AI-assisted qualitative analysis is fast, but fast is not the same as finished. Any researcher who takes their methodology seriously knows that a first-pass categorisation, however good, needs to be interrogated: codes that are too broad, distinctions that matter but weren't captured, categories that overlap, individual quotes that landed in the wrong place.
Skimle is designed for this. The AI generates a starting point you can work from rather than a fixed black box output you have to accept. This is a major differentiator vs. typical "AI assisted analysis tools" that are not really suitable for serious analysis.
This article explains the full range of manual editing tools available in Skimle, and then covers how to export your finished coding scheme to NVivo, MAXQDA, ATLAS.ti, or any other REFI-QDA-compatible software once the analysis is done.
This article is aimed at academic researchers who know their way around QDA software and want to understand exactly how Skimle fits into (or extends) their existing workflow.
Why manual control matters for rigour
There is a version of AI-assisted analysis that treats the AI as an oracle: it produces themes, you report them. This is not a methodology. Qualitative rigour requires iterative refinement, reflexive engagement with the data, and the ability to justify every coding decision.
Skimle's position is that the AI should do the time-consuming mechanical work (reading every document, identifying candidate quotes, proposing initial categories) while the researcher retains full authority over the final coding scheme. We believe AI tools should augment experts, not be used to produce AI slop at record speed. The manual editing tools exist to make that authority practical rather than theoretical: you should be able to reshape the category structure as easily as you can accept it.
This also matters for collaborative research. When you share an analysis with co-authors or present it for audit, you need to be able to show that the coding reflects your intellectual judgement, not just what an AI produced unsupervised. Manual refinement is part of demonstrating that.
Manual editing: what you can do
Renaming categories
Category names generated by AI are functional but might not capture your preferences or existing theoretical conventions. You can rename any category directly in the interface by clicking on its name. Skimle stores both a short name (used in the category tree and charts) and a longer descriptive label; both are editable. Renaming is immediate and propagates everywhere the category appears.
This is the most common manual operation. Researchers often find that the AI produces names that are accurate but not the terms they would use in their own work: "concerns about management responsiveness" might better be rendered as "management accountability" in the context of a specific theoretical framework.
Creating new categories and manual coding
You can create categories from scratch at any point during analysis. More importantly, you can manually code any passage of text from any document into any category, including ones you have just created.
In the document view, you select a passage of text, give it a label (Skimle offers an AI-generated suggestion based on the passage content, but you can overwrite it), and assign it to one or more categories. The label becomes the "insight" text that appears in your category summaries; the highlighted passage in the original document becomes the linked quote.
This is full manual coding in the traditional sense: you are directly assigning meaning to text passages and building a category scheme from the ground up if you choose to. The AI coding and manual coding coexist in the same project with no distinction in how they are stored or exported. After you have edited the insights in a specific category, remember to recreate the category summaries in the Category view.
You can also add notes to any individual insight: useful for recording your reasoning about a particular coding decision, flagging something for a second opinion, or noting a theoretical connection.
Moving and copying insights
Insights can be moved from one category to another or copied into multiple categories. The interface offers both options explicitly: moving removes the insight from its current category, copying keeps it in both.
This handles the common situation where a passage is genuinely relevant to more than one code, or where you have decided a category should be split and need to redistribute its contents manually. You can also reassign insights in bulk when restructuring larger sections of a coding scheme.
Merging categories
The merge operation combines two or more categories into one. All insights from the source categories are reassigned to the merged result, which you name yourself. Skimle will also suggest an AI-generated name for the merged category if you want one.
A useful feature here: when you select a single category and open the merge dialogue, Skimle automatically suggests the most semantically similar other category in your project (this computed from the semantic embeddings of the insights in each category) and shows a similarity percentage. If you have ended up with two codes that cover overlapping ground, this surfaces them for you rather than leaving you to spot the overlap manually.
Removing categories and reassigning their contents
Deleting a category does not have to mean losing the insights it contains. The remove-and-reassign operation lets you specify where each category's contents should go when it is removed. Skimle suggests the most similar remaining category for each item (using embedding similarity), but you can override any suggestion.
Insights with a confidence weight below a threshold are automatically removed during this operation, since they represent lower-confidence AI codings that are unlikely to be worth manually reassigning. The interface tells you how many will be affected before you proceed.
Trimming a category scheme
If you have run an inductive analysis on a large dataset, you may end up with a category scheme that is more granular than you need: 80 categories where 30 would serve your research questions better. The trim operation addresses this systematically.
Skimle computes a dispensability score for each category based on its insight count, its internal semantic coherence, and its relationships to neighbouring categories. The trim dialogue shows all categories ranked by dispensability and lets you incrementally reduce the scheme, watching the count update as you go. Removed categories' contents are automatically reassigned to the nearest remaining category. This is considerably faster than manual consolidation and produces a defensible rationale for the choices made.
Deleting insights and documents
Individual insights can be deleted with a confirmation step. Documents can also be removed from a project if they turn out to be out of scope or duplicates. These operations are straightforward but the confirmation requirement is deliberate: deletion is not undoable, which is standard in QDA software generally.
