ai-ops-cli 0.1.24 → 0.2.0

package/README.md

@@ -1,42 +1,38 @@
 # ai-ops-cli
 
-CLI for managing AI tool rules and presets across projects.
+CLI for managing core AI tool rules and agent skills across projects.
 
 ## Why this exists
 
 `ai-ops-cli` reduces configuration drift across AI coding tools.
 
-- Different tools use different file layouts and loading models.
-- Manual sync across tools is error-prone over time.
-- Teams need a deterministic, repeatable setup for AI pair-programming rules.
+- different tools use different file layouts and loading models
+- manual sync across tools is error-prone over time
+- teams need a deterministic setup for shared core rules and installable skills
 
-The CLI uses centralized YAML rules as SSOT and renders tool-native files into the current project.
+The CLI treats these as separate sources of truth:
 
-> **📌 Core Concept**
->
-> Instead of directly managing platform-specific files, manage **abstract metadata** as SSOT and achieve **Asset Centralization** across multiple AI environments.
+- `apps/cli/data/rules/*.yaml`: always-loaded core rules only
+- `apps/cli/data/skills/skill-registry.json`: install/catalog metadata for skills
+- `apps/cli/data/skills/reference-skills/<skill-id>/`: installable reference skills
+- `apps/cli/data/skills/task-skills/<skill-id>/`: installable task skills
+- `apps/cli/data/presets.yaml`: preset-to-core-rule mapping
 
 ## What this CLI provides
 
-- Interactive installation (`ai-ops init`)
-- Source drift checks (`ai-ops diff`)
-- Deterministic re-apply (`ai-ops update`)
-- Managed cleanup (`ai-ops uninstall`)
-- Project-only operation (no global scope)
-
-## What this CLI does not provide
-
-- Hosted backend or remote state
-- In-CLI rule authoring UI
-- IDE plugin management
+- interactive project initialization (`ai-ops init`)
+- skill package installation and lifecycle management (`ai-ops skill ...`)
+- source drift checks (`ai-ops diff`)
+- deterministic re-apply (`ai-ops update`)
+- managed cleanup (`ai-ops uninstall`)
 
 ## Supported tools and output layout
 
-| Tool | Single project output | Monorepo output |
-| --- | --- | --- |
-| Claude Code (`claude-code`) | `.claude/rules/<rule-id>.md` | Shared rules in `.claude/rules/*.md`, domain rules in `<workspace>/CLAUDE.md` |
-| Codex (`codex`) | `AGENTS.md` and `AGENTS.override.md` | Root `AGENTS.md` and `<workspace>/AGENTS.override.md` |
-| Gemini CLI (`gemini`) | `GEMINI.md` | Root `GEMINI.md` and `<workspace>/GEMINI.md` |
+| Tool                        | Project rules output                                  | Skill output                 |
+| --------------------------- | ----------------------------------------------------- | ---------------------------- |
+| Claude Code (`claude-code`) | `.claude/rules/<rule-id>.md`, `<workspace>/CLAUDE.md` | `.claude/skills/<skill-id>/` |
+| Codex (`codex`)             | `AGENTS.md`, `<workspace>/AGENTS.override.md`         | `.agents/skills/<skill-id>/` |
+| Gemini CLI (`gemini`)       | `GEMINI.md`, `<workspace>/GEMINI.md`                  | `.agents/skills/<skill-id>/` |
 
 Optional settings files:
 
@@ -53,27 +49,40 @@ npm install -g ai-ops-cli
 ## Usage
 
 ```bash
-# Initialize rules for the current project
+# Initialize the current project
 ai-ops init
 
-# Check drift against current source hash
+# Install a skill globally (user scope by default)
+ai-ops skill install skill-load-check --tool codex
+
+# Install a skill only for the current project
+ai-ops skill install skill-load-check --project --tool codex
+
+# Inspect or update installed skills
+ai-ops skill list
+ai-ops skill diff
+ai-ops skill update
+ai-ops skill uninstall skill-load-check
+
+# Check project drift
 ai-ops diff
 
-# Re-apply installed rules (or force)
+# Re-apply current project state
 ai-ops update
 ai-ops update --force
 
-# Remove installed files and manifest
+# Remove project-managed files
 ai-ops uninstall
 ```
 
