its thread arrives with the thread around it.
The switch, and what it turns on
- Runs on
- On-device and Claude
- Where you switch it
- Skill editor and agent chat
- The 2 calls behind it
- search_messages · get_message_context
Where it runs
Searches your local message database only. In the agent chat it is held to the chats you scoped, enforced inside the tool rather than asked of the model.
Chat Discovery
What it does
Find the conversation rather than the message: filter chats by platform, unread status, pinned, archived, favourite, participant count and activity date; find the groups you and one contact are both in; and browse the forum topics inside a Telegram supergroup, with their names, message counts and pinned status. Results are titles and metadata, never message content.
The switch, and what it turns on
- Runs on
- On-device and Claude
- Where you switch it
- Skill editor and agent chat
- The 3 calls behind it
- list_chats · get_mutual_chats · list_forum_topics
Where it runs
Reads your local conversation list. Nothing leaves your device, and the scope you set in the agent chat is enforced here too.
Contacts
What it does
Everything about the people, from four directions: your device address book — names, numbers, relationships, work info; the platform contact caches, which find someone by username or number even when they were never saved to your phone; whose birthday is coming up; and who is in a group chat, with roles and admin status.
The switch, and what it turns on
- Runs on
- On-device and Claude
- Where you switch it
- Skill editor and agent chat
- The 4 calls behind it
- search_device_contacts · search_birthdays · search_platform_contacts · list_group_members
Where it runs
Reads your local address book and your local platform contact caches. Nothing leaves your device.
Asks for Contacts permission first
This one reaches a protected iOS surface, so the first time a Skill uses it iOS puts up its own Contacts prompt. Decline and the tool simply stays unavailable — the AI carries on without it and says so inline.
Calendar & Events
What it does
Search events by title, date or location — up to 90 days back and a year ahead. Detects all-day events and shows which calendar each one is on. Read-only: it never creates or edits an event. Writing is an action card’s job, and that needs your tap.
The switch, and what it turns on
- Runs on
- On-device and Claude
- Where you switch it
- Skill editor and agent chat
- The call behind it
- search_calendar
Where it runs
Reads your local calendar. Nothing leaves your device.
Asks for Calendar permission first
This one reaches a protected iOS surface, so the first time a Skill uses it iOS puts up its own Calendar prompt. Decline and the tool simply stays unavailable — the AI carries on without it and says so inline.
Reminders & Tasks
What it does
Read your task lists, reminders and to-dos — what is due, what is overdue, what is on the shopping list. Read-only, like the calendar: a reminder gets created by an action card you confirm.
The switch, and what it turns on
- Runs on
- On-device and Claude
- Where you switch it
- Skill editor and agent chat
- The call behind it
- search_reminders
Where it runs
Reads your local reminders. Nothing leaves your device.
Asks for Reminders permission first
This one reaches a protected iOS surface, so the first time a Skill uses it iOS puts up its own Reminders prompt. Decline and the tool simply stays unavailable — the AI carries on without it and says so inline.
Location & Maps
What it does
Three things behind one switch: where you are now, what is near you, and how to get somewhere. It is what turns “where should we meet?” into a suggestion with a name and an address instead of a question back. Your position is cached for ten minutes, so a conversation that asks twice does not wake the GPS twice.
The switch, and what it turns on
- Runs on
- On-device and Claude
- Where you switch it
- Skill editor and agent chat
- The 3 calls behind it
- get_location · search_nearby_places · get_directions
Where it runs
Uses iOS Core Location. Your position is resolved on device; a place search sends the query and a coarse area, never your message.
Asks for Location permission first
This one reaches a protected iOS surface, so the first time a Skill uses it iOS puts up its own Location prompt. Decline and the tool simply stays unavailable — the AI carries on without it and says so inline.
Files & Documents
What it does
Search and read files in the locations you grant access to. Reads content from PDF, Word (DOCX), Excel (XLSX), PowerPoint (PPTX), Pages, Keynote, Numbers, RTF and HTML; notes and data files (TXT, Markdown, reStructuredText, JSON, XML, CSV, TSV, YAML, LOG, PLIST, INI, TOML, ENV, CONF, CFG, .properties); and source code (Swift, JavaScript, TypeScript, JSX/TSX, Python, Java, Kotlin, Ruby, Go, Rust, C, C++, Objective-C, C#, PHP, Shell/Bash/Zsh, SQL, R, Scala, Dart, Lua, Perl, Gradle, Groovy). It also works with the Files API, so a document can be uploaded once and referenced across conversations instead of re-sent every time.
