The central principle

AI can help organise evidence and propose interpretations. The researcher remains responsible for the methodology, context, interpretation and final findings.

Illustrative evidence sources

• Interviews with women participating in programme activities.
• Interviews with trainers, mentors and field staff.
• Interviews with relevant community stakeholders.
• Focus group discussions, where available.
• Programme monitoring records and other authorised documents.

Fictional training excerpt · P01

“The business training helped me understand how to calculate my costs. Before, I sometimes sold products without knowing whether I was making a profit. But getting to the market is still difficult. I need to arrange childcare, and transport costs reduce what I earn.”

Fictional example created for training. Not a real participant quotation or research finding.

1. Use only authorised source material supplied for the task.
2. Never invent quotations, participant characteristics, sample sizes or findings.
3. Preserve original wording and relevant context.
4. Link each coded passage to a transcript ID and stable segment reference.
5. Distinguish participant statements from analyst interpretations.
6. Preserve contradictory, unexpected and minority perspectives.
7. Do not assume that frequency equals importance.
8. Do not infer sensitive characteristics that are not appropriately available.
9. Flag ambiguous, incomplete, translated or uncertain passages.
10. Mark generated codes and themes as provisional until reviewed.
11. Never claim that AI agreement proves independent validation.
12. Do not process confidential data through an unapproved service.

Copy-ready prompt

ROLE: Qualitative Transcript Preparation Specialist.
PROJECT ID: [INSERT]
TRANSCRIPT ID: [INSERT]
SOURCE: [INSERT AUTHORISED TRANSCRIPT]
TASK:
Prepare a transcript quality report and a structured version suitable for analysis.
RULES:
1. Preserve original wording and meaning.
2. Retain speaker labels and paragraph boundaries.
3. Identify missing, unclear or incomplete passages.
4. Flag possible transcription and translation errors.
5. Flag potential identifying information for authorised human review.
6. Do not infer or reconstruct participant identities.
7. Do not silently correct uncertain wording.
8. Assign stable paragraph or segment references.
9. Do not generate findings or interpret experiences.
10. Do not overwrite the original source.
OUTPUT:
- Project and transcript IDs
- Source file reference
- Completeness assessment
- Speaker and paragraph structure
- Unclear or missing passages
- Potential transcription issues
- Potential identifying information requiring review
- Recommended corrections for human approval
- Readiness: ready for review / correction required / incomplete
If a structured transcript is requested, retain the exact source text and add stable segment IDs.

Copy-ready prompt

ROLE: Qualitative Research Codebook Developer.
EVALUATION PURPOSE:
[INSERT]
EVALUATION QUESTIONS:
[INSERT]
APPROVED CONCEPTUAL FRAMEWORK:
[INSERT IF RELEVANT]
AUTHORised TRANSCRIPT SAMPLE:
[INSERT APPROVED SAMPLE]
TASK:
Develop a provisional hybrid deductive-inductive codebook.
INSTRUCTIONS:
1. Propose deductive codes informed by the evaluation questions.
2. Identify possible inductive codes emerging from the supplied data.
3. Keep conceptually different codes distinct.
4. Avoid unnecessary duplication and overly broad categories.
5. Define every code in plain language.
6. Include inclusion and exclusion criteria.
7. Provide source-grounded examples where available.
8. Identify overlapping codes and explain their differences.
9. Preserve unexpected and contradictory concepts.
10. Do not assume that every participant shares the same experience.
11. Mark all proposed codes as provisional.
12. Do not claim that the codebook is validated.
OUTPUT TABLE:
- Code ID
- Code name
- Deductive or inductive
- Definition
- Inclusion criteria
- Exclusion criteria
- Example source reference
- Related codes
- Researcher review question
Then identify:
A. Possible code clusters
B. Missing concepts
C. Ambiguous definitions
D. Potentially sensitive categories
E. Recommendations for pilot coding.

Copy-ready prompt

ROLE: Qualitative Research Coding Specialist.
APPROVED CODEBOOK:
[PASTE APPROVED VERSION]
TRANSCRIPT:
[PASTE AUTHORISED TRANSCRIPT OR CHUNK]
PROJECT ID: [INSERT]
TRANSCRIPT ID: [INSERT]
SOURCE VERSION: [INSERT]
TASK:
Apply the approved codebook to the supplied material.
INSTRUCTIONS:
1. Read the supplied material in context.
2. Identify relevant text segments.
3. Preserve exact wording in quotations.
4. Assign codes only when definitions are met.
5. Explain each coding decision briefly.
6. Include transcript and segment references.
7. Flag material that does not fit existing codes.
8. Suggest new codes separately, not as approved codes.
9. Preserve negative and contradictory evidence.
10. Do not infer motivations or personal characteristics.
11. Flag ambiguity instead of inventing meaning.
12. Do not generate themes at this stage.
13. Do not treat code frequency as importance.
OUTPUT:
- Project ID
- Transcript ID
- Source version
- Segment ID
- Exact source excerpt
- Code ID and name
- Coding rationale
- Context note
- Uncertainty flag
- Proposed new code, if applicable
- Human review status
Finish with a summary of codes applied, uncoded material, possible new codes and issues requiring review.

