mirror of
https://github.com/mintlify/docs.git
synced 2026-09-14 13:35:46 +08:00
017d7c7377
* fix: reduce Vale false positives via vocab and config updates - Make English-word vocab entries case-insensitive so sentence-start capitalization and normal prose usage stop flagging (agents, rest, cursor, setup, endpoints, etc.) - Add vocab entries for filenames and code identifiers that appear in frontmatter and JSX contexts (docs.json, llms.txt, CLAUDE.md, etc.) - Ignore openapi frontmatter lines, filenames, JSX attributes, email addresses, and internal link targets via TokenIgnores - Skip inline code scope and indented code fences - Add Headings exceptions for proper nouns (Claude Code, GitHub Actions, Route 53, GA4, etc.) - Disable linting for the all-code vercel-json-generator snippet Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: resolve all Vale warnings and errors across docs Content fixes: - Reword sentences using 'will', first person, spaced em dashes, 'e.g.', Latin abbreviations, and hyphenated adverbs - Sentence-case headings that started with dotted filenames - Backtick literal API values instead of bolding them - Move periods inside quotation marks - Fix Oxford comma rule misfires by restructuring sentences False-positive suppression: - Vale toggles around example user questions, keyboard shortcut keys (Cmd+I), UI labels, and code samples in JSX contexts that Vale misparses - Exclude vale toggle comments from the brace Token/BlockIgnores so in-document commands actually reach Vale (the greedy brace pattern was also silently swallowing large regions; now lazy) - Vocab entries for code identifiers (internal_id, handleSubmit, etc.) - Per-file rule disables for component docs with dotted JSX names and files where link-target linting ignores in-document toggles Result: vale --minAlertLevel warning is clean repo-wide; only suggestion-level items (Passive, Semicolons, Acronyms) remain. mint broken-links passes. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: correct WordList rule instead of degrading prose - Restore sentence-start "Email" in advanced-support; the rule now only flags hyphenated e-mail/E-mail forms - Restore the idiom "above all else"; the above->preceding swap now exempts "above all" Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: restore deliberate prose flagged by blunt rules - Restore spatial 'above the navbar' / 'above a page title' in custom-scripts; the above->preceding swap now exempts 'above the/a/an' - Restore the '= ...' in the react-components named-export example (the ellipsis is inside inline code) with an Ellipses toggle - Restore SLA phrasing 'will use commercially reasonable efforts' with a Will toggle - Restore the quoted developer question in the GEO guide intro with a FirstPerson toggle, matching the file's other example questions Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: remove WordList swaps that flag legitimate English - Drop tablet->device and firewalls->firewall rules; both words are correct in ordinary prose - Narrow touch->tap to UI-instruction phrasing (touch the/a) so 'keep in touch' and 'touch devices' stop flagging Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: reposition vale toggles to wrap whole blocks Toggle comments placed between list items split the lists (restarting ol numbering in deployments) and comments flush against tables risked breaking GFM table parsing. Wrap entire lists/tables with blank-line separation instead. Verified rendering with mint dev: single ol with two items, tables intact, no comments in visible DOM. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * refactor: prune accept.txt to load-bearing entries Empirically removed 166 vocab entries (575 -> 409) whose removal causes no Vale flags across all 908 English pages: dictionary words that never needed listing (agents, setup, endpoints, webhooks, yaml), lowercase entries the speller already accepts, and filename entries made redundant by inline vale toggles. Kept every case-enforcing entry (API, JSON, GitHub, ...) so casing policy is unchanged, plus entries that double as capitalization exceptions for headings (mcp, md, auth, cursor, txt). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * Update ai/mintlify-mcp.mdx * Update api/agent/v2/create-agent-job.mdx * chore: alphabetize Headings exceptions and accept.txt Case-insensitive sort, ignoring the (?i) prefix; also drops a duplicate Scala entry. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix: re-apply OxfordComma rewrite lost in branch merge Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * docs: resolve Vale suggestions batch 1 (acronyms, semicolons, passive voice) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * docs: resolve remaining Vale suggestions (passive voice batch 2) Rewrite ~90 passive constructions to active voice across deploy, guides, migration-services, and organize docs. Toggle the deliberate passive examples in the style-and-tone guide ('by zombies' test). Add axios/lodash vocab entries for a repositioned code example. vale . is now fully clean: 0 errors, 0 warnings, 0 suggestions. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
73 lines
3.7 KiB
Plaintext
73 lines
3.7 KiB
Plaintext
---
|
|
title: "Add assistant skills"
|
|
description: "Tune the assistant's responses with topic-specific Markdown guides that the assistant loads on demand to answer specific questions with your team's guidance."
