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 →
Claude proposes the hypothesis table, drafts the experiment plan and cross-checks it. You check both gates and decide.
Assemble the materials
Build hypothesis table
Check the table
Draft and cross-check
Check and decide
Use: You
Assemble the approved ad concepts, baseline aggregate data, supplied traffic constraints and decision objective, retaining authorised source IDs and dates.
Check each item is the approved version, dated and permitted; list anything missing as still to confirm.
Use: Claude
Ask Claude for a hypothesis table: the main measure, what could skew the result and the comparisons that can be run, with source references.
What to paste into the tool
Task: Hypothesis table specifying the main measure, what could skew the result and the comparisons that can be run; source references; uncertain or absent facts; Supplied inputs: Approved ad concepts, baseline aggregate data, supplied traffic constraints and decision objective. Use only the supplied packet. Build a labeled table with source IDs, dates, uncertainty and permitted-use limits on each row. Check the intended message contrast, the main measure, traffic limits, what could skew the result and whether the plan is feasible against approved concepts and baseline aggregate data. Retain the source packet as an unchanged attachment in the output bundle. Keep unresolved facts and approval dependencies explicit. Do not resolve a missing fact by inventing one. Produce a draft only; do not publish, send, deploy, change accounts, spend or launch a campaign
Use: You
Compare every table row with the source materials and correct it; resolve or retain missing facts before drafting.
Check the message contrast, main measure, traffic limits, skew factors and feasibility against the approved concepts and baseline data.
Use: Claude
Ask Claude to draft the experiment plan with the allocation proposal and stopping rules, then cross-check it against the checked table and correct it.
What to paste into the tool
Task: Draft ad experiment plan with allocation proposal, stopping rules for qualified review and no launch; Open questions and approval dependencies retained beside the recommendations; Reviewer-checked hypothesis table specifying the main measure, what could skew the result and the comparisons that can be run; source references; uncertain or absent facts; Supplied constraints, retained open questions and approval dependencies with source IDs and dates; Supplied inputs: Approved ad concepts, baseline aggregate data, supplied traffic constraints and decision objective. Propose one prospective message contrast and primary decision, the main measure, what could skew the result, allocation and stopping assumptions for qualified analyst review; do not launch. Retain the checked table, constraints and source packet unchanged in the output bundle. Keep unresolved facts and approval dependencies explicit. Do not resolve a missing fact by inventing one. Produce a draft only; do not publish, send, deploy, change accounts, spend or launch a campaign. Second message, after your check: Task: Corrected draft with a change list and unresolved mismatches flagged for the reviewer. Compare the draft with the checked hypothesis table and the retained source packet. Correct every mismatch you find in the draft, keep source IDs, dates, uncertainty and named exceptions, and list each change. Flag anything you cannot settle from the supplied material; do not invent facts, figures, policies or claims. Return preparation material only; publishing, outreach, audience changes, tracking deployment and financial commitments are outside this task.
Use: You
Compare the corrected draft and change list with the checked table, confirm the assumptions, then decide; launching stays outside this workflow.
Check isolation and statistical assumptions get analyst approval. If a check fails, correct the draft or leave the workflow incomplete.
ChatGPT proposes the hypothesis table, drafts the experiment plan and cross-checks it. You check both gates and decide.
Assemble the materials
Build hypothesis table
Check the table
Draft and cross-check
Check and decide
Use: You
Assemble the approved ad concepts, baseline aggregate data, supplied traffic constraints and decision objective, retaining authorised source IDs and dates.
Check each item is the approved version, dated and permitted; list anything missing as still to confirm.
Use: ChatGPT
Ask ChatGPT for a hypothesis table: the main measure, what could skew the result and the comparisons that can be run, with source references.
What to paste into the tool
Task: Hypothesis table specifying the main measure, what could skew the result and the comparisons that can be run; source references; uncertain or absent facts; Supplied inputs: Approved ad concepts, baseline aggregate data, supplied traffic constraints and decision objective. Use only the supplied packet. Build a labeled table with source IDs, dates, uncertainty and permitted-use limits on each row. Check the intended message contrast, the main measure, traffic limits, what could skew the result and whether the plan is feasible against approved concepts and baseline aggregate data. Retain the source packet as an unchanged attachment in the output bundle. Keep unresolved facts and approval dependencies explicit. Do not resolve a missing fact by inventing one. Produce a draft only; do not publish, send, deploy, change accounts, spend or launch a campaign
Use: You
Compare every table row with the source materials and correct it; resolve or retain missing facts before drafting.
