Manual Mode
Generate a single focused skill by entering details by hand.
Manual Mode generates a single skill from details you type in directly. No codebase scan, no premade catalog, just the analyzer pointed at one specific topic.
//When To Use
- The technology isn't auto-detected (private internal library, unusual stack).
- You want a skill for a domain concept rather than a framework, e.g. "billing-rules", "video-encoding-pipeline", "ranking-algo".
- You're documenting an internal pattern your team uses across the codebase.
- You want to prototype a skill before committing to it.
For team integrations like Stripe / Linear / Sentry, prefer /saikit-workflows, it ships hand-written ops skills that are deeper than what a one-off analyzer pass produces.
//Workflow
Pick "Create a focused skill by entering details" from the mode menu. The CLI walks you through five fields, then generates the skill in seconds.
1. Skill Name
> react-query
Lowercase, hyphens only, no leading/consecutive hyphens, ≤ 64 chars (Agent Skills spec).
2. Description
> Fetching and caching server state in React with TanStack Query
What the skill does, in one sentence. This is what Claude/Cursor/Codex matches user intent against, pack it with high-signal keywords.
3. Technology
> TanStack Query v5 with oRPC integration
The specific tech, version, and any integration context.
4. Use Cases
> Data fetching, mutations, optimistic updates, query invalidation
The kinds of work the skill should activate on.
5. Confirm & Generate
Review the inputs, pick a model (Sonnet for most things, Opus for complex frameworks), and watch the skill stream in.
//Fields
| Field | What good looks like |
| --- | --- |
| Name | react-query, webhook-replay, feature-flag-hygiene |
| Description | Specific verbs + nouns the user will actually say |
| Technology | Include the version. "React 19 with Server Components" beats "React" |
| Use cases | Three to five concrete tasks, not abstract domains |
//Best Practices
Be specific.
❌ "Help with database queries"
✅ "PostgreSQL with Drizzle ORM for the user/post schema"
Include versions.
❌ "React"
✅ "React 19 with Server Components and the use() hook"
Front-load keywords in the description. Claude matches semantically, but exact-keyword overlap is still the strongest signal. Mention the library name, the version, the file extension, and the most common verbs.
One skill per concern. A skill that covers "database, auth, and billing" will get matched on too many requests and fire when it shouldn't. Split.
//Manual vs. Analyze vs. Workflows
| Mode | Source | Best for |
| --- | --- | --- |
| Analyze | Codebase scan | First-time setup, framework migration, full repo refresh |
| Manual | Your input | Internal library, domain skill, pattern you can describe but the analyzer won't see |
| /saikit-workflows | Detection + your answers | Team operational tools (Stripe, Linear, Sentry, …) and the way you ship |
These compose. A typical setup runs Analyze first, runs /saikit-workflows second, and reaches for Manual only when something specific is missing.