# AI

> AI — Index

- Skill: `persimmon-automation-labs/ai` (Agent Skill)
- Install (CLI): `npx skillmds@latest add persimmon-automation-labs/ai`
- Raw SKILL.md: https://api.skillmd.com/api/skills/persimmon-automation-labs/ai/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: Persimmon-Automation-Labs (https://skillmd.com/u/persimmon-automation-labs)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/persimmon-automation-labs/ai

---


# AI — Index

Persimmon AI = the Anthropic Claude SDK funneled through one wrapper, prompts kept in one library, and RAG grounded on pgvector. This mother is a map; follow the child for the actual work.

## Trigger

- "Add AI to this feature" / "Call Claude for…"
- "Write the prompt for…" / "Where do prompts live?"
- "Retrieve context" / "RAG" / "embed and search"

## The child skills

| Skill | When to use | Owns |
|---|---|---|
| `ai-sdk-wrapper` | Any Claude call, or fixing retry/cost/caching | `src/lib/ai/claude.ts` — retries (3, exp backoff), prompt caching, model routing (`pickModel`), token accounting, streaming |
| `ai-prompt-library` | Writing/editing any prompt | `src/lib/ai/prompts.ts` — centralized prompt templates; no inline prompt strings in business logic |
| `ai-rag-retrieval` | Retrieval-augmented features | Chunking, embedding, pgvector retrieval, reranking, citation grounding |

## How to route

1. **Calling Claude?** → `ai-sdk-wrapper` (always — never import `@anthropic-ai/sdk` directly elsewhere).
2. **Writing a prompt?** → `ai-prompt-library` (no inline prompt strings).
3. **Need grounded context?** → `ai-rag-retrieval` (uses `data-prisma-pgvector` for the vector store).

## Persimmon AI defaults — one-screen summary

- **One wrapper.** All Claude traffic through `src/lib/ai/claude.ts`. Never import the SDK elsewhere.
- **One prompt home.** All prompts in `src/lib/ai/prompts.ts`. No inline prompts in business logic.
- **Model routing**: Sonnet for classification / per-item / volume; Opus only for genuine deep reasoning (composition, multi-doc synthesis).
- **Prompt caching** (`cache_control: { type: "ephemeral" }`) on stable system/doctrine blocks meeting the per-model minimum (Sonnet 4.6 = 2,048 tokens; Opus 4.6 / Haiku 4.5 = 4,096). Below the threshold it silently no-ops.
- **Retries**: 3 with exponential backoff (1s → 4s → 16s); SDK default of 2 is too low under 529 load.
- **Persist every AI output.** Never regenerate if a result exists — expensive and non-deterministic.

## Anti-patterns banned

- Importing `@anthropic-ai/sdk` outside the wrapper
- Inline prompt strings in components/actions
- Opus where Sonnet suffices (cost)
- Regenerating an existing AI result
- Ungrounded generation when a RAG citation is required (see `legal-brief-composer`)

## Relationship to other mothers

| Mother | Connection |
|---|---|
| `data` | RAG vector store lives in pgvector via `data-prisma-pgvector` |
| `domain-legal` | Legal generation is RAG-grounded; uses `ai-*` under the hood |
| `quality` | Prompt hygiene/output checked by `quality-review-prompt-output` |