|
|
keywords: ["assistant", "skills", "loadSkill", "guides", "system prompt"]
|
|
---
|
|
|
|
Skills are Markdown files that direct the assistant's responses on specific topics. The assistant has access to any skills you add to your project in the `.mintlify/assistant/skills/` directory. It loads a skill only when a user's question matches it, so you can provide curated answers for specific topics without adding context to every response.
|
|
|
|
Use skills to:
|
|
|
|
- Provide specific answers for topics where search results alone are unreliable.
|
|
- Encode playbooks, troubleshooting flows, or migration guides that aren't structured as documentation pages.
|
|
- Guide users through multi-step or nuanced procedures.
|
|
|
|
## Add a skill
|
|
|
|
Add a `skill.md` file inside a directory named for the skill under `.mintlify/assistant/skills/`:
|
|
|
|
```text Example directory structure
|
|
.mintlify/
|
|
└── assistant/
|
|
└── skills/
|
|
├── migrate-to-v3/
|
|
│ └── skill.md
|
|
└── debug-webhooks/
|
|
└── skill.md
|
|
```
|
|
|
|
Each `skill.md` file needs YAML frontmatter with a `name` and `description`, followed by the skill instructions:
|
|
|
|
```markdown .mintlify/assistant/skills/migrate-to-v3/skill.md example skill file
|
|
---
|
|
name: Migrate to SDK v3
|
|
description: Step-by-step guide for upgrading from SDK v2 to v3, including breaking changes, code mods, and rollout checklist.
|
|
---
|
|
|
|
# Migrating to SDK v3
|
|
|
|
Follow these steps to upgrade an existing v2 integration to v3.
|
|
|
|
## 1. Update dependencies
|
|
|
|
...
|
|
```
|
|
|
|
The assistant uses the `name` and `description` to decide when to load a skill, so write descriptions that state when the skill applies. Keep the instructions concise. The assistant truncates skill content at approximately 6,000 tokens.
|
|
|
|
<Note>
|
|
After you add or change a `skill.md` file, redeploy your site to make the skill available to the assistant.
|
|
</Note>
|
|
|
|
## How the assistant uses skills
|
|
|
|
The assistant lists every available skill by name, slug, and description in its system prompt. When a user's question matches a skill description, the assistant loads the skill and treats it as high-priority guidance that takes precedence over general search results.
|
|
|
|
The assistant can load up to three skills per response.
|
|
|
|
On follow-up messages, the assistant removes previously loaded skill content from the conversation history to preserve the context window. It reloads a skill if it needs it again.
|
|
|
|
Skills work alongside [custom assistant instructions](/assistant/customize) in `.mintlify/Assistant.md`. Use `Assistant.md` for tone, persona, and rules that apply to every response. Use skills for topic-specific guidance that the assistant should load only when relevant.
|
|
|
|
## Write effective skills
|
|
|
|
- Give each skill a distinct `description`. Overlapping descriptions make it harder for the assistant to load the right skill.
|
|
- Use a unique, URL-safe slug for each directory name. Slugs must contain only lowercase letters, numbers, and hyphens. You cannot publish skills with duplicate slugs.
|
|
- Structure skills with clear headings and numbered steps.
|
|
- Keep skills self-contained. The assistant loads one skill at a time, so a skill should not depend on another skill.
|
|
- Use prose and code samples. Do not use screenshots.
|
|
|
|
## Disable skills
|
|
|
|
Skills are on by default when at least one `skill.md` file is present. To remove all skills, delete the `.mintlify/assistant/skills/` directory and redeploy your site. To turn off the skills feature for your deployment, contact [support](mailto:support@mintlify.com).
|