You have coded some of your interviews by hand and you want the rest coded the same way: with your labels, and with the same sense of which claims count. Code Like Me is how the Causal Map app does this. You code a few sources yourself, then ask MapCat, the app's assistant, to work out how you code and to set up AI coding to continue in your style, for example: "Look at how I coded A1, A2 and B3, and code C1, C4 and D2 the same way." This page is about the method; the steps in the app are on AI coding.

Choose the sources you code by hand#

The AI learns only from what you show it, so which sources you code first matters more than how many.

• Code a few whole sources (we have tested this with three to five).
• Choose them at random within each group your analysis will compare: some women and some men, some rural and some urban households, each district or each year. The Random N / Group buttons in the Sources Bar choose them, or MapCat can choose them for you. A sample drawn from one group teaches the AI how that group talks.
• Code each source all the way through. What you leave uncoded is part of how you code: the claims you pass over, the passages you treat as background, the hypothetical "what I would do if..." answers you skip.

If you ask MapCat when you have coded only one source, or only one kind, it suggests coding a few more first.

What MapCat learns#

MapCat learns two things from your coding, and handles them differently.

• Your labels. MapCat copies the labels you used, exactly as you wrote them, into the project's codebook, and the AI uses them wherever they fit.
• Your rules, meaning everything the labels cannot show: which claims you code and which you leave out; where the text gives a chain (A led to B, which led to C), whether you also link A straight to C; how much text one link covers; and how you build a label, including how it shows which way something changed. MapCat writes these as rules in its own words, never as examples from your coding, because a coder given examples tends to copy them.

MapCat then tries the coding on a whole source you coded, with your labels, and sets the result beside your coding. It reads the two as a coder would: whether they tell the same story, chain claims through the same factors, code a repeated claim once where you do, and give each label the right direction. It adjusts the rules and tries again. In our tests the AI followed a rule about what to leave out only loosely when it read the rule before coding, so where a trial shows the AI still coding something you skip, MapCat adds a later pass, which goes through the links just produced and removes those that break the rule.

General principles first, your habits second#

Every good coder has habits of their own, and many are matters of preference rather than right or wrong. Code Like Me follows yours, within the rules that apply to all causal coding in the app: one cause and one effect per link; a quote, copied word for word, that shows the claim; no claim the text does not make; and no label that says the opposite of what the text says. The basic rules say nothing either way about hypothetical claims, such as "if a colleague broke the code, I would report it", so whether to code them is part of your own style, which MapCat tries to follow from your coding. When the AI's coding differs from yours, ask whether it breaks one of those rules. If it does, it is an error. If it does not, it is a difference of style, which a change to the instructions or a later pass can bring closer to yours.

What to expect#

• MapCat codes the sources you name. If you name none and many are still uncoded, it asks before coding them all and suggests a random sample first, so you can read and correct that coding before the rest. Coding usually goes in rounds.
• If you have coded many sources by hand and do not say which to learn from, MapCat reads a random sample of about five and says which; the codebook still takes the labels from all of them.
• The AI usually codes more links per source than a person does. In our tests most of the extra links were claims the text does make, which the person had chosen not to code. Extra links are easy to find and delete.
• MapCat sets a coding model that followed hand coding more closely than the default in our tests and costs several times as much. Each trial on a whole source costs about what coding that source in a run would.
• Before relying on counts from the AI's coding, read one coded source beside your own.

This page was drafted by the Causal Map team with the assistance of Claude, an AI model from Anthropic; the content and judgements are our own.