Vibe Coding a Data Import Script

Describe the source, the field mapping, and the dry run before a builder writes the import

Voiskey Voiskey Team
Voice Vibe CodingAI Code PromptsVibe Coding Data

Profile the Source File First

Vibe Coding a Data import script starts with a careful look at the source file, not with a prompt. Open the file and record its shape: the columns, their order, their types, and the rows that look unusual. Count the rows and note the empty cells, the mixed formats, and the values that do not match their column. Save a small sample that includes both clean rows and strange ones. A script built without this profile will fail on the first real file, and the failure will look mysterious because nobody described the input. Also note the file size and the date it arrived, because both decide how long the first run should take.

Name the destination and what a correct row looks like when it arrives. Say the table, the fields, and the required values. Describe how each source column maps to a destination field, and what should happen to columns that have no destination. When a value must be transformed, say the rule and give an example, such as trim the spaces and use lowercase for the email field. The mapping is the heart of the import, and it deserves a written form before any code exists. If a destination field is required, say so beside the mapping, because an empty required field is the first thing a bad row will hit.

Describe the volume and the rhythm of the import. Say how many rows a normal file contains, how often the file arrives, and whether the import runs once or repeatedly. A file with a hundred rows can be checked by eye; a file with a million rows needs different handling and a different test plan. When the same file may be imported twice, say so, because duplicates change the whole design. Volume and rhythm decide how much the script must automate and how much can stay manual. If the import runs on a schedule, say who watches it and what a missed run looks like to the team.

Write the questions you cannot answer and assign each one to a person. Which row wins when two rows share an identifier, which fields are allowed to be empty, and what to do with a value that is out of range. Do not invent answers to keep moving. A short list of named questions is better than a script that quietly chooses a policy the owner never approved. Resolve the list before the build, or mark the unresolved items as scope for a later version. Resolve the questions before the first build, or mark them clearly as scope for the second version.

Speak the Field Mapping Aloud

Speak the field mapping as two columns in a sentence, source on the left and destination on the right. Say the source column, the destination field, and the transformation in that order. For example, customer name maps to full name with spaces trimmed. Read the whole list aloud and listen for a column that appears twice or a destination with no source. Exact mapping language prevents the builder from guessing which field a value belongs to, and it gives the reviewer a list that can be checked line by line. When a destination has no source, mark it as a default or a question, never a guess.

Describe each transformation with an example rather than a category. Say a date in month slash day slash year becomes year dash month dash day, and read one real value through the change. Examples settle questions about leading zeros, time zones, and empty values that a category name leaves open. When a transformation can fail, say what the script should show and whether the row continues. A mapping with examples is a specification; a mapping with labels alone is a wish. Write one real example for each rule in the table so the reviewer can test the rule quickly today.

Name the fields that must be unique and the fields that may repeat. Say the email must be unique, and the phone may repeat. State what the script should do when a uniqueness rule is broken: stop the file, skip the row, or update the existing record. Each choice changes the result, so the owner must make it. Speak the decision and the reason into the task. A clear uniqueness rule prevents a duplicate from becoming a support ticket weeks later. Say whether a failed uniqueness check stops the whole file or only skips the later duplicate copy. State the choice in a sentence the owner can approve or reject.

Ask for a mapping table in the plan before asking for code. The table should list each source column, its destination, its transformation, and its handling when empty. Read the table aloud and check that every destination appears. When a source column has no destination, say whether it is ignored on purpose. Approve the table, then build from it. A table that both sides accept is the cheapest way to avoid a rewrite after the first real import. If any row in the table has an empty transformation column, the mapping is not complete enough to build the script from safely.

Handle Bad Rows and Duplicates

Describe a bad row as a row the script cannot safely import. Say what makes it bad: a missing required field, a value that does not match its type, a broken reference, or a duplicate. For each kind, say what the script should do and where the row should go. A review file with the original row and a short reason is more useful than a count of failures. The owner can then fix the row and rerun the import without searching for what went wrong. Say where the review file lives and who is expected to open it before the next run begins.

Decide whether the import stops at the first bad row or continues and reports at the end. Say the choice and the reason. Continuing is faster for a large file with a few problems; stopping is safer when one bad row means the file is wrong. Either way, the script should leave a readable summary. Speak the summary fields: rows read, rows imported, rows skipped, and the reason for each skip. A clear summary turns a failed import into a short repair list. Say whether a skipped row can be corrected and reimported without rerunning the whole file from the start.

Handle duplicates with an explicit rule, not an accident. Say what counts as a duplicate, which copy wins, and whether the older record is updated or left alone. When the source carries a timestamp, say whether it decides the winner. Speak one example of a duplicate pair and the expected result. If the rule depends on business policy, name the owner who decides it. A duplicate rule that lives only in someone's memory will be applied differently by the next person. Say whether the winner is chosen by the newest timestamp or by the first row seen in the whole file, and state the rule plainly.

