Loading Accelerated Logic…
GA Release

The Multi-Agent WebGPU Engine & Interactions API are now generally available.

Part 12: Skills & Personas — Custom Instructions You Toggle Like Superpowers

Teach AI agents specialized roles and behavioral personas in seconds. Stack multiple technical skills or switch tones dynamically with slash commands.

Category: Customization • Read Time: 25 min read • Updated: 2026-11-10

1. Skill vs Persona — Key Differences

Both Skills and Personas inject specialized system instructions into model prompts, but they govern distinct layers of agent behavior:

⚡ Skill (Stackable Capability)

An additive tool or domain expertise. You can stack multiple skills simultaneously (e.g. Security Auditor + Performance Optimizer + Accessibility Checker). Displayed as blue chips above the composer.

🎭 Persona (Exclusive Identity)

A tone or behavioral identity. Only ONE persona can be active at a time. Enabling a new persona automatically deactivates the previous persona. Displayed as indigo chips with message icons.

2. How to Find & Equip Skills

Method 1: Composer Slash Commands (Fastest)

Type / in any chat input. A popover menu lists all custom and preset skills. Use arrow keys to navigate and press Enter or click to equip.

Method 2: Skills Manager (Full Control)

Open Sidebar → Skills (Sparkles icon) to view your full library, create custom skills, edit system prompts, and duplicate or delete entries.

3. The Built-in Preset Library

The Skills Manager groups reusable instructions by common work patterns. Available items can vary with the current build and your own saved skills:

Analysis & Quality Security Auditor, Performance Auditor, Accessibility Checker
Writing & Media Technical Documentation Writer, Image Prompt Writer, Story Architect
Tool-aware instructions Reusable guidance for configured providers and tools; a skill alone does not connect external services.
Personas Friendly Tutor, Senior Architect, Rubber Duck Debugger

4. Creating Custom Skills in 1 Minute

Custom Skill Authoring Pattern:

# Name: Tailwind Accessibility Auditor

# Description: Reviews UI code and enforces WCAG contrast & responsive classes

You are a Tailwind CSS Accessibility Auditor.

- Review code for low contrast (<4.5:1), missing focus rings, and aria-labels.

- Do not alter core business logic; update class names only.

- Output a bulleted list of issues followed by the updated code block.

5. Active Skills Staging Bar

When skills or personas are enabled, an active staging bar appears directly above the chat composer. Each chip displays the skill name with a quick X button to remove or disable it instantly.

6. How System Prompts Incorporate Active Skills

During model inference, active skills sit at a high priority within the system instruction stack:

Base System Prompt

+ Folder Memory & Relevant Facts

+ Active Skills System Prompts (Concatenated)

+ Active Persona Tone Prompt

+ Attached Reference Files & Tools

+ User Prompt Message

7. AI Self-Generated Skills

When you ask for a reusable skill or persona and the skill proposal permission is enabled, an agent can emit a <create_skill name="..." description="..."> proposal. Review it above the composer and choose Keep to save it; otherwise discard it or let it expire after 60 seconds.

8. Duplication, Backup & Cloud Sync

In Skills Manager, click the Duplicate (Copy) icon to create variations of existing personas (e.g. Friendly Tutor - Spanish Edition). Skills synchronize automatically to Google Drive in my_ai_settings.json when cloud sync is enabled.

9. Recommended Custom Skills Library

Tailwind Stylist Enforces mobile-first responsive design, contrast, and clean layout utility classes.
Security Auditor Scans code for XSS, SQL injection vulnerabilities, and insecure storage calls.
Accessibility Checker Enforces WCAG AA standards, aria-labels, and keyboard navigation traps.
Senior Architect (Persona) Direct, concise output with zero conversational fluff and immediate code solutions.

10. Skills Troubleshooting

Skill Ignored? Ensure the skill chip is active in the composer staging bar and verify that the system prompt contains specific output constraints rather than generic text.

Personas Overwriting Each Other? Remember that Personas are mutually exclusive. Enabling a new persona automatically replaces the active one by design.

11. Anatomy of a High-Quality Skill Prompt

Skill prompts are injected verbatim. A skill that changes behavior is written like a contract — role, scope, constraints, output shape — not like a job description aspiration:

# Role

You are a release-note writer for a developer-tools company.

# Scope

Only rewrite the changelog entries the user pastes. Do not invent features.

# Hard constraints

- One bullet per user-visible change; group under Added / Fixed / Changed.

- Each bullet names the affected surface (CLI flag, config key, API route).

- No marketing adjectives ("blazing", "seamless", "delightful").

# Output contract

Markdown. Ready to paste into CHANGELOG.md. End with "Version bump: <type>".

The four load-bearing elements: a narrow role (not "helpful assistant" but "release-note writer for dev tools"), an explicit scope boundary (what the skill will NOT do), measurable constraints (word bans, format bans), and an output contract the model can self-check against. Skills that fail tend to be missing the scope line — the model defaults to doing the whole job superset.

12. Composing Multiple Skills Without Them Fighting

Skills stack by concatenation — the active skill prompts land back-to-back in the system instruction. Two rules keep a stack coherent:

Stack capabilities, not identities

Security Auditor + Accessibility Checker compose cleanly (two review lenses on the same code). "Brutal Senior Reviewer" + "Supportive Mentor" do not — both try to set tone, and the model averages them into mush.

Watch for scope collisions

If two active skills both constrain output format (e.g. "always answer in bullets" + "always answer in prose"), the last-applied wins opaquely. Keep at most one format-governing skill per turn, or make the precedence explicit inside one of them.

Order intuition: treat the stack like nested code review — broad correctness skills (Security, A11y) outermost, formatting skills innermost, personas exclusive of everything. When in doubt, disable all skills, confirm the baseline answer is sound, then add skills back one at a time and watch which one changes behavior.

13. Test-Driving a Skill Before You Trust It

Thirty seconds of evaluation prevents a skill that quietly degrades every future answer. A minimal protocol:

  1. Baseline first. Ask your candidate prompt with no skills active; save the answer.
  2. Equip and re-ask the identical prompt. Diff the two. The skill should change behavior along its declared axis (e.g. surfaces XSS issues) without collateral damage (e.g. ignoring half the original question).
  3. Probe the edges: one prompt where the skill should abstain ("review this poem for SQL injection") and one where it must trigger unambiguously.
  4. Regression-check with two other task types. If a "security auditor" suddenly stops answering non-security questions at all, its constraints are too aggressive — narrow them.
  5. Demote load-bearing skills to memory or folder memory? If the "skill" is really a standing preference rather than a task lens, it belongs in Memory (Part 11), not the chips row.

14. Backing Up & Moving Your Skills Library

Skills are plain data in your browser. Keep them portable:

  • Google Drive sync: when Drive backup is enabled, settings (including skills) sync inside the hidden AppData vault — your own Google account, cross-device (Part 17).
  • Snapshot export: the full workspace snapshot includes skills; the per-slice export lets you carry just the skills file between machines.
  • Sharing with teammates: copy the skill's prompt text out of the manager and hand it over — skills have no secret fields, which is what makes them shareable. There is no one-click publish flow; treat curated skills like shared internal documentation.
  • Versioning habit: when you improve a skill, keep the old one duplicated as Name (v1) for one week as a rollback — prompt edits can regress behavior silently.