Error guessing has a reputation problem. The name suggests a tester poking at the screen and hoping. The technique, as described in ISTQB material, is nothing of the sort: you use knowledge of how software like this tends to fail, and of what this team and this product have got wrong before, to decide what to try. It is guided by a list, even if the list lives in your head.
The best way to see it is to follow one session from the first look to the last note. Here is the feature.
The feature
An HR portal for a mid-sized firm in Coimbatore lets an admin upload a spreadsheet of new joiners as a CSV file. Columns: employee name, date of joining, department, monthly salary in rupees, mobile number. The system creates the records and shows “Upload successful”. The spec says only: “Admin can bulk upload employees.” No boundaries, no error rules.
Equivalence classes and boundaries will give you some cases. The rest come from asking, of each part of the feature, “how do things like this usually break?”
Step one: list the inputs and the places they are handled
The inputs are the file itself and each of five columns. The places they pass through are the parser, the validation, the database and the confirmation message. Bugs live at the seams, so for each input we ask about each seam.
Step two: apply the usual suspects to the file
Experienced testers carry a mental list of file troubles. Applying it here gives:
- An empty file, and a file with headers only.
- A file saved from Excel with a byte order mark at the start, which can make the first header unreadable.
- A name containing a comma, such as “Iyer, Karthik”, which breaks naive splitting unless it is quoted.
- Names with Tamil or other non-English characters, to check the encoding survives.
- Windows versus Unix line endings.
- A very large file, say 50,000 rows, to see whether it times out.
- The wrong file type renamed to .csv, and a real Excel .xlsx file.
None of these were in the spec. All are drawn from how CSV handling commonly fails.
Step three: apply the usual suspects to each field
Dates: 30 February; a date written as 03/04/2025, which is 3 April in India but 4 March in the American reading; a joining date in the future; a date stored by Excel as a serial number. Salary: zero, negative, a decimal like 45000.50, a value with a rupee sign or commas such as “₹1,20,000” (Indian digit grouping will wreck a simple number parser), and an absurdly large number. Mobile numbers: a leading zero, a +91 prefix, nine digits, eleven digits, and a number Excel has turned into scientific notation, like 9.88E+09. That last one is a real and common spreadsheet behaviour. Name: empty, only spaces, 300 characters, an apostrophe as in “D’Souza”.
Step four: apply the usual suspects to the process
Here the questions are about what happens across rows and across attempts. What if row 7 of 20 is invalid: are the other 19 saved, none, or does the user get no clue which row? What if the same file is uploaded twice: duplicates, or rejection? What if two rows have the same mobile number? What if the admin double-clicks Upload? What if the connection drops halfway: is there a half-imported batch? What if a non-admin user posts to the upload address directly?
This is the heart of the technique. We are not testing random inputs. We are replaying known failure patterns: partial failure, repeated submission, state left half-changed, permission checks done only in the interface.
Step five: run, observe, and note
Suppose the run finds three things: the scientific-notation mobile is saved as 9880000000 with no warning, the 20-row file with one bad row says “Upload successful” while saving only 19, and the message never says which row failed. Each is a real defect with an obvious next test, and each one came from a hunch with a reason behind it.
Write the cases down afterwards, briefly. A line per case, with what you tried and what happened. Otherwise a successful guess evaporates when you leave the team.
Where the list comes from
Be honest about the limits. The list is only as good as the experience behind it, so two newcomers will find less than one veteran. You can widen it deliberately: keep a team checklist of past defects and their patterns, read published failure collections and the OWASP material for security habits, and look at the defect history of this product to see what it keeps getting wrong. A module that has leaked bugs about dates three times will probably do it again.
Error guessing also does not give you a coverage claim. You cannot say you have tested all dates. It complements the systematic techniques, covering the cases that no formula produces, and it works best when the technique is timeboxed and the findings are recorded, which is exactly what a session charter is for.
Figures and names in the examples are illustrative.
Leave a Reply