Fernhill Claude connector kit

Companion: https://quickchat.ai/post/claude-connectors-explained
Fictional demonstration data, prepared 30 October 2026.

These are editable reference files, not an import format. No account credential or installable connector URL is included. Use your own Agent URL from Connect MCP.

1. Create an Agent named Fernhill Help. Put main-prompt.txt in Main Prompt.
2. Add fernhill-help-center.txt as a Knowledge Base article and wait for Ready.
3. Keep custom Actions, Human Handoff and Smart Data Gathering off.
4. In Testing, choose Create Dataset > Paste messages. Paste test-messages.txt including the standalone --- separators and select Create. Inside the dataset, expand Evaluation criteria, paste evaluation-criteria.txt and Save. Select Run > Start Run to check the three separate cases.
5. In External Apps > MCP, enable MCP, keep Require sign-in (private) on, and copy mcp-settings.json and mcp-description.txt into the corresponding fields.
6. Copy your actual Connect MCP URL. In Claude, open Settings > Connectors > + Add > Add custom connector. Enter Fernhill Help and the URL, then Continue. Keep Sign in now and Register automatically, leave request headers empty, and select Add. On the next screen select Connect and authorize with an account that can access the Agent.
7. Check for Disconnect and Ask Fernhill Help. Keep Needs approval for the first tests. Start a new chat and enable Fernhill Help in + > Connectors.
8. Send claude-test-instruction.txt, then the three questions separately. Use Allow once if Claude asks for tool permission. Compare the completed Claude answer with the matching Quickchat AI Inbox question and Agent reply.
9. Fill evidence-template.csv only with your actual results. The blank rows are intentionally not marked as passing. Local Agent answers and Claude's final wording are separate checks.

The third question deliberately asks about an undocumented destination. The expected answer is uncertainty, not an invented yes or no. The return answer must retain the original-tags condition and the refund clock starting at receipt. Generated replies and Simulations use the Agent owner's AI credits.

RECORDED RESULTS

local-receipts/ records our local demonstration: the original missing-fact failure, revised prompt, complete-reference evaluation run, private anonymous rejection and public tool discovery. These historical observations have UTC timestamps. None is a Claude call or a completed OAuth grant. The final local configuration was restored to private.

observed-results/hosted-agent.json records the actual hosted Agent-side questions/replies and read-only configuration comparison. observed-results/claude-comparison.json pairs these with the three completed Claude answers transcribed from the author's screenshots. All three questions reached the Agent unchanged in one Claude conversation.

The warranty and Switzerland replies preserve the relevant facts. The return reply adds that the original tags must be attached. Neither the source nor the Agent reply says that. This case is deliberately recorded as added_policy_condition, not a full pass. suggested-return-recheck.txt is a follow-up prompt to try; it was not tested in the captured run and is not a verified fix. Keep the source unchanged unless the real policy owner confirms a correction.

The phone image shows the same conversation on iPhone. No separate mobile or Claude Code call was observed. The article provides those as optional extensions.

The tuning/ folder preserves the exact initial Main Prompt and incomplete evaluation criteria from the first local runs. Use the top-level files for the finished configuration; the initial files are only for reproducing the comparison.