Check the message contrast, main measure, traffic limits, skew factors and feasibility against the approved concepts and baseline data.
Use: ChatGPT
Ask ChatGPT to draft the experiment plan with the allocation proposal and stopping rules, then cross-check it against the checked table and correct it.
What to paste into the tool
Task: Draft ad experiment plan with allocation proposal, stopping rules for qualified review and no launch; Open questions and approval dependencies retained beside the recommendations; Reviewer-checked hypothesis table specifying the main measure, what could skew the result and the comparisons that can be run; source references; uncertain or absent facts; Supplied constraints, retained open questions and approval dependencies with source IDs and dates; Supplied inputs: Approved ad concepts, baseline aggregate data, supplied traffic constraints and decision objective. Propose one prospective message contrast and primary decision, the main measure, what could skew the result, allocation and stopping assumptions for qualified analyst review; do not launch. Retain the checked table, constraints and source packet unchanged in the output bundle. Keep unresolved facts and approval dependencies explicit. Do not resolve a missing fact by inventing one. Produce a draft only; do not publish, send, deploy, change accounts, spend or launch a campaign. Second message, after your check: Task: Corrected draft with a change list and unresolved mismatches flagged for the reviewer. Compare the draft with the checked hypothesis table and the retained source packet. Correct every mismatch you find in the draft, keep source IDs, dates, uncertainty and named exceptions, and list each change. Flag anything you cannot settle from the supplied material; do not invent facts, figures, policies or claims. Return preparation material only; publishing, outreach, audience changes, tracking deployment and financial commitments are outside this task.
Use: You
Compare the corrected draft and change list with the checked table, confirm the assumptions, then decide; launching stays outside this workflow.
Check isolation and statistical assumptions get analyst approval. If a check fails, correct the draft or leave the workflow incomplete.
Gemini Notebook only locates cited source passages. You build the question table and write the experiment plan yourself.
Assemble the materials
Build passage inventory
Build the plan
Check and decide
Use: You
Assemble the approved ad concepts, baseline aggregate data, supplied traffic constraints and decision objective, retaining authorised source IDs and dates.
Check each item is the approved version, dated and permitted; list anything missing as still to confirm.
Use: Gemini Notebook
Add the materials as notebook sources and ask for a cited passage inventory for the main measure, traffic constraints, baseline values and skew factors.
What to paste into the tool
Task: Source passage inventory for the main measure, traffic constraints, baseline values and documented things that could skew the result with citation locators and missing-evidence questions; Supplied inputs: Approved ad concepts, baseline aggregate data, supplied traffic constraints and decision objective. Extract passages for the decision objective, the main measure, baseline aggregate values, traffic constraints and documented things that could skew the result. Include the source ID, date and citation locator beside each passage. Identify missing evidence rather than proposing allocation or stopping rules. Retain the source packet unchanged in the output bundle. Keep unresolved facts and approval dependencies explicit. Do not resolve a missing fact by inventing one. Produce a draft only; do not publish, send, deploy, change accounts, spend or launch a campaign
Use: You
Open each cited passage, correct the inventory, then build the hypothesis table and the experiment plan locally, with allocation and stopping assumptions.
Check statistical assumptions directly; the source inventory is not a design decision.
Use: You
Check the table and plan against the source evidence and assemble the reviewed draft; launching stays outside this workflow.
Check isolation and statistical assumptions get analyst approval. If a check fails, correct the draft or leave the workflow incomplete.
Best when: You want AI proposals for table, plan and a first cross-check
What it does: Claude proposes, drafts and cross-checks; you check both gates
Strong at: Drafts and edits text in many formats
Limit: A table error can reach the plan; check at the checkpoint
Free plan: Yes, limited (5-hour and weekly usage caps)
Best when: You want AI proposals for table, plan and a first cross-check
What it does: ChatGPT proposes, drafts and cross-checks; you check both gates
Strong at: Analyses uploaded files; writes and rewrites text
Limit: Free plan limits file uploads, voice and data analysis
Free plan: Yes, limited (file upload and analysis caps)
Best when: You want to inspect citations before taking on the full design
What it does: Gemini Notebook finds passages; you build the table and plan
Strong at: Output is an inspectable passage inventory, not a design
Limit: Citations do not settle statistical assumptions; you still design
Free plan: Yes
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.
The named reviewer compares working records and final drafts against retained source evidence; unresolved questions stay visible.
These are proposed preparation workflows; execution, publication, sending, deployment, account changes and spending are outside scope.
Check the result against your original inputs. Approve it before sharing or acting.