diff --git a/instructions/aws-appsync.instructions.md b/instructions/aws-appsync.instructions.md
new file mode 100644
index 00000000..40a2bec5
--- /dev/null
+++ b/instructions/aws-appsync.instructions.md
@@ -0,0 +1,190 @@
+---
+description: 'Production-grade guidance for AWS AppSync Event API handlers using APPSYNC_JS runtime restrictions, utilities, modules, and datasource patterns'
+applyTo: '**/*.{graphql,gql,vtl,ts,js,mjs,cjs,json,yml,yaml}'
+---
+
+# AWS AppSync Event API Instructions
+
+Use these instructions when implementing AWS AppSync **Event API** handlers (`onPublish`, `onSubscribe`) with the `APPSYNC_JS` runtime.
+
+## Scope And Contract
+
+- Design handlers around channel namespace flow: `onPublish` runs before broadcast, `onSubscribe` runs on subscription attempts.
+- Keep event contracts explicit and stable. Treat channel path and payload shape as API contracts.
+- Prefer additive changes for payload fields and avoid breaking existing subscribers.
+
+## Data Sources Map (Event API)
+
+Use data sources intentionally based on event workflow needs:
+
+- Lambda: custom compute, transformation, orchestration, external AWS/service integrations.
+- DynamoDB: low-latency event/state persistence and key-based reads/writes.
+- RDS (Aurora): relational checks, joins, and stronger relational integrity use cases.
+- EventBridge: route events into broader event-driven architectures.
+- OpenSearch: search and analytics over event data.
+- HTTP endpoints: external APIs or AWS service APIs over HTTP.
+- Bedrock: model inference and AI enrichment in event pipelines.
+
+Prefer combining multiple data sources only when each hop has a clear reason (auth, persistence, enrichment, routing).
+
+## Data Source Setup And IAM (Required)
+
+- Create data sources at the Event API level, then attach them as namespace integrations.
+- If using a service role, grant only required actions (least privilege).
+- Trust policy principal must allow `appsync.amazonaws.com` to assume the role.
+- Restrict trust with conditions when possible:
+  - `aws:SourceAccount` to your account.
+  - `aws:SourceArn` to a specific AppSync API ARN (or tightly scoped pattern).
+- Do not reuse broad, cross-service IAM roles for AppSync data source access.
+
+## Runtime Restrictions (Must Follow)
+
+The `APPSYNC_JS` runtime is a constrained JavaScript subset. Write code for this environment, not for full Node.js.
+
+- Do not use async patterns: no promises, `async/await`, or background async workflows.
+- Do not use unsupported statements/operators: `try/catch/finally`, `throw`, `while`, C-style `for(;;)`, `continue`, labels, unsupported unary operators.
+- Do not rely on network or file system access from runtime code. Use AppSync data sources for I/O.
+- Do not use recursion or pass functions as function arguments.
+- Do not rely on classes or advanced runtime features outside documented support.
+- Prefer `for-of` / `for-in` loops when iteration is needed.
+
+## Handler Flow Patterns
+
+- For handlers without data source integration, return transformed `ctx.events` directly.
+- For handlers with data sources, use object form with `request(ctx)` and `response(ctx)`.
+- Use `runtime.earlyReturn(...)` when business logic decides to skip data source invocation and response mapping.
+- Use `ctx.info.channel.path`, `ctx.info.channel.segments`, `ctx.info.channelNamespace.name`, and `ctx.info.operation` to drive routing logic.
+- For `onPublish` with data source integration, return the event list to broadcast from `response(ctx)`.
+- For `onSubscribe` with data source integration, include a `response(ctx)` function (it can be empty when no follow-up mapping is needed).
+
+### `ctx.prev.result` vs `ctx.stash` (Pipeline Guidance)
+
+- If resolver/functions execute step-by-step and the next step depends on the previous step output, use `ctx.prev.result`.
+- Use `ctx.prev.result` as the default data handoff mechanism between consecutive pipeline functions.
+- Use `ctx.stash` when you need shared data across multiple pipeline stages that is not just the immediate previous result.
+- Store only small, intentional metadata in `ctx.stash` (for example flags, IDs, correlation context), not large payload copies.
+- Do not duplicate full previous results into `ctx.stash` when `ctx.prev.result` already provides the needed value.
+
+## Error And Authorization Flow
+
+- Do not use `throw` in handlers. Use `util.error(...)` and `util.appendError(...)` patterns supported by AppSync runtime.
+- For publish failures, return explicit runtime errors with safe messages (no internals).
+- For business-level authorization rejection at handler level, use the documented unauthorized utility in handler code.
+- Keep error payloads non-sensitive. Never expose secrets, raw stack traces, or internal identifiers.
+
+## Built-In Utilities
+
+Use `util` for runtime-safe helpers.
+
+- Encoding utilities:
+  - `util.urlEncode`, `util.urlDecode`
+  - `util.base64Encode`, `util.base64Decode`
+- Runtime utility:
+  - `runtime.earlyReturn(obj)` to stop current handler execution and skip data source + response evaluation.
+
+## Built-In Modules
+
+Use official modules from `@aws-appsync/utils` and keep code declarative.
+
+- DynamoDB module import:
+  - `import * as ddb from '@aws-appsync/utils/dynamodb'`
+- RDS module import:
+  - `import { ... } from '@aws-appsync/utils/rds'`
+
+### DynamoDB Usage
+
+Prefer module helpers over handwritten request objects where possible.