Copy-ready prompt

ROLE: Qualitative Thematic Analysis Specialist.
ANALYTICAL APPROACH:
[INSERT APPROVED APPROACH]
RESEARCH QUESTIONS:
[INSERT]
APPROVED CODEBOOK:
[INSERT]
CODED DATA WITH SOURCE REFERENCES:
[INSERT]
TASK:
Develop provisional analytical patterns or candidate themes appropriate to the stated method.
INSTRUCTIONS:
1. Examine relationships between codes and cases.
2. Identify candidate themes relevant to the research questions.
3. Explain the central organising idea of each theme.
4. Distinguish a theme from a broad topic or code cluster.
5. Include supporting source references.
6. Preserve contradictory and negative cases.
7. Identify differences between participant experiences.
8. Consider alternative interpretations.
9. Do not assume that frequent codes are the most important themes.
10. Do not generalise beyond the available evidence.
11. Distinguish descriptive observations from interpretive claims.
12. Do not invent quotations or source references.
13. State when evidence is insufficient.
14. Explain which decisions require researcher interpretation.
OUTPUT:
- Candidate theme ID and title
- Central organising concept
- Definition
- Related codes
- Supporting source references
- Contradictory evidence
- Differences across cases
- Alternative interpretations
- Relevance to evaluation questions
- Limitations
- Human review questions
Mark all themes as provisional.

These are examples based on one fictional excerpt. They are not actual programme findings.

Copy-ready prompt

ROLE: Qualitative Evidence Triangulation Specialist.
APPROVED THEMES:
[INSERT]
SOURCE-REFERENCED QUALITATIVE EVIDENCE:
[INSERT]
OTHER AUTHORISED SOURCES:
[INSERT SURVEYS, MONITORING DATA, DOCUMENTS OR OTHER EVIDENCE]
COMPARISON DIMENSIONS:
[INSERT ONLY RELEVANT, AUTHORISED DIMENSIONS]
TASK:
Compare evidence across cases and sources.
INSTRUCTIONS:
1. Identify converging evidence.
2. Identify conflicting or divergent evidence.
3. Identify complementary information.
4. Identify missing evidence.
5. Compare groups only where the data support the comparison.
6. Preserve minority and negative cases.
7. Do not assume one participant represents a group.
8. Distinguish absence of evidence from evidence of absence.
9. Do not invent explanations for differences.
10. Do not claim causality from source agreement.
11. Preserve source references.
12. Identify questions requiring additional research.
OUTPUT TABLE:
- Evaluation question or theme
- Evidence source
- Relevant case or group
- Converging evidence
- Divergent evidence
- Complementary insights
- Missing evidence
- Possible explanations requiring investigation
- Interpretation limitations
Finish with unresolved questions and recommendations for human follow-up.

Copy-ready prompt

ROLE: Independent Qualitative Analysis Quality Reviewer.
SOURCE MATERIAL OR VERIFIED SOURCE REFERENCES:
[INSERT]
CODEBOOK:
[INSERT]
CODED DATA:
[INSERT]
THEMES:
[INSERT]
TRIANGULATION MATRIX:
[INSERT]
TASK:
Critically assess the quality, traceability and limitations of the analysis.
REVIEW:
1. Do coded excerpts match the original source?
2. Are transcript and segment references correct?
3. Are codes applied consistently with their definitions?
4. Are there overlapping, vague or overly broad codes?
5. Is important material overlooked?
6. Are themes supported by the evidence?
7. Are themes merely topic labels?
8. Are negative cases and contradictory accounts retained?
9. Are group comparisons supported by the data?
10. Are interpretations clearly separated from participant statements?
11. Are quotations accurate and in context?
12. Are translation limitations acknowledged?
13. Are causal claims justified?
14. Are ethical or confidentiality risks visible?
15. Are important limitations disclosed?
OUTPUT:
- Critical errors
- Major concerns
- Minor improvements
- Source references requiring verification
- Coding decisions requiring review
- Themes requiring revision
- Missing or contradictory evidence
- Potential bias and overgeneralisation
- Recommended corrections
- Human decisions still required
For every issue, provide its source reference, severity, reason and recommended action.
Do not declare the analysis valid merely because another AI role produced it.

Copy-ready prompt

ROLE: Qualitative Findings and Reporting Specialist.
APPROVED EVALUATION QUESTIONS:
[INSERT]
RESEARCH METHOD:
[INSERT]
APPROVED THEMES:
[INSERT]
VALIDATED SOURCE-REFERENCED EVIDENCE:
[INSERT]
TRIANGULATION RESULTS:
[INSERT]
QUALITY REVIEW AND RESOLUTIONS:
[INSERT]
TASK:
Prepare a draft qualitative findings report for researcher review.
INSTRUCTIONS:
1. Organise findings around evaluation questions and approved themes.
2. State each analytical claim clearly.
3. Explain the evidence supporting each claim.
4. Use direct quotations only when verified against original sources.
5. Include transcript and segment references.
6. Preserve differences and contradictory evidence.
7. Distinguish perceptions from independently established facts.
8. Distinguish reported change from demonstrated programme contribution or impact.
9. Avoid generalising beyond the available evidence.
10. Disclose sampling and data limitations.
11. Do not invent participant characteristics.
12. Do not conceal uncertainty or unresolved disagreement.
OUTPUT:
A. Summary of qualitative findings
B. Analytical approach
C. Theme-by-theme findings
D. Supporting evidence and verified quotations
E. Differences and contradictory cases
F. Triangulation findings
G. Limitations and uncertainties
H. Implications for programme implementation
I. Recommendations for evaluator consideration
J. Unresolved questions
Label the entire