Clean a customer file in five steps
The task: clean customers.csv. Learn its columns, split the addresses, find duplicates, propose merges that keep every row, then hand over a checked copy.
The five steps
Recorded results. Each step searches, takes the exact version it chose and writes one output the next step reads.
Show the folders for
{{skillPlace}} inside each step folder.
- 1
Profile the columns
Count what each column holds before anything changes.
Searchprofile a text column before cleaningChosenprofile_text_column_before_cleaningSkill, MIT, 2.9 KBsha256 1058606a…Folderstep-1-profile/Skill in {{skillPlace}}Skill named in the task file Outputprofile.json - 2
Split the addresses
Split each address into number, street, city, region and postal code.
Searchsplit address lines in a customer fileChosensplit_address_lines_into_componentsSkill, MIT, 3.0 KBsha256 4d7494fd…Folderstep-2-split/Skill in {{skillPlace}}Skill named in the task file Outputaddresses.csv - 3
Find duplicates
Find customers entered more than once.
Searchfind duplicate customer recordsChosenfind_duplicate_records_with_blocking_keysSkill, MIT, 2.7 KBsha256 984e29c5…Folderstep-3-duplicates/Skill in {{skillPlace}}Skill named in the task file Outputduplicate-pairs.csv - 4
Propose merges
Propose one survivor for each duplicate group, keeping every row.
Searchpropose a dedupe without deleting rowsChosenpropose_a_dedupe_without_deleting_rowsSkill, MIT, 2.6 KBsha256 3935a528…Folderstep-4-merge/Skill in {{skillPlace}}Skill named in the task file Outputmerge-proposal.csv - 5
Hand over the copy
Write the cleaned copy and the notes its reader needs.
Searchhand over a cleaned file its consumer can useChosenhand_over_a_result_its_consumer_can_useSkill, MIT, 2.7 KBsha256 2a2726d3…Folderstep-5-handover/Skill in {{skillPlace}}Skill named in the task file Outputcustomers.cleaned.csv
Recorded searches
These searches replay the starter library as it was recorded for this page. A fresh search can return newer results and versions, so take the digest from a fresh search before downloading.
Your agent chooses
This demonstration takes the first result at each step. In your own run, your agent reads the results and picks the best fit. The folders show where each harness reads the skill and how each output feeds the next step.
Savings depend on the task, the model and the setup.
Measured in the open.
Our data cleanup study ran each step three times in a fresh Pi process, and an independent scorer checked every output. The full report is public, every result included.
143
synthetic rows
72
counted steps
3
runs per step, each in a fresh Pi process
4
task families, scored independently
Measured September 22 to 23, 2026, with Pi 0.73.1.
Run these steps yourself.
Pro is $29 a month for the whole library. Cancel any time from your account page.
Pro
$29a month
Includes Agent Feeds + Skills and agent files
- The whole library, 1,560,856 files
- The exact version and checksum on every download
Model usage stays with your provider. Cancel any time from your account page.