Manage Skills¶
Skills are folders of task instructions and helper scripts that the model reads on demand when a request calls for them, such as building a Word document, auditing an Excel workbook, or designing a poster. All chats and models share the same skill folders.
How skills reach the model¶
- The first user message of each chat includes a compact
<available_skills>index with one line per active skill (its name and when to use it). Because the index only changes when skills or tier permissions change, it stays inside the provider's prompt cache across turns. - When a request matches a skill, the model calls
read_skillto readSKILL.mdand any reference files it links to. - When Python code in
code_interpreterreferences/mnt/skills/<name>, that skill's folder is mounted read-only in the sandbox for that run. Helper scripts can be imported after adding/mnt/skills/<name>/scriptstosys.path, and their allowlisted package dependencies load automatically. Both code execution and the Skills tier permission are required for sandbox mounts.
Skill helpers run inside the standard Python sandbox with outbound networking disabled by default.
Built-in skills¶
| Skill | Use |
|---|---|
data-analysis |
Load, profile, clean, and analyze uploaded CSV, Excel, or JSON data |
docx |
Create, edit, and inspect Word documents |
event-planning |
Plan events ranging from a dinner party to a wedding |
financial-calculator |
Loans, savings, retirement, rent-versus-buy, and other financial math |
gif-creator |
Build small animated GIFs |
learn |
Guided tutoring, practice exercises, quizzes, flashcards, and study plans |
pdf |
Read, split, merge, fill, stamp, and create PDFs |
poster-design |
Create posters, flyers, cards, and generative art as PNG or PDF |
pptx |
Create and edit PowerPoint slide decks |
shopping-research |
Compare products and offers before buying |
work-updates |
Status reports, newsletters, incident notes, and announcements |
writing-coauthor |
Co-write longer documents step by step |
xlsx |
Build, edit, and audit Excel workbooks |
If a skill requires a tool or sandbox package that is disabled on the server (or
unavailable in the active chat or sub-agent), it is omitted from both
<available_skills> and read_skill's name enum, and the admin page shows
what is missing. When no skills are usable or skills are switched off in Admin,
both <available_skills> and the read_skill tool schema are omitted.
Switch skills on or off¶
- Admin page: open Models and scroll to Skills. The top toggle turns all skills on or off, and individual rows toggle single skills. Changes take effect on the next request without restarting the server.
- Server config:
SKILLS_ENABLED=falsedisables all skills and removes theread_skilltool at startup;SKILLS_DENYLISThides specific skills by name. - Tiers: the Skills feature on the Tiers admin page controls whether
users in that tier can use
read_skilland/mnt/skills/mounts. - If you customize
TOOLS_ALLOWLIST, keepread_skillin the list.
See Configuration for the environment settings.
Add a custom skill¶
Built-in skills live in src/numi_chat/skills/<name>/. Operators can add
instance-specific skills without modifying the repository by placing a folder
under state/skills/<name>/ (or the directory set in SKILLS_PRIVATE_DIR) and
restarting Numi. A private skill with the same name as a built-in skill
replaces it.
state/skills/meal-prep/
├── SKILL.md required
├── reference/*.md optional, read on demand via read_skill
└── scripts/*.py optional Python helpers for code_interpreter
SKILL.md starts with single-line YAML frontmatter:
---
name: meal-prep
description: Plan a week of meals with a shopping list. Use when the user asks for meal plans or batch cooking.
requires: code_interpreter
packages: pandas
---
# Meal prep
Instructions for the model...
| Field | Rule |
|---|---|
name |
Matches the folder name: 1 to 64 lowercase letters, digits, and hyphens, starting with a letter or digit |
description |
Single line, up to 300 characters. State what the skill does and when to use it |
requires |
Optional comma-separated tool names required by the skill |
packages |
Optional comma-separated sandbox packages (each must be in TOOLS_CODE_INTERPRETER_DEPS_ALLOWLIST) |
Keep SKILL.md concise and put longer reference material in reference/*.md.
read_skill reads text files up to 200,000 bytes; binary assets and helper
scripts are accessed from code_interpreter.
Document styling defaults¶
Unless the user asks for a specific palette or typeface, the built-in office and PDF helpers use a clean default style: Arial (or Helvetica in PDFs), dark headings, and Numi blue accents for rules, table headers, and charts. When the user specifies their own colors or fonts, the helpers use those while keeping body and heading text legible. Uploaded files keep their existing styling when edited.
Security and sharing notes¶
- Private skills are available to all accounts whose tier grants Skills. Never place secrets or credentials inside skill files, as their text is sent to the active model provider when read.
- Symlinks that point outside the skill directory are rejected by both
read_skilland sandbox mounting. - Public share links omit
read_skillcalls and raw skill text, while still showing the assistant's final reply and any generated files referenced in the shared chat.