How do I do thematic analysis step by step?

The six-phase description is standard and slightly misleading, because it looks linear and the analysis is not. You will return from phase four to phase two more than once, and that recursion is part of the method rather than a sign you did it wrong. Say so in your write-up.

Phase one, familiarization, is the phase most often skipped under time pressure and the one that most affects the result. Reading the whole dataset before coding anything, and taking notes on impressions, is what stops your codes being shaped entirely by the first three transcripts.

Coding across the whole dataset before building themes matters for the same reason. Building themes early means fitting the rest of the data into a structure derived from a fraction of it, and the analysis then confirms an early impression rather than testing it.

A codebook makes your coding decisions reviewable. Coates et al. recommend defining each code, stating what belongs inside it, and attaching examples taken from the raw data, and they note that a codebook can incorporate the relevant literature and improves consistency when several people code the same material (Coates et al., 2021, p. 3). Their team sequence is also worth copying: independent preliminary coding, a meeting to agree the coding scheme, an assessment for saturation, formal individual coding, then a dispute-resolution meeting before anyone names a theme (Coates et al., 2021, p. 2). Coding is iterative in practice, not one pass. In Thomas and Harden's synthesis of eight studies, every sentence received at least one code and most received several, all the text under each code was rechecked for consistency, and the first round produced 36 codes (Thomas and Harden, 2008, p. 5).

Separate descriptive themes from analytical ones and say in your methods which you are reporting. Thomas and Harden grouped their 36 codes into a hierarchical tree of 12 descriptive themes that stayed close to the source text, then built analytical themes that went beyond it (Thomas and Harden, 2008, pp. 6, 8-9). Their tree also shows that themes need not be mutually exclusive, since one finding about children's food preferences sat in two places at once, because it concerned both reactions to food and behavior when choosing it (Thomas and Harden, 2008, p. 6). The analytical stage is where the interpretation appears. Theirs asked how fruit and vegetable consumption could be promoted, identified six issues in children's views, and reported that interventions targeting people with particular risk factors worked better, while interventions combining physical activity with healthy eating worked less well than those focused on healthy eating alone (Thomas and Harden, 2008, p. 7).

The write-up is where thematic analyses most often disappoint. Extracts are evidence, not illustration, so each one should be doing analytical work in the paragraph around it. A results section that presents three quotations per theme with a sentence of framing is a report of what people said rather than an analysis of what it means.

What is the difference between a code and a theme?

A code labels a segment of data. A theme is a pattern of shared meaning across codes that says something about your research question. Codes are descriptive and numerous, themes are interpretive and few. A list of topics is not a set of themes, however neatly it is organized.

The diagnostic is whether the theme has a point. "Communication" is a topic. "Participants treated communication failures as evidence of institutional indifference rather than individual error" is a theme, because it says something.

Three to six themes is typical for a study. More than about eight usually means you have collated codes rather than interpreted them, and the fix is to look for the shared meaning connecting several of them.

How do I know a theme is real and not something I imposed?

Check it against the coded extracts and then against the whole dataset. A real theme has internal coherence, clear boundaries against other themes, and enough supporting data across participants. If it holds together only in the extracts you chose to look at, it is not yet a theme.

Look specifically for disconfirming data. A theme that no case contradicts is either genuinely strong or a sign that you stopped noticing the cases that did not fit, and going looking for the exceptions distinguishes the two.

Agreement between coders is a weaker check than it looks. Bonner et al. mapped 583 descriptions of childhood vaccination barriers onto a deductive framework and reached 89% agreement on the first pass, and the remaining discrepancies still forced them to write more precise definitions at the construct level (Bonner et al., 2021, p. 1). The same 583 descriptions coded inductively produced 74 barriers in seven categories and differentiated the data more finely, while the deductive pass exposed four domains of the 14-domain framework that the data did not reach. They recommend combining both approaches when you want comprehensive coverage, rather than treating either one as generally superior (Bonner et al., 2021, p. 2).

Reflexivity is the other half. Write down what you expected to find before you began, and check your themes against that note. Where they match perfectly, look harder. That note is also worth including in the methods, since it is evidence of the awareness examiners are testing for.