-## CLI surface
+## CLI Surface
 
 ```text
 ai-ops [command]
 
 Commands:
   init       Initialize AI tool rules for a project
+  skill      Manage agent skills
   update     Update installed rules
   diff       Show diff between installed and current rules
   uninstall  Remove installed rules and manifest
@@ -86,32 +95,46 @@ Options:
 
 Notes:
 
-- `--scope` is deprecated and explicitly rejected. The CLI is project-only.
-- The installation state is tracked in `.ai-ops-manifest.json` at project root.
+- Project installation state is tracked in `.ai-ops-manifest.json`.
+- User-scope skill state is tracked in `~/.ai-ops/skills-manifest.json`.
+- `ai-ops skill` defaults to user scope. Use `--project` to keep a skill local to the current repo.
 
-## How install/update/uninstall behave
+## Install / Update / Uninstall Behavior
 
-- Managed files are wrapped in an `ai-ops` section with metadata (`sourceHash`, `generatedAt`).
-- If a file already has an `ai-ops` section, only that section is replaced.
-- If a file has no managed section, generated content is appended and user content is preserved.
-- `uninstall` removes only managed sections for appended files and keeps user-authored content.
+- Managed project rule files are wrapped in an `ai-ops` section with metadata (`sourceHash`, `generatedAt`).
+- If a rule file already has an `ai-ops` section, only that section is replaced.
+- If a rule file has no managed section, generated content is appended and user content is preserved.
+- Skill packages are written into dedicated directories and replaced as full package trees on update.
+- `uninstall` removes only project-managed rule files and project-installed skill directories.
+- User-scope skills are never removed by `ai-ops uninstall`.
 
-## Init flow summary
+## Init Flow Summary
 
 `ai-ops init` prompts for:
 
 1. Tool selection (`claude-code`, `codex`, `gemini`)
 2. Monorepo confirmation
-3. Preset selection per workspace
-4. Domain rule fine-tuning per workspace
-5. Optional settings installation
+3. Workspace selection for monorepos
+4. Preset selection per workspace
+5. Locked core rules review
+6. Preset-linked `reference` skills only:
+   - already-installed global skills are shown separately
+   - only installable skills can be deselected
+7. One shared install scope for selected installable skills (`user` default or `project`)
+8. Optional settings installation
+
+Important behavior:
 
-Preset and rules are loaded from:
+- core rules come from the preset directly and are not fine-tuned in `init`
+- `init` shows only preset-linked `reference` skills
+- `task` skills are excluded from `init` and are managed with `ai-ops skill install/uninstall`
+- globally available skills are not reinstalled by default
+- when `user` scope is chosen, selected skills are written only to the global skill registry
+- when `project` scope is chosen, selected skills are recorded in `.ai-ops-manifest.json`
 
-- `apps/cli/data/presets.yaml`
-- `apps/cli/data/rules/*.yaml`
+Skill authoring rules live in `apps/cli/data/skills/README.md`.
 
-## Local development
+## Local Development
 
 From repo root:
 
@@ -129,7 +152,46 @@ npm run build --workspace=apps/cli
 npm run test --workspace=apps/cli
 ```
 
-## Related docs
+## Local Skill Loading Check
+
+Use the built-in `skill-load-check` task skill before publishing to npm.
+
+Recommended local flow:
+
+```bash
+# 1. Build the CLI
+npm run build
+
+# 2. Use an isolated user home so you do not pollute your real ~/.agents or ~/.claude
+export AI_OPS_HOME="$(mktemp -d)"
+
+# 3. Install the sample skill globally
+node apps/cli/dist/bin/index.js skill install skill-load-check --tool codex
+
+# 4. Verify files exist
+find "$AI_OPS_HOME/.agents/skills/skill-load-check" -maxdepth 2 -type f | sort
+
+# 5. Run the sample script manually
+node "$AI_OPS_HOME/.agents/skills/skill-load-check/scripts/loaded.js"
+```
+
+Expected output:
+
+```text
+A Skill loaded
+```
+
+Project-scope verification:
+
+```bash
+node apps/cli/dist/bin/index.js skill install skill-load-check --project --tool codex
+find ./.agents/skills/skill-load-check -maxdepth 2 -type f | sort
+node ./.agents/skills/skill-load-check/scripts/loaded.js
+```
+
+After file placement is verified, use a real tool prompt that should trigger `skill-load-check` and confirm the tool discovers the skill metadata. If the tool caches skill discovery, restart that tool session before re-checking.
+
+## Related Docs
 
