Instructions
- Confirm the target. This skill is for the
rowboatlabs/rowboatapp, not an arbitrary knowledge graph. Pin repository facts to the commit inreferences/source.json; this does not establish current runtime behavior. - Classify the work surface: app install/config; source synchronization; graph build/curation; or a bug in a UI/core module. Inspect the relevant pinned files and current checkout before giving commands.
- Use the app's actual model. Rowboat ingests work context (including connected email/meetings and assistant conversations) into a Markdown, backlinked knowledge base. The graph builder tracks source-file state and processes one source file per agent batch to reduce cross-file entity contamination; it is not evidence of an independent graph database.
- Diagnose before changing state. For stale/missing notes, check source sync completion, changed-file detection, graph state, and the note destination before proposing rebuild or re-import. Preserve source files and state unless the user approves a destructive reset.
- Check source-grounded implementation details. Follow
references/upstream.mdfor exact pinned setup/configuration, graph lifecycle, operational checks, and limitations. Treat commands there as documentation, not permission to execute. - Gate effects separately. Ask for approval before downloading/installing the app or dependencies, changing configuration or login state, connecting accounts, importing/private-email or meeting data, accessing cloud resources, writing to a database, starting a public service, or paid uploads. Read-only local inspection needs no such effect approval.
- Verify claims honestly. The offline preflight helper can compare listed public-file hashes only. Report
runtime_verified: falseas no runtime proof; never claim installation, account connection, synchronization, or graph correctness from source hashes alone. - Link
references/source.json,references/upstream.md, andevals/evals.jsonwhen deeper provenance or edge cases matter. Authored eval scenarios are not executed test evidence.
Examples
- "In rowboat, new meeting notes aren't appearing in the knowledge graph. Trace the sync/build path without changing my files." → Inspect state and sync/build boundaries read-only; do not re-import automatically.
- "Set up Rowboat and connect my Gmail." → Explain the pinned setup path, then gate download/install and account authorization/import separately.
- "Design a Neo4j ontology for a research archive." → Route to ontology/database design, not this application skill.
Best practices
- Keep app workspace identity and owner identity explicit; do not infer the user's identity from email headers.
- Preserve source provenance and distinguish a Markdown note graph from catalog metadata, vector search, or relational storage.
- Use the current checkout's package scripts and configuration as truth when they differ from this pinned snapshot.
- For troubleshooting, state the last verified layer and the next read-only check before proposing a write.
References
- Pinned operational guide
- Pinned source inventory and hashes
- Evaluation scenarios
- Offline source preflight (parent-provided; source hashes are not runtime verification)