diff --git a/.agents/SKILLS.md b/.agents/SKILLS.md
new file mode 100644
index 0000000..3e7f777
--- /dev/null
+++ b/.agents/SKILLS.md
@@ -0,0 +1,90 @@
+# SwiftBox Agent Instructions
+
+## What SwiftBox Is
+
+SwiftBox is a cross-platform automation runtime and workflow engine. It is a dependency, not a product. Other products depend on it for safe system automation.
+
+It is not the UI. It is not the business logic of any single app. It is the shared execution layer underneath multiple products.
+
+## What SwiftBox Does
+
+SwiftBox turns a situation into a safe action loop:
+
+```
+input → normalize → decide → act → state → verify
+```
+
+It reads config and state, matches situations to workflows, checks permissions, executes through adapters, verifies results, and writes logs.
+
+## Repository Map
+
+| Path | Purpose |
+|------|---------|
+| `config/hosts/` | Per-host identity and execution settings |
+| `config/healthchecks/` | What to check on a given system |
+| `config/workflows.yaml` | What to do when checks fail |
+| `config/permissions.yml` | What actions are allowed and by whom |
+| `config/default.yml` | Global fallback values |
+| `core/schemas.py` | All internal data structures |
+| `core/detect.py` | Run checks, produce IssueReports |
+| `core/plan.py` | Map issues to ActionRequests |
+| `core/verify.py` | Confirm action outcomes |
+| `core/engine.py` | Coordinate the full loop |
+| `adapters/linux/` | Disk, network, process, service checks |
+| `adapters/notify/` | stdout and filelog output |
+| `adapters/repo/` | Fetch and sync scripts from trusted remote |
+| `scripts/bootstrap/` | Initial environment setup |
+| `scripts/diagnose/` | Read-only inspection scripts |
+| `scripts/fix/` | Repair actions |
+| `state/` | Persistent memory between runs |
+| `logs/` | Append-only operational logs |
+| `cli/swiftbox.py` | CLI entrypoint |
+| `tests/` | Behavior contracts |
+
+## What an Agent May Do
+
+- Read any file in this repo.
+- Add config entries to `config/` files.
+- Add new workflows to `config/workflows.yaml`.
+- Add new scripts to `scripts/`.
+- Add new adapters to `adapters/`.
+- Add tests to `tests/`.
+- Modify `core/` files when a clear functional gap exists.
+- Update `config/permissions.yml` to add new roles or allowed actions.
+
+## What an Agent Must Not Do
+
+- Add business logic, UI logic, or product-specific behavior to `core/`.
+- Hardcode values that belong in `config/`.
+- Run destructive actions without a dry-run check.
+- Delete or truncate `state/history.jsonl` or any log file.
+- Add actions to `permanently_blocked` bypass paths in code.
+- Modify more than one layer at a time without verifying tests still pass.
+- Make silent changes to `config/permissions.yml` that expand allowed actions.
+
+## How to Choose the Next Action
+
+1. Read the current structure — do not assume file contents.
+2. Identify the smallest change that solves the problem.
+3. Check if the change belongs in `config/`, `adapters/`, `scripts/`, or `core/`. In that order of preference.
+4. If modifying `core/`, locate all callers and dependencies first.
+5. Write or update tests before or alongside the change.
+6. Verify the full test suite passes: `pytest tests/ -v`.
+
+## How to Verify Success
+
+- `pytest tests/ -v` must pass with no failures.
+- Dry-run the affected workflow and confirm the log output is correct.
+- Check `state/last_actions.json` after a run to confirm the record is written.
+
+## Key Design Constraint
+
+**Config decides behavior. Code executes behavior.**
+
+If you are adding a rule, a threshold, a limit, or a policy — it goes in `config/`, not in `core/`.
+
+If you are adding platform-specific behavior — it goes in `adapters/`, not in `core/`.
+
+If you are adding a system action — it goes in `scripts/`, exposed via a workflow in `config/workflows.yaml`.
+
+`core/` should remain small, deterministic, and testable.
\ No newline at end of file