 - Master blueprint: [`docs/plan.md`](../../docs/plan.md)
 - Implementation playbook: [`docs/implementation-playbook.md`](../../docs/implementation-playbook.md)

package/data/presets.yaml

@@ -6,13 +6,6 @@ frontend-web:
     - code-philosophy
     - plan-mode
     - naming-convention
-    - engineering-standards
-    - typescript
-    - react-typescript
-    - shadcn-ui
-    - nextjs
-    - graphql
-    - libs-frontend-web
 
 frontend-app:
   description: '앱 프론트엔드 프로젝트를 위한 프리셋'
@@ -22,10 +15,6 @@ frontend-app:
     - code-philosophy
     - plan-mode
     - naming-convention
-    - engineering-standards
-    - flutter
-    - graphql
-    - libs-frontend-app
 
 backend-ts:
   description: 'TypeScript 백엔드 프로젝트를 위한 프리셋'
@@ -35,13 +24,6 @@ backend-ts:
     - code-philosophy
     - plan-mode
     - naming-convention
-    - engineering-standards
-    - typescript
-    - nestjs
-    - prisma-postgresql
-    - graphql
-    - nestjs-graphql
-    - libs-backend-ts
 
 backend-python:
   description: 'Python 백엔드 프로젝트를 위한 프리셋'
@@ -51,8 +33,3 @@ backend-python:
     - code-philosophy
     - plan-mode
     - naming-convention
-    - engineering-standards
-    - python
-    - fastapi
-    - sqlalchemy
-    - libs-backend-python

package/data/skills/README.md

@@ -0,0 +1,157 @@
+# Skill Authoring Guide
+
+This directory is the source of truth for installable agent skills.
+
+## Terms
+
+### Reference Skill
+
+A `reference skill` is a lazy-loaded knowledge pack.
+
+- Canonical detail lives in `references/reference.md`
+- `SKILL.md` stays thin and tells the agent when to use the skill and what to read first
+- Use it for standards, stack guidance, and large domain references
+
+### Task Skill
+
+A `task skill` is a procedural workflow.
+
+- Canonical procedure lives in `SKILL.md`
+- `references/` is optional supporting material only
+- Use it for repeatable checks, actions, and guided workflows
+
+## Directory Shape
+
+```text
+apps/cli/data/skills/
+  README.md
+  skill-registry.json
+  reference-skills/
+    <skill-name>/
+      SKILL.md
+      agents/       # optional
+      references/   # required for reference skills
+      assets/       # optional
+      scripts/      # optional
+  task-skills/
+    <skill-name>/
+      SKILL.md
+      references/   # optional
+      assets/       # optional
+      scripts/      # optional
+```
+
+## Authoring Rules
+
+1. Directory name must exactly match frontmatter `name`.
+2. `SKILL.md` must start with YAML frontmatter.
+3. `kind`, `supported_tools`, `install_scopes`, preset grouping, and `source_path` live in `skill-registry.json`.
+4. A `reference` skill must live under `reference-skills/` and include `references/reference.md`.
+5. A `task` skill must live under `task-skills/` and keep its executable procedure in `SKILL.md`.
+6. Do not duplicate the same detailed content across `SKILL.md` and `references/`.
+7. Tool-specific metadata is authored in the skill source and copied as-is by the CLI.
+8. Codex and Gemini install to `.agents/skills/<skill-name>`. Claude installs to `.claude/skills/<skill-name>`.
+9. The CLI copies the whole skill directory tree as-is, so any file under `agents/`, `references/`, `assets/`, or `scripts/` is installed unchanged.
+
+## Frontmatter Fields
+
+`SKILL.md` frontmatter is now agent-facing only. The CLI validates the required fields below and ignores extra tool-specific frontmatter fields.
+
+| Field         | Required | Example                                       | Meaning                                     |
+| ------------- | -------- | --------------------------------------------- | ------------------------------------------- |
+| `name`        | Yes      | `graphql-contract`                            | Unique skill name and install directory key |
+| `description` | Yes      | `Use when changing GraphQL schema contracts.` | Discovery/autotrigger summary               |
+
+## Registry Fields
+
+`skill-registry.json` is the install/catalog SSOT.
+
+| Field                 | Required | Example                              | Meaning                                           |
+| --------------------- | -------- | ------------------------------------ | ------------------------------------------------- |
+| `id`                  | Yes      | `graphql-contract`                   | Canonical skill id                                |
+| `kind`                | Yes      | `reference` / `task`                 | Skill category                                    |
+| `supported_tools`     | Yes      | `["claude-code", "codex", "gemini"]` | Where the skill may be installed                  |
+| `install_scopes`      | Yes      | `["project", "user"]`                | Allowed install scopes                            |
+| `groups`              | Yes      | `["frontend-web"]`                   | Display/discovery grouping                        |
+| `included_in_presets` | Yes      | `["frontend-web", "backend-ts"]`     | Presets that surface this skill in `ai-ops init`  |
+| `source_path`         | Yes      | `reference-skills/graphql-contract`  | Relative directory that contains the skill source |
+
+## Content Placement
+
+### Reference Skill
+
+- `SKILL.md`: short routing note
+- `references/reference.md`: full detailed content
+
+### Task Skill
+
+- `SKILL.md`: full actionable procedure
+- `references/`: optional background material
+- `scripts/`: optional executable helpers
+
+## Tool-Specific Invocation Control
+
+If a skill must be explicit-only, the skill author must add the tool-specific metadata directly.
+
+### Codex
+
+Create `agents/openai.yaml` yourself.
+
+```yaml
+allow_implicit_invocation: false
+```
+
+- Use this when Codex should not auto-trigger the skill from the description.
+- Because the CLI copies the full directory tree, `agents/openai.yaml` is installed as-is.
+
+### Claude Code
+
+Add `disable-model-invocation: true` to the `SKILL.md` frontmatter yourself.
+
+```yaml
+---
+name: deploy-script
+description: Runs the deployment workflow.
+disable-model-invocation: true
+---
+```
+
+- Use this when Claude should only load the skill through explicit `/skill-name` invocation.
+- The CLI preserves this extra frontmatter field and does not generate it for you.
+
+### Gemini CLI
+
+Gemini CLI does not currently provide a skill-level frontmatter flag for disabling implicit invocation.
+
+- If you need explicit-only behavior in Gemini, use a custom command instead.
+- Do not add a fake Gemini-only field to `SKILL.md`; Gemini will not interpret it.
+
+## Examples
+
+### Reference Skill Skeleton
+
+```text
+reference-skills/graphql-contract/
+  SKILL.md
+  agents/
+    openai.yaml      # optional
+  references/
+    reference.md
+```
+
+### Task Skill Skeleton
+
+```text
+task-skills/skill-load-check/
+  SKILL.md
+  scripts/
+    loaded.js
+```
+
+## Temporary Validation Skills
+
+`skill-load-check` is an example of a temporary validation skill.
+
+- Keep the description narrow and test-focused
+- Keep the body short and procedural
+- It is acceptable to delete this skill later once the install/load workflow is proven

