AI & agents

Converting reports with an assistant

Use the welldot-converter Agent Skill to turn water well reports in any language into validated .well files with Claude, ChatGPT, Gemini, Cursor or Codex.

welldot-converter is an Agent Skill: a written procedure that an AI assistant loads and follows. It turns a water well report (PDF, Word document, scan, photo or text, in Portuguese, English, Spanish or another language) into a .well file, then reports what it found and what it left out.

The skill lives in the repository at .agents/skills/welldot-converter/. SKILL.md is the procedure, and references/well-spec.md is an offline copy of the spec for assistants that cannot browse.

What it does

  1. Gets the spec. Reads the current specification, or the bundled copy when offline.
  2. Reads the report, with the assistant's own file and vision tools.
  3. Extracts the well: identity, location, construction, lithology, tests, analyses, operation, permits and history, section by section (what goes where).
  4. Validates the result: required fields, intervals, units, datetimes with offsets, closed vocabularies.
  5. Delivers the file with a summary of the sections found, the fields left out on purpose and the free-text values it kept as written.
  6. Asks about ambiguities instead of guessing.

Its rules match the conversion guide: never estimate a number, keep the report's wording, convert units exactly, write datetimes with an offset, keep the drilled depth out of well_depth.

Set it up

Clone the repository, or copy .agents/skills/welldot-converter/ into your own project. These clients discover skills in .agents/skills/ (Claude Code through the .claude/skills/ link) with no configuration. The assistant activates the skill when you ask it to convert a report.

Close the loop with validation

Connect the same assistant to the docs MCP server. The skill tells it to run validate_well on every file it produces and to fix each reported error before handing the file over. With MCP connected, it can also look up a field in the spec with search_docs instead of relying on memory.

Review the result

The assistant's summary tells you where to look. Then:

  • Open the file in welldot.org and compare the drawn profile with the report's drawing.
  • Check the numbers in the tables: depths, diameters, levels, rates.
  • Read the fields it says it left out, and fill them only from the report.

The review checklist has the full list.

Conversion quality depends on the assistant and on the scan. Handwritten logs, low-resolution scans and tables split across pages cause most errors. Treat every converted file as a draft until a person has checked it.