Exporting to REFI-QDA
Once your analysis is complete, you may want to take it into NVivo, MAXQDA, or ATLAS.ti. This can be useful if you or your organisation has a preferred software environment, if you are collaborating with researchers who still use legacy QDA tools, or if you want to use a feature that Skimle does not offer yet, such as a specific type of matrix query or a visualisation format you have used before.
REFI-QDA is the open international standard for QDA software interoperability. Skimle exports a fully compliant .qdpx file, which all the major QDA packages can open directly. You can export the file from the export menu in Skimle and open the file directly in your QDA software of choice.
What the export contains
The export is comprehensive. It is not a surface-level summary.
All your documents in the project are included as plain text files with the markdown formatting stripped for compatibility. Every document you analysed in Skimle is present.
Your full category hierarchy is preserved, including parent and child categories at every level. The hierarchical structure you have built, whether AI-generated, manually created, or a combination, comes across intact.
All codings are included with exact character positions. Every quote assigned to every category is exported with its precise start and end position in the source document. When you open the file in NVivo or MAXQDA, you will see your documents with the coded passages highlighted, exactly as they appeared in Skimle.
Category summaries are exported as memos linked to their respective codes. If Skimle has generated a summary of what a category contains, that text is attached to the code in the exported file and will appear as a memo in the QDA software.
Frequently asked question: Is it just a superficial export like with MAXQDA Tailwind? Note that Skimle's exports are much more comprehensive than, for example, MAXQDA's Tailwind or other AI-additions in legacy QDA tools. Tailwind exports only contain the documents and codes, with no coding applied to the documents itself. The export is surface-level: it transfers category names and some thematic content, but does not include the underlying codings with precise quote positions mapped back to your source documents. You get a set of theme labels rather than a fully coded project, which means you still need to do the actual coding work in MAXQDA after the import.
Skimle's REFI-QDA export includes the complete coding: every quote, at every character position, in every category, linked back to its source document. When you open it in MAXQDA (or NVivo or ATLAS.ti), you have a coded project, not a set of starting points that require you to re-code everything manually. The analysis you did in Skimle is preserved in full. This distinction matters if you are planning to do further analysis in QDA software, share the project with a collaborator for independent auditing, archive the analysis for replication, or provide supplementary materials for publication.
A possible workflow for academic research
The practical combination of these features looks something like this:
Step 1: Import and run AI analysis. Upload your interview transcripts or documents and run the initial thematic analysis. Skimle produces a first-pass category scheme with insights and quotes assigned to categories across all your documents. For 40 transcripts this takes minutes rather than weeks.
Step 2: Review and refine. Work through the category scheme systematically. Rename categories that do not fit your theoretical framing. Merge categories that overlap. Split ones that are doing too much work. Move individual insights that have been miscoded. Add manual codes for passages the AI missed or for a new category that emerged during your review. Use the trim function if the scheme needs consolidation. This step is where your analytical contribution is most visible.
Step 3: Add notes and memos. Use the note function on individual insights to record your reasoning on disputed or theoretically important codings. These notes travel with the insight and are available when you look back at the analysis later.
Step 4: Check patterns. If you have set up metadata fields, review whether any metadata variables explain meaningful variation in your categories. This can surface patterns that warrant another pass through the data.
Step 5: Export. When the coding scheme is finalised, export to REFI-QDA. Open the file in your QDA software if you need to run additional queries or share with collaborators, or archive it as a permanent record of the analysis.
This workflow is compatible with established qualitative methodologies. The AI accelerates the time-consuming parts (initial coding across all documents) without removing the researcher's engagement with the data. The manual editing tools ensure that the final coding scheme is one you have shaped, not one you have simply approved.
A note on trustworthiness and audit trails
One concern raised about AI-assisted coding is the audit trail: if a reviewer or examiner asks how a particular coding decision was made, can you explain it?
In Skimle, every insight is traceable to the original passage in the original document. The two-way transparency between categories and source text is preserved throughout the analysis, including after manual edits. When you move an insight, the link to its source quote moves with it. When you create a manual coding, the passage you selected and the label you assigned are both stored.
This means you can answer the question "why is this passage coded under this category?" for any coding in the project, regardless of whether it was AI-generated or manually created. That is the minimum requirement for a defensible qualitative methodology, and it is what distinguishes serious qualitative analysis tools from simpler text-summarisation tools.
Ready to combine AI efficiency with full methodological control? Try Skimle for free and bring your own QDA workflow to your next project, whether you finish the analysis in Skimle or export it to NVivo, MAXQDA, or ATLAS.ti via REFI-QDA.
Want to learn more about how Skimle handles qualitative analysis? Read our guide on how Skimle's end-to-end workflow handles qualitative data, how to analyse documents in multiple languages, and setting up interviews for AI-assisted transcription and analysis.
About the authors
Henri Schildt is a Professor of Strategy at Aalto University School of Business and co-founder of Skimle. He has published over a dozen peer-reviewed articles using qualitative methods, including work in Academy of Management Journal, Organisation Science, and Strategic Management Journal. His research focuses on organisational strategy, innovation, and qualitative methodology. Google Scholar profile
Olli Salo is a co-founder at Skimle and former Partner at McKinsey & Company where he spent 18 years helping clients understand the markets and themselves, develop winning strategies and improve their operating models. He has done over 1000 client interviews and published over 10 articles on McKinsey.com and beyond. LinkedIn profile