A spoken procedure for inspecting, cleaning, and validating a table.

Dictating Prompts for Data cleaning tasks begins with a careful look at the table. Open the file and note the number of columns, the header row, the first and last row, and any column that mixes formats. Say these facts at the start of your prompt, because a cleaning request that skips them usually cleans the wrong thing. A table with a title row above the header is common, and it breaks rules that assume the header is first. Two minutes of inspection saves a full rebuild. Keep the inspection note beside the file, because the same facts will be needed again if the pass has to be rerun.
List the problems you can see before you ask for a fix. Say which columns have blanks, which have duplicates, which use inconsistent capitalisation, and which mix text and numbers. If a column has trailing spaces or odd separators, say so. Do not ask for a general clean; ask for a specific list of fixes in a specific order. A specific list is checkable, while a general request produces changes you did not ask for and may not notice. Put the most damaging problem first, so the first fix addresses the issue that would otherwise distort every later step. Number the fixes so the order is explicit.
Decide which problems are out of scope. Some rows look wrong but are correct, such as a zero that means no activity rather than missing information. Say which values must not be changed. A cleaning pass that silently corrects valid entries is worse than no pass at all, because the error is invisible afterward. Name the protected columns and the values that are allowed to stay unusual. When you are unsure about a value, mark it for review rather than choosing a rule, because a flagged row can be checked later and a silently changed row cannot. Keep the review list short and give each flagged row a reason.
Write down the target shape of the clean table. Say what one row should represent, which columns must be filled, and what the final column order should be. If the table feeds a report or an upload, say that too, because the required shape may be stricter than a readable table. A stated target lets you compare the result with a clear standard instead of judging it by appearance. Name the required column order exactly, since a correct value in the wrong column can break the upload just as badly as a missing one. Write the order as a numbered list, and check it after the clean.
State the rule for duplicates before you ask for them to be removed. Say which columns define a duplicate: a full-row match, a match on one identifier, or a match on a combination. Say which copy to keep, such as the most recent or the one with the fewest blanks. Without this rule, a duplicate pass can delete the record you needed. A written rule also lets a reviewer check the result against the same standard. Write one line per rule, so the prompt and the review sheet can be compared side by side. Keep each rule to one sentence, and name the tie-breaker.
Give an explicit rule for blanks. Say whether a blank should be left, filled with a default, filled from another column, or flagged. Each choice changes the meaning of the table. Filling a blank with zero can be right for a count and wrong for a price, so say which columns allow which treatment. When the rule depends on the column, list the columns and the rule beside each one. If two columns share a rule, group them, but never assume that a rule for one column applies to another without saying so. Group only what you name. Write the grouping beside the rule.
Say how to handle near-duplicates and inconsistent labels. Two spellings of the same name, a trailing space, or a different capitalisation can hide a duplicate. Ask for a normalisation step before the duplicate check, and say what normalisation means for each column. Normalising text before matching is the difference between an accurate merge and a missed one. Write the rule in plain words and give one example. When a label has a legitimate variant, such as a regional spelling, say whether to merge it or keep it, so the normalisation does not erase a real difference. Add that exception to the rule.
Ask for a report of what changed, not just a clean file. Request a list of rows removed, values filled, and labels normalised, with counts. A change report is how you verify the pass and how you explain it to someone else. If the counts look wrong, stop and inspect before using the result. A cleaning pass with no report is a guess disguised as a fix. Keep the report with the file, so a later reviewer can see what the pass did without rerunning it. Name each change type in the report, sort it by rule, and that order matches your prompt.
Attach a check to every cleaning step. If a step removes duplicates, ask for the before and after row count. If a step fills blanks, ask for the number of cells changed. If a step normalises labels, ask for the list of distinct labels before and after. A step with a check can be trusted or rejected; a step without one is an act of faith. The checks also make the prompt longer in a useful way. Put each check beside its step, so the order of the checks matches the order of the changes. Keep the checks numeric where possible, because numbers are hard to argue with.
Order the steps so each check still makes sense. Remove duplicates before filling blanks, or the fill will count rows that are about to disappear. Normalise labels before merging, or the merge will miss matches. Say the order explicitly and ask the assistant to keep it. When the order is wrong, the counts will not reconcile, and reconciliation is exactly what the checks are for. Write the order as a numbered list at the top of the prompt, so both the assistant and the reviewer work from the same sequence. Do not let the order drift, and rerun from the start.
Ask for one small sample to be shown at each step. A few affected rows before and after a change are easier to review than a summary count alone. The sample shows whether the rule did what you meant, not just how many times it ran. When a sample looks wrong, fix the rule before running the rest of the table. A sample is a cheap way to catch a rule that is technically applied and practically wrong. Choose sample rows that include an edge case, because a clean sample proves less than a difficult one. Ask for five rows at most.
