Choose one of three approaches. The steps for that option appear below.
Choose an option above. On a phone, swipe the tabs left or right to see all options →
Cursor scopes sources and drafts rules; you approve the plan, check citations and reason through known rows without execution.
Collect the dictionary and rule
Propose source files and output
Check permission
Draft constraints, cited rules
Check source lines
Use: You
Have ready: line-numbered data dictionary, team rule-format sample and known test rows in an isolated folder; leave originals unchanged.
Use: Cursor
Ask Cursor to locate only the specified sources, propose scoped files and output paths, and create an isolated branch without modifying sources.
What to paste into the tool
Task: Proposed source packet, output paths and isolated draft branch. Search the isolated folder for the data dictionary, the rule-format sample and the known test rows. List the exact files you would read and the output file paths you would write. Create an isolated draft branch. Do not modify any source file, run production pipelines, install packages, push or access secrets.
Use: You
Approve repository permission, exact source files, output paths and allowed read-only commands.
Check: Confirm repository permission and tool plan. No remote push, package install, secret access or production run. Preserve source files Confirm the proposed scope names the exact source files and output paths before any drafting.
Use: Cursor
Ask Cursor to inspect the approved sources and draft constraints, team-format rules, line citations, candidate test rows and vague-definition gaps.
What to paste into the tool
Read the named dictionary and format sample. Quote the exact dictionary line for every rule. Mark vague or absent constraints as gaps; do not invent defaults. Draft rules and candidate test rows in new files. Do not run production pipelines, install packages, push or modify sources.
Use: You
Check every rule against its dictionary line and team format; manually reason through five known good and five known bad rows.
Steward performs the manual known-row checks. This workflow does not run rules, tests or production pipelines. Any later execution needs separate approval. The tool does not establish completeness or correctness. It ends at a reviewed draft in an isolated copy, not a merge or production change.
ChatGPT extracts constraints, then formats your approved table; you check citations, disputed definitions and candidate rows without running rules.
Collect the data dictionary
Draft constraints with exact dictionary
Check each constraint
Draft rules and candidate test
Check coverage
Use: You
Have ready: data dictionary, team rule format and known test rows.
Use: ChatGPT
Ask ChatGPT for a field-by-field constraint table quoting exact dictionary lines, plus vague or missing definitions.
What to paste into the tool
Task: Draft constraint table: field, constraint, exact dictionary line; Draft gaps list of vague lines. Read this data dictionary. For every field, extract each constraint as a row: field, constraint, and the exact dictionary line it comes from, quoted word for word. List vague or missing definitions as gaps. Do not invent constraints and do not soften vague ones.
Use: You
Correct or remove wrong rows using past definition disputes; approve the constraint table and gaps list.
Every approved constraint cites its dictionary line; vague lines go on the gaps list Check every extracted constraint against its dictionary line before approving the table.
Use: ChatGPT
Ask ChatGPT to convert your approved constraints into the team rule format with citations and candidate known-good and known-bad rows.
What to paste into the tool
Here is my approved constraint table with dictionary-line citations. Convert each constraint into a rule in our rule format, keeping the exact line citation on each rule. Do not add constraints I did not list and do not soften vague ones - my gaps list stays as is. Then draft candidate test rows for each rule: rows that should pass and rows that should fail.
Use: You
Check rules against five known good and five known bad rows; keep the final cited rule list and gaps list.
Check the rules against five known good and five known bad rows; confirm each dictionary constraint has a rule or is listed as untestable; rules are not run by the workflow the rules are not run by this workflow.
ChatGPT drafts; a fresh Claude session critiques citations. You reconcile the two and check known rows without running rules.
Collect the data dictionary
Draft constraints, rules and gaps
Check omissions
List citation failures and candidate
Reconcile critiques
Use: You
Have ready: data dictionary, team rule format and known test rows.
Use: ChatGPT
Ask ChatGPT for the constraint table, team-format rule list and gaps list, all with dictionary-line citations.
What to paste into the tool
Here is our data dictionary. Extract every constraint per field with the exact dictionary line quoted, convert each into a rule in our rule format keeping the line citation, and list vague or untestable lines on a gaps list. Quote the dictionary line for every rule. Do not run anything.
Use: You
Skim for obvious omissions; move the dictionary and draft into a fresh Claude session without the drafter's reasoning.
Keep the two sessions separate so the critic reads the draft without the drafter's reasoning. Skim for omissions before the handoff.
Use: Claude
Ask Claude to check citations independently, listing unsupported rules, missing constraints, uncited rules and gaps-list errors, with candidate test rows.
What to paste into the tool
Here is a data dictionary and a draft rule set derived from it. Independently verify: for each rule, does the quoted dictionary line actually support it? List rules whose citation does not support them, dictionary constraints with no rule, rules with no citation, and vague lines wrongly treated as testable. Then draft candidate good and bad test rows for the rules that survive.
Use: You
Reconcile the critique with the draft; check five known good and five known bad rows; retain the cited final rules and gaps.
Check the final rules against five known good and five known bad rows from the real data (source check); confirm each dictionary constraint has a rule or is listed as untestable; rules are not run by the workflow (source warning). Reconcile the critique with the draft before finalising.
Best when: You want repository-local source search and isolated draft files.
What it does: Cursor scopes and drafts; you approve scope and check rules
Strong at: Reads scoped repository files and drafts rules with source-line references.
Limit: No rule execution; later runs need separate approval.
Free plan: Yes, limited agent features
Best when: You want extraction drafted before you settle disputed definitions.
What it does: ChatGPT extracts and formats; you check constraints and known rows
Strong at: Extracts cited constraints and converts an approved table into rule format.
Limit: Rules are not run; you must check coverage and known rows.
Free plan: Yes, limited uploads and data analysis
Best when: You want a separate-context check of dictionary citations.
What it does: ChatGPT drafts, Claude critiques; you reconcile and check known rows
Strong at: Claude critiques ChatGPT's draft without seeing the drafter's reasoning.
Limit: Both sessions rely on the same supplied dictionary; rules remain unrun.
Free plan: ChatGPT: Yes, limited uploads and data analysis; Claude: Yes, limited usage windows
Before you paste into any AI tool, remove any confidential information. Put it back at the end.
Deciding between ChatGPT and Claude for this? See the comparison.
Every rule cites the exact dictionary line it comes from (source constraint).
Each dictionary constraint has a rule or is listed as untestable on the gaps list.
The steward checks the rules against five known good and five known bad rows.
Rules are not run by the workflow.
Check the result against your original inputs. Approve it before sharing or acting.