package/data/skills/reference-skills/ai-llm-python-runtime/SKILL.md

@@ -0,0 +1,10 @@
+---
+name: ai-llm-python-runtime
+description: Use when implementing LLM workflows in Python, especially structured outputs, retries, prompt versioning, and cost controls.
+---
+
+# AI LLM Python Runtime
+
+Use this skill when building Python LLM workflows or runtime integrations.
+
+Read `references/reference.md` before changing prompt storage, structured outputs, retries, or token budgeting.

package/data/skills/reference-skills/ai-llm-python-runtime/references/reference.md

@@ -0,0 +1,26 @@
+# AI LLM Python Runtime
+
+## Constraints
+
+- Do not parse model outputs with regex or string splitting when schema output is required.
+- Do not hardcode prompts inline.
+- Do not call sync SDK methods from async application paths.
+- Do not ignore token limits.
+- Do not log raw responses containing PII.
+
+## Guidelines
+
+- Use Pydantic-based structured outputs or provider-native schema modes.
+- Apply retry with exponential backoff for transient provider failures.
+- Centralize provider routing through one abstraction layer.
+- Track model, tokens, and latency per call.
+- Version prompts and include the prompt version in logs.
+- Stream user-facing responses when possible.
+- Define fallback model chains for rate limits or provider outages.
+
+## Decision Rules
+
+- When output must match a strict schema, use structured output with schema validation.
+- When selecting a model, prefer the smallest model that meets quality requirements.
+- When input exceeds the context window, chunk with overlap and aggregate.
+- When multiple providers are required, route them through one unified abstraction.