+
+- Core helpers include: `get`, `put`, `remove`, `update`, `query`, `scan`, `sync`.
+- Batch helpers: `batchGet`, `batchPut`, `batchDelete`.
+- Transaction helpers: `transactGet`, `transactWrite`.
+- For `update`, prefer operation helpers like increment/append/add/remove for safe patch-style mutations.
+- Model keys and indexes for query-first access. Avoid `scan` unless justified.
+- Use conditions for correctness and optimistic concurrency when needed.
+- For bursty publish flows, prefer `batchPut`/`batchDelete` (or `transactWrite` when atomicity is required) over many single-item operations.
+- Keep DynamoDB batch sizes within service/API limits and chunk inputs deterministically.
+
+### Lambda Usage
+
+For Event API Lambda data source requests, use:
+
+- `operation: 'Invoke'`
+- optional `invocationType: 'RequestResponse' | 'Event'`
+- `payload` shaped explicitly for the Lambda contract
+
+Guidance:
+
+- Use `RequestResponse` when handler flow depends on Lambda output.
+- Use `Event` only for fire-and-forget side effects.
+- Validate `ctx.result` in `response(ctx)` and map to the exact outgoing event shape.
+- In Event API handlers, Lambda operation support is `Invoke`; do not rely on GraphQL-style `BatchInvoke` here.
+- If you need batching with Lambda in Event API flows, send an array payload in one `Invoke` and implement item-level aggregation/partial-failure handling inside Lambda.
+
+### Direct Lambda Integration (No Handler Code)
+
+You can configure namespace handlers with direct Lambda integrations (`Behavior: DIRECT`) instead of writing `onPublish`/`onSubscribe` code.
+
+- `REQUEST_RESPONSE` mode:
+  - `onPublish` Lambda returns `{ events?: OutgoingEvent[], error?: string }`.
+  - `onSubscribe` Lambda returns `null` for success or `{ error: string }` for rejection.
+- `EVENT` mode:
+  - Invocation is asynchronous; AppSync does not wait for a Lambda response.
+  - For publish, events continue broadcasting as usual.
+- If Lambda returns `error` in request/response mode, it is logged when logging is enabled, and not sent back as a detailed internal error payload.
+
+Prefer direct Lambda integration when the entire namespace behavior can be centralized in Lambda and you do not need APPSYNC_JS request/response mapping logic.
+
+### HTTP/EventBridge/RDS/OpenSearch/Bedrock
+
+When using non-DynamoDB data sources:
+
+- HTTP: return `resourcePath`, `method`, optional `params` (`headers`, `query`, `body`); check `ctx.result.statusCode`, `ctx.result.body`, and `ctx.error`.
+- EventBridge: use `operation: 'PutEvents'` and build deterministic event entries from `ctx.events`.
+- RDS: prefer SQL helpers and `createPgStatement`/`createMySQLStatement`; do not interpolate unsafe SQL.
+- OpenSearch: keep request path/params explicit and map only required fields from `ctx.result`.
+- Bedrock: define `operation` (`InvokeModel` or `Converse`) explicitly and include prompt-injection safeguards.
+
+## Batch Operations (Required Guidance)
+
+- Prefer batching where the target data source natively supports it and event semantics allow grouping.
+- DynamoDB:
+  - Use `batchGet`, `batchPut`, `batchDelete` for non-atomic bulk operations.
+  - Use `transactGet`, `transactWrite` when atomic all-or-nothing behavior is required.
+  - Validate and cap per-request item counts; chunk large batches.
+- Lambda:
+  - Event API JS handler requests use `operation: 'Invoke'` with optional `invocationType`.
+  - There is no Event API `BatchInvoke` operation in handler request objects.
+  - For pseudo-batch Lambda patterns, send list payloads to one invoke and return deterministic per-item result structures.
+- Keep ordering guarantees explicit: if downstream consumers depend on order, preserve and document ordering keys.
+
+## Security And Data Safety
+
+- Treat `ctx.identity`, headers, and payload fields as untrusted input.
+- Enforce least-privilege IAM per data source.
+- Add validation before write operations and before forwarding transformed events.
+- Never hardcode secrets in handler code.
+- For public usage, keep defaults conservative (deny/unauthorized on invalid states).
+
+## Tooling, TypeScript, And Build
+
+- Use `@aws-appsync/eslint-plugin` (`plugin:@aws-appsync/base` at minimum).
+- Use `plugin:@aws-appsync/recommended` when TypeScript tooling is configured.
+- TypeScript is not executed directly by AppSync runtime. Transpile to supported JavaScript before deployment.
+- Bundle with externalized `@aws-appsync/utils` imports and source maps for debugging.
+
+## Observability And Operations
+
+- Enable CloudWatch logging for handlers and datasource integration.
+- Log with structured, low-cardinality fields (channel namespace/path, operation, request id).
+- Add alarmable signals: handler errors, datasource errors, latency regression.
+- Keep response transformations deterministic and test with multi-event payloads.
+
+## Minimum Quality Checklist
+
+- [ ] Uses only APPSYNC_JS-supported runtime features.
+- [ ] No `throw`, no async/promise usage, no unsupported loop/control constructs.
+- [ ] Error flow uses runtime-supported utilities and returns non-sensitive messages.
+- [ ] `onPublish` and `onSubscribe` behavior is explicit and tested.
+- [ ] Data source request/response mapping is deterministic and schema-safe.
+- [ ] Lambda/DynamoDB contracts are documented and validated.
+- [ ] Linting with `@aws-appsync/eslint-plugin` is enabled.