The switch, and what it turns on
- Runs on
- On-device and Claude
- Where you switch it
- Skill editor and agent chat
- The 2 calls behind it
- search_files · read_file_content
Where it runs
Local file reads stay on device. Files API uploads go directly to Anthropic and need your own Anthropic key — they are not part of what AI Credit covers.
Worth knowing
On device the toggle is hidden until the Skill actually grants a file location, because the file tools can do nothing without one.
Image Generation
What it does
Draw an original picture on your iPhone with Apple Intelligence Image Playground, in one of three styles: Animation, Illustration or Sketch. What comes back is an action card — save it to Photos, share it, or send it into a chat.
The switch, and what it turns on
- Runs on
- On-device and Claude
- Where you switch it
- Skill editor and agent chat
- The call behind it
- generate_image
Where it runs
Generated entirely on device by Apple Intelligence — the prompt and the picture both stay there, whichever engine asked for it.
Generate Message Reply
What it does
The switch behind the action cards. With it on, the assistant can turn what it worked out into something tappable — an event, a reminder, a contact, a note, a draft message — instead of leaving it as text you have to act on yourself. Turn it off and you still get the answer; you just get it as prose.
The switch, and what it turns on
- Runs on
- On-device and Claude
- Where you switch it
- Skill editor and agent chat
- The call behind it
- request_card
Where it runs
Cards are built and rendered on your device, and nothing on one is sent or created until you tap it.
Worth knowing
On device this is not a tool the model calls: it is a second pass over the finished answer that pulls the cards out of it. That pass costs a model run per turn, which is why it can be switched off.
Task Tracking
What it does
Claude’s own checklist for a job with several steps, separate from your Reminders. It writes the plan down before it starts and ticks each item off as it finishes, so you can see which step it is on and what is still outstanding. The list exists only for the length of the task and is never saved to Reminders.
The switch, and what it turns on
- Runs on
- Claude only
- Where you switch it
- Skill editor and agent chat
- The call behind it
- update_todos
Where it runs
The list is part of the conversation with Claude — there is no on-device equivalent, so the toggle does nothing while a feature is set to Apple Intelligence.
Memory
What it does
Claude saves facts that persist after a conversation ends — a name, a preference, the way you like something done — and checks what it already knows before it starts the next one. The second time you ask for the same kind of work, you do not have to brief it again.
The switch, and what it turns on
- Runs on
- Claude only
- Where you switch it
- Skill editor only
- The call behind it
- memory
Where it runs
Memories are stored locally on your device and sent only as part of a request you made.
Sub-Agents
What it does
Claude splits a job across helper agents that run at the same time, each with its own separate context. This helps when a question needs six searches rather than one: the searches run concurrently, and the working notes from each one stay out of the main conversation.
The switch, and what it turns on
- Runs on
- Claude only
- Where you switch it
- Agent chat only
- The call behind it
- spawn_subagent
Where it runs
Inherits the privacy model of the active provider — a sub-agent sees only the scope its parent had.
Code Execution
What it does
Claude writes Python and runs it in a secure sandbox on Anthropic’s servers, then returns the result rather than the code. Use it for calculations the model should not do in its head, for analysing a spreadsheet or CSV you attach — the figures are computed from the actual rows instead of estimated — and for producing a finished file: a workbook, a document, a PDF, a chart or an HTML page, downloaded to your device as a card you can preview and share. It also handles format conversions and cleaning up messy data. It is the foundation Agent Skills run on, so it is switched on automatically whenever a Skill needs it.
The switch, and what it turns on
- Runs on
- Claude only
- Where you switch it
- Skill editor and agent chat
- The call behind it
- code_execution
Where it runs
Runs in Anthropic’s sandbox, not on your device. The container is isolated and has no internet access of its own. Container time is free on any turn that also carries web search or web fetch — which is every turn with Web Search on. Leave Web Search off and Code Execution on and that exemption lapses, and container time bills against your own key.
Worth knowing
The name says code, but most of what this tool does is not programming. It is the tool behind spreadsheet analysis: attach an Excel file or a CSV and Claude loads the real rows, computes the totals and trends, and can return a chart or a cleaned-up workbook. Without it the model can only work from what fits into the conversation, and estimate.