package/data/skills/reference-skills/backend-python-fastapi-runtime/SKILL.md

@@ -0,0 +1,10 @@
+---
+name: backend-python-fastapi-runtime
+description: Use when implementing FastAPI routes, dependency injection, Python backend libraries, logging, packaging, or service/runtime boundaries.
+---
+
+# Backend Python FastAPI Runtime
+
+Use this skill for Python backend service work built on FastAPI.
+
+Read `references/reference.md` before changing routers, models, dependency injection, packaging, logging, or test setup.

package/data/skills/reference-skills/backend-python-fastapi-runtime/references/reference.md

@@ -0,0 +1,43 @@
+# Backend Python FastAPI Runtime
+
+## FastAPI Constraints
+
+- Do not use plain dict or TypedDict as request or response models.
+- Do not run blocking I/O in async handlers.
+- Do not return ad-hoc error dicts from handlers.
+- Do not place business logic in routers.
+- Do not hardcode CORS origins.
+
+## FastAPI Guidelines
+
+- Use reusable `Depends()` providers for DB session, auth context, and services.
+- Use `APIRouter` per domain with clear prefixes and tags.
+- Set `response_model` explicitly.
+- Use `Annotated[T, Depends(...)]` for typed dependency injection.
+- Use Pydantic Settings for startup-time env validation.
+- Use lifespan context managers for startup and shutdown.
+
+## Backend Python Library Constraints
+
+- Do not use `requirements.txt` as the primary dependency spec.
+- Do not use `print()` for logging.
+- Do not use `unittest.TestCase`.
+- Do not rely on bare `assert` in production code.
+
+## Backend Python Library Guidelines
+
+- Use uv for dependency and environment management.
+- Use pytest with pytest-asyncio and pytest-cov.
+- Use httpx for sync and async HTTP clients.
+- Use structlog and bind request-scoped context.
+- Use ruff for lint and format.
+- Use mypy or pyright in strict mode.
+- Use pydantic-settings for env loading and validation.
+- Use tenacity for retry and backoff around transient external calls.
+
+## Decision Rules
+
+- When an endpoint is CPU-bound, use `def` and let FastAPI run it in the threadpool.
+- When an endpoint is I/O-bound, use `async def` with async drivers.
+- When shared resources need setup or teardown, use lifespan context managers.
+- When package management is needed, use uv with `pyproject.toml` and a lock file.

package/data/skills/reference-skills/backend-service-standards/SKILL.md

@@ -0,0 +1,10 @@
+---
+name: backend-service-standards
+description: Use when designing backend API contracts, service boundaries, request tracing, idempotency, health checks, or graceful shutdown behavior.
+---
+
+# Backend Service Standards
+
+Use this skill when the work touches backend API contracts, identifiers, timestamps, request tracing, idempotency, health and readiness endpoints, or graceful shutdown.
+
+Read `references/reference.md` and use it as the primary source of truth for detailed standards.

package/data/skills/reference-skills/backend-service-standards/references/reference.md

@@ -0,0 +1,29 @@
+# Backend Service Standards
+
+## Constraints
+
+- Do not use floating-point for money. Use minor-unit integers with ISO 4217 currency.
+- Do not expose sequential ids in external APIs. Use UUIDs, preferably UUID v7.
+- Do not store or transmit timezone-naive timestamps.
+- Do not use magic numbers or strings inline when the value is policy or protocol relevant.
+- Do not return inconsistent error shapes.
+- Do not accept unbounded input at API boundaries.
+
+## Guidelines
+
+- Use UTC end-to-end across API, logs, and persistence.
+- Wrap API responses in a consistent envelope with `data`, `error`, and `meta`.
+- Propagate `X-Request-Id` across gateway, service, logs, DB comments, and outbound calls.
+- Validate environment variables at startup and fail fast.
+- Use domain-specific error codes in `DOMAIN_ACTION_REASON` format.
+- Support `Idempotency-Key` for POST/PATCH where replay safety matters.
+- Expose `GET /health` and `GET /ready` separately.
+- Return empty collections instead of `null`.
+- Handle `SIGTERM` gracefully by stopping intake, draining in-flight work, closing resources, and then exiting.
+
+## Decision Rules
+
+- When a new entity needs a primary key, use UUID v7 and avoid auto-increment ids.
+- When an endpoint returns an error, return the standard error envelope and avoid ad-hoc error fields.
+- When systems exchange timestamps, use ISO 8601 UTC in APIs/logs and timezone-aware DB types.
+- When an API needs versioning, prefer URL versioning over header-only versioning.

