Translate and deliver

Step 1: Draft the findings summary

Direct AI to draft a findings summary for Budi. Before it writes, give it constraints:

Draft a findings summary for Budi. He's a shrimp farmer, not a statistician. No jargon. Answer his three questions: (1) which water quality parameters matter most, (2) does sensor data explain production differences, (3) should he install sensors in the other five ponds. Be honest about what the data can and cannot support.

Review the draft. Check that the language is practical, that uncertainty is communicated honestly, and that Budi would understand every sentence.

Step 2: Edit for Budi's context

AI's first draft will likely be too formal or too technical in places. Edit it.

Budi knows dissolved oxygen matters -- he has lost shrimp after heavy rain when the aerators could not keep up. Use language that connects to his experience. "The data shows that dissolved oxygen dips are the strongest signal for harvest problems" means more to him than "minimum DO per cycle exhibits the highest Pearson correlation with survival rate."

The summary should answer three questions clearly:

• Which parameters matter most? (Dissolved oxygen, especially minimum values during a cycle)
• Does water quality explain production differences? (The patterns suggest yes, but three ponds is too few to be certain)
• Should he install more sensors? (The evidence supports it, with a qualification about the confound)

Step 3: Send to Budi

Send the findings summary to Budi. Let the conversation unfold.

He may ask a follow-up about feed data or about specific ponds. Address the scope extension explicitly. The feed analysis is a reasonable addition using the same data infrastructure, or it could be a separate engagement. Either decision is valid -- what matters is that you make it deliberately rather than letting scope drift.

Step 4: Address the scope extension

Budi asks about feed conversion ratios. Decide: include it as a small addition (the data is already in the production records) or flag it as separate work. Both approaches are professional. Make the choice, explain your reasoning to Budi, and proceed.

Step 5: Final commit and push

Direct AI to finalize the project:

Create a final commit with a descriptive message summarizing the analysis. Push to GitHub. Update the README with a project summary.

The project repository should contain: the DuckDB database, the analysis code, the findings summary, and the data dictionary.

Step 6: Reflect on the MCP experience

This project started with you describing data to AI and ended with AI querying a database directly. The shift was not incremental -- it was categorical. AI went from "working with what you told it" to "working with what it can read."

That shift introduced a new verification responsibility. You checked the SQL queries, the join logic, the temporal alignment, the sensor calibration. Verifying the data path -- not just the statistical output -- is the discipline this project adds to your practice.

MCP is a standard. The same protocol works regardless of which AI coding agent connects to the database. The specific configuration syntax for Claude Code is a lookup. The understanding -- that tool connectivity changes what AI can do, and that connected AI requires data path verification -- transfers everywhere.