Start with bounded context, not the whole repository
A naming agent rarely needs build output, dependencies, binary files, credentials, or every source file. Begin with the README and the few product documents that explain audience, pain, outcome, mechanism, tone, constraints, and future scope. Preview exactly what will leave the machine.
NamingSignal's local CLI and stdio MCP workflow are designed around this boundary. Detected secrets and obvious credential files are excluded, and the approved brief becomes the portable input to hosted generation and research.
Ask the agent to inspect before it generates
The first output should be an editable naming brief, not a list of names. Tell the agent to identify uncertainties and contradictions, preserve avoided brands and roots, and ask for a material correction only when the missing choice would change the work.
- What is the product in one precise sentence?
- Who chooses it, and what painful job makes them switch?
- Which mechanisms or differentiators are genuinely name-worthy?
- Which channels make spelling, pronunciation, or package collisions important?
- How far might the product expand beyond the current feature?
Use one key, but not one permanent key
Create a separate user-owned API key for each device, agent, or CI environment and store it in NAMING_SIGNAL_API_KEY. The user signs in through Clerk to create and revoke the key; the local agent sends it as an Authorization Bearer token. Your Clerk secret key is never distributed to users or agents.
Separate keys make incident response simple: revoke the affected client without signing every other tool out. Usage is charged to the same monthly account credit ledger regardless of whether the request came from a browser session or a user key.
Run the sequence, not a one-shot prompt
Compile and confirm the brief. Generate a broad field in distinct veins. Check the selected domain endings in bounded batches. Rank survivors without hiding registered or needs-review leaders. Research only one to three finalists, then export the full record with sources, timestamps, fallback labels, and remaining manual actions.
The agent should never convert a timeout, unsupported TLD, failed calibration, ambiguous namesake, or empty search response into availability. Those are unresolved states for a person or a later retry.
A ready-to-paste instruction
The public developer documentation publishes a current ready-to-paste setup instruction for Codex or Claude. It points the agent to the OpenAPI contract and agent guide, tells it where the user-owned key belongs, and requires a small smoke test before a full naming sprint.