Keep the checks in the prompt, not in a separate note. A prompt that contains its own checks is easier to rerun and easier to hand to someone else. If a step fails, the check tells you where to look. When you improve a check, update the prompt so the next run uses the better version. A cleaning prompt with built-in checks becomes a repeatable procedure rather than a one-off request. If a check passes for the wrong reason, rewrite it until a known bad row would fail, then return to the real table. Test the test before you trust it, because a check that never fails is useless.
Compare the cleaned copy with the original before you use it. Check the row count and the column count, then check a few identifiers by hand. Pick three rows and confirm that the values match the original except where a rule changed them. If a value changed without a rule, the pass did something you did not ask for. A short manual comparison catches the class of error that counts never show. Keep the comparison in a note beside the file, so the next person can repeat it without rebuilding the check from memory. Name the three rows you used.
Use Voiskey to describe any mismatch in plain language. Put the insertion point in the message window, press the shortcut, and say the row, the column, the original value, the new value, and the rule you expected. The app removes filler and false starts, so the report is easy to read and easy to act on. A clear description gets a clear fix, whether the fix comes from the assistant or a colleague. If the mismatch is a pattern rather than a single row, say how many rows share it, because a pattern needs a rule change and a single row may need a manual correction. State the pattern count first, because it decides the fix.
Check the edge cases separately. Look at the rows with blanks, the rows that were nearly duplicates, and the values that looked unusual but valid. These are the rows most likely to be changed by mistake. If a protected value changed, restore it and tighten the rule. When you set up another computer, download Voiskey there and keep the same check rows available, so the validation does not depend on one device. Write the edge rows into the check note, because they are the cases a future run is most likely to get wrong again. Give each one a quick verdict.
Get a second person to run one check on the final copy. Ask them to verify a handful of rows against the original using the change report. A second pair of eyes finds the mistake the first pass assumed away. Keep the report and the check rows with the file, so the validation can be repeated later. A clean table is only trustworthy when someone can reproduce the check. Ask the second reviewer to write down what they checked, not just whether they were satisfied, because a written check can be repeated and a feeling cannot. Store that note beside the final copy.
Save the prompt that produced a clean table. Keep the inspection notes, the list of fixes, the duplicate and blank rules, and the checks together. The next import starts from a tested procedure instead of a blank request. Name the saved prompt by the kind of table, such as monthly sales export or customer list import. A clear name means you can find the right procedure when the file arrives under time pressure. Add the date of the last successful run, so you can see at a glance whether the procedure is current. Update the date after each clean. Note the source file name beside the date.
Note the quirks of each source. One export may put a title in the first row, another may use a different date format, and a third may add a totals row at the bottom. These quirks are the reason a generic cleaning prompt fails. Record them beside the saved prompt and check them when a new file arrives. A short list of known quirks is more valuable than a long list of general advice. When a new quirk appears, add it to the list before the next import, so the procedure improves from real files rather than from guesses. Recheck the quirk list before every import.
Use Voiskey to update the prompt by voice when a new quirk appears. Open the note, press the shortcut, and speak the problem and the rule you want. The entry lands beside the existing rules, so the next run includes the fix. Read the changed section once before saving. This keeps the procedure current without a full rewrite after every unusual file. If the fix affects the order of steps, update the numbered list as well, because a rule in the wrong place can pass its own check and still break the next one. Check the order after each edit, then run one test file to confirm the fix.
Share the saved prompts with anyone who runs the same imports. Give them the procedure, the change report template, and the check rows. Show one clean result and one corrected result, so the standard is visible. When a colleague sets up a second machine, download Voiskey there and copy the procedure folder in the same structure, then run one test file through it. A shared procedure makes the output consistent across a team. Ask each person to add their own source quirks to the shared note, so the procedure grows from every import rather than from one person's memory. One shared note beats many copies.
Say which columns define a duplicate and which copy to keep, then ask for a before and after row count. Voiskey turns the spoken brief into an AI-ready prompt and removes filler, so the rule list stays clear. Review the change report before using the cleaned sheet.
Yes. Name each column and say whether a blank should stay, take a default, or copy from another column. Voiskey adapts formatting to the request, which keeps a per-column rule list readable. Ask for the number of cells changed so you can check the result.
Ask for a normalisation step first, then a duplicate check, and give one example of the label change you expect. Voiskey keeps the spoken steps in order and learns names, jargon, and acronyms, so repeated labels stay stable. Include the distinct label count before and after.
Compare row and column counts with the original, then check three identifiers by hand. Read the change report and confirm that every changed value traces to a rule. Test the rows that had blanks or near duplicates. Voiskey places corrections in the same field, so the final comparison stays in one place.

Download Voiskey for Mac and Windows and speak your next prompt in seconds.
Use it across macOS, Windows, iOS, and Android