Describe the order of operations when the import touches several tables. Say which table is written first, which depends on which, and what happens if a later step fails. If the script can write half a file and then stop, say how the team should clean up. A short rollback note is worth more than a perfect happy path, because imports fail at the worst moment. Keep the note beside the run steps so the operator can find it under pressure. Say which step is safe to repeat on a retry and which must run only once for each input file.

Dry Run, Review, and Download

When the mapping and rules are agreed, use Voiskey to dictate the run notes at the insertion point in the task document. Speak the input file, the command, the summary fields, and the check that confirms success, and let the app remove filler so the notes are easy to follow. Because it learns your field names and paths, the notes stay consistent with the mapping table. Run the script on a small copy first. When the dry run matches, download the notes and the script together. Keep the notes beside the mapping table so both can be read together in one pass.

Run the dry run on a sample that includes clean rows, bad rows, and a duplicate. Compare the summary with the expected numbers, not only the imported total. If the skipped count is higher than expected, inspect the review file before changing the script. When a rule is wrong, fix the rule and rerun the same sample. Download the updated script after each fix so the reviewed version is the one that produced the numbers. If the numbers disagree, write the expected value beside the observed one before changing the code or the rule at all, and keep the sample file.

Ask the owner to check the review file rather than the whole import. The review file lists the rows the script refused, with a reason for each. If the reasons are clear, the owner can decide whether to fix the source or change a rule. Speak each decision into the task and update the mapping table when a rule changes. Download the approved table with the script. A clear review file turns a long argument about correctness into a short list of decisions. Keep the review file beside the run notes so a decision can be traced to a single row later.

Test the import twice with the same file to confirm it behaves the same way on a repeat run. If the second run creates duplicates, the uniqueness rule is incomplete. If the second run imports nothing and says so, the repeat behavior is understood. Record the result in the notes. When the repeat behavior is correct, download the final script, the mapping table, and the run notes as one package for the team. If the second run imports new rows instead of skipping them, the uniqueness rule needs another careful look and a small fresh test case in the shared set right away.

Keep the Import Repeatable

Keep the mapping table and the run notes with the script so the next person can change one without guessing the others. When a source file changes, open the table, read the mapping aloud, and mark the columns that moved or disappeared. Use Voiskey to dictate the change at the insertion point, then check that every destination still has a source. Download the updated package and note the date. A change that touches only the table is much cheaper than a change that touches the code. Keep the package name short so the next person can find it in the task list.

Watch the summary counts over time. A rising skip count usually means the source format has drifted or a new bad value has appeared. Read a sample of the skipped rows and decide whether the rule or the source should change. Speak the finding into the notes and assign the next action. Download the notes after each review. Counts that nobody reads hide a slow decline that only shows up when someone finally needs the imported records. Keep the notes current so the counted rows mean something when two months of imports are compared directly side by side with each other in the same review.

Keep a small set of test files beside the script: one clean sample, one with bad rows, and one with a duplicate. Run all three after every change and compare the summaries with the expected numbers. This takes minutes and catches most regressions. When a new failure appears in a real file, add a sample that reproduces it, then fix the script. Download the test files with the script so the set stays together. If a test file becomes outdated, replace it with a small current sample instead of deleting it from the set silently and note the change in the notes.

Retire the script when the source disappears or a better path replaces it. Say so in the task and archive the script, the table, the notes, and the test files together. Point the team to the replacement and note the date. A retired import with a clear record is better than one that runs by habit and nobody can explain. Review the retirement at the next planning session so the old job is not quietly restarted. Say why the import stopped and what replaces it before the archive is closed for good or the team agrees to keep it running.

FAQ s

Speak each source column, its destination field, and the transformation in that order at the insertion point in your plan. Voiskey removes filler so the mapping reads as a clean list. Add one real example for each transformation. A clear spoken mapping keeps the builder from guessing which value belongs to which field.

Check the source file first: its columns, their order, the empty cells, and the rows that look unusual. Save a sample with clean and strange rows. Then say the destination, the mapping, and what a correct row looks like. A profile of the input prevents a script that fails on the first real file.

Yes. The app learns the field names and paths you repeat, so the mapping and the run notes stay aligned. Say each name clearly once, check the inserted line, and reuse it instead of a synonym. Consistent names make it easier to see which column maps to which field.

Download the script with its mapping table and run notes as one package, then attach it to the task. Include the clean sample, the bad-row sample, and the duplicate sample. Ask the owner to check the review file and confirm each skipped row has a reason. One package keeps the import easy to rerun.

Say less. Get more
Say less. Get more
Say less. Get more
Say less. Get more
VoiskeyVoiskey

Your voice, amplified.

Download Voiskey for Mac and Windows to describe your import mapping and run notes in clear, consistent lines.

Use it across macOS, Windows, iOS, and Android

Use it across