Agent Skills
What it does
Pre-built capabilities Claude loads into its container to produce a finished document: PowerPoint presentations (.pptx), Excel spreadsheets (.xlsx), PDF reports (.pdf) and Word documents (.docx). The file is downloaded to your device and appears as a card you can preview and share. Custom Skills you have uploaded through the Skills API are produced the same way.
The switch, and what it turns on
- Runs on
- Claude only
- Where you switch it
- Set per conversation, not in the tool list
Where it runs
Files are generated in Anthropic’s code-execution container using your own API key, so this one needs a key — it is not part of what AI Credit covers. The finished document is downloaded straight to your device.
Worth knowing
This is the one entry here that is not a switch in the tool list: Agent Skills are attached to a conversation, and the app forces Code Execution on underneath them for as long as they are active.
How to use Skills
Two places run an entry from your library. Both start it the same way — you pick the Skill and add an instruction — and both put the tools you granted it to work.
The Skill button sits above the keyboard in every chat. Pick an entry, add an instruction, say which messages it reads — the reply comes back as a draft you can keep talking to.
The agent chat runs a Skill across the chats you scope rather than one thread, with the model, the tools and the scope all set per conversation and changeable mid-thread.
By default a Skill runs on Apple Intelligence, on your iPhone. Flip the Skill feature to Cloud AI in Settings and the same button drafts with Claude instead — more on choosing between the two.
Writing one well
The instructions you write are the biggest lever you have over what comes back. Start from a template, and know the handful of things that quietly stop a Skill working.
Prompt
For a voice you want on every reply in a given relationship.
Example — You are a supportive running coach who celebrates small wins. Reply in 1–2 sentences using an encouraging voice. The recipient just started training for their first 5K.
Skill
For a task that repeats — something to look up, work out, and draft in one go.
Example — Description: Use when someone asks to meet up and I need to propose a time. Instructions: Check my calendar for the next seven days and find two or three free evening slots. Avoid days I already have an evening event. Offer them in a relaxed, friendly voice and ask which suits.
Things to watch out for
- Leaving the description vague. If it doesn't name a situation, there is nothing for the AI to match against — and the Skill sits in your library unused.
- Building one Skill that does everything. Split it: several small, sharply described entries beat a single sprawling one.
- Contradicting your tone sliders — writing "keep it very casual" while Formality is set high.
- Writing a wall of text, especially for Apple Intelligence. Three focused sentences beat three paragraphs.
- Listing too many rules. Past about five, some quietly stop being followed — decide which ones actually matter.
- Using in-house jargon without explaining it once.
- Giving an example so specific it gets copied verbatim instead of generalised from.
- Saying the same thing three different ways. It fills the context window and confuses smaller models.
- Turning on every tool. Enable the ones the job needs and leave the rest off.
How much you write depends on which engine the entry drafts with — the on-device model wants a few focused sentences where Claude can take examples and edge cases. Writing for each engine has that side by side.
Skills & Tools — questions, answered.
A wide range of formats:
- PDFs
- Microsoft Office — Word (DOCX), Excel (XLSX), PowerPoint (PPTX)
- Apple iWork — Pages, Keynote, Numbers
- Text & markup — rich text (RTF), Markdown, HTML, plain text
- Data & code — JSON, XML, CSV, Swift, JavaScript, Python, and more
- Images — AI Reply Assistant extracts text from those too
Each file can be up to 50 MB, with a total limit of 100 MB per Skill.
RAG stands for Retrieval-Augmented Generation, and it is the machinery behind the Knowledge Base above. The whole pipeline runs on-device. When you add a document:
- It’s broken into small chunks
- Each chunk gets a mathematical representation (an embedding)
- The embeddings are stored on your device
Then, when you ask for a reply:
- Your conversation is converted into a query
- The most relevant chunks are found using similarity search
- Those chunks are fed to the AI alongside the conversation
- The AI drafts a reply informed by both the thread and your documents — all on your device
RAG works with both. On iOS 26, Apple Intelligence powers the document processing pipeline — chunking, embedding, query optimisation, and retrieval all run on-device using Apple Foundation Models. The retrieved context is then passed to whichever AI provider your Skill uses (Apple Intelligence or Claude) for the final reply.
Import when somebody has already written the procedure you need and you would rather edit than start from a blank page — that is most people, most of the time. Write your own when the context is specific to you and repeats: a particular client, a community with its own jargon, a professional domain like legal or medical, or any situation where you want the AI to follow your guidelines every time.
Your personal AI assistant for messaging.
Download AI Reply Assistant on iOS and try it free.