package/data/skills/reference-skills/backend-ts-nestjs-runtime/SKILL.md

@@ -0,0 +1,10 @@
+---
+name: backend-ts-nestjs-runtime
+description: Use when implementing NestJS modules, controllers, DTO validation, logging, HTTP clients, or backend TypeScript service structure.
+---
+
+# Backend TS NestJS Runtime
+
+Use this skill for TypeScript backend service work built on NestJS.
+
+Read `references/reference.md` before changing controllers, modules, validation, logging, HTTP integrations, or service boundaries.

package/data/skills/reference-skills/backend-ts-nestjs-runtime/references/reference.md

@@ -0,0 +1,43 @@
+# Backend TS NestJS Runtime
+
+## NestJS Constraints
+
+- Do not use `@Res()` or `@Response()` for normal route handling.
+- Do not put business logic in controllers.
+- Do not use `forwardRef()` as a default circular dependency fix.
+- Do not read `process.env` directly in services or controllers.
+- Do not use `console.log` for app logging.
+
+## NestJS Guidelines
+
+- Organize by feature module.
+- Use DTO validation at request boundaries.
+- Register global validation, filters, and guards in bootstrap.
+- Use custom `HttpException` types for domain failures.
+- Use Swagger decorators for REST contracts.
+
+## Backend TS Library Constraints
+
+- Do not use moment/dayjs.
+- Do not use jsonwebtoken.
+- Do not handle Express req/res directly in NestJS handlers.
+- Do not import lodash as a full bundle.
+- Do not use node-fetch/got in NestJS services.
+- Do not use winston, morgan, or console logging.
+
+## Backend TS Library Guidelines
+
+- Use class-validator and class-transformer DTO validation.
+- Use jose for JWT sign/verify and JWKS workflows.
+- Use pino via nestjs-pino for structured logs.
+- Use rxjs operators in interceptors, guards, and event streams.
+- Use Vitest with @nestjs/testing and supertest.
+- Use zod outside DTO boundaries.
+- Keep TypeScript strict mode enabled.
+
+## Decision Rules
+
+- When auth or roles depend on handler metadata, use guards.
+- When response transformation or timing is needed, use interceptors.
+- When JWT auth is needed, use jose.
+- When structured logging is needed, use pino via nestjs-pino.

package/data/skills/reference-skills/data-pipeline-python-performance/SKILL.md

@@ -0,0 +1,10 @@
+---
+name: data-pipeline-python-performance
+description: Use when building Python data pipelines that need Polars, DuckDB, schema control, or out-of-core execution.
+---
+
+# Data Pipeline Python Performance
+
+Use this skill for Python ETL, analytics pipelines, large dataset transforms, and out-of-core execution work.
+
+Read `references/reference.md` before changing dataframe transforms or data processing architecture.

package/data/skills/reference-skills/data-pipeline-python-performance/references/reference.md

@@ -0,0 +1,24 @@
+# Data Pipeline Python Performance
+
+## Constraints
+
+- Do not iterate DataFrame rows in Python loops for core transformations.
+- Do not use Pandas `.apply(axis=1)` for production-scale transforms.
+- Do not load datasets larger than memory in one shot.
+- Do not rely on implicit dtype inference in production pipelines.
+- Do not mutate DataFrames in place.
+
+## Guidelines
+
+- Prefer Polars for new pipelines, especially lazy mode.
+- Use DuckDB for local SQL analytics on Parquet/CSV and out-of-core workloads.
+- Use streaming or chunked reads for large sources.
+- Write partitioned Parquet outputs for downstream pruning.
+- Enforce explicit schemas at I/O boundaries.
+
+## Decision Rules
+
+- When transforming medium or large tables, prefer Polars lazy pipelines.
+- When ad-hoc SQL analysis on local files is needed, use DuckDB directly on the source files.
+- When data exceeds available memory, use out-of-core execution.
+- When custom per-row logic is unavoidable, use controlled vectorized or mapped APIs rather than general row iteration.

package/data/skills/reference-skills/db-prisma-postgresql/SKILL.md

@@ -0,0 +1,10 @@
+---
+name: db-prisma-postgresql
+description: Use when modifying Prisma schemas, transactions, raw queries, indexes, or PostgreSQL migration safety.
+---
+
+# Prisma PostgreSQL
+
+Use this skill before changing Prisma schema, raw SQL usage, migration flow, or PostgreSQL transaction settings.
+
+Read `references/reference.md` and use it as the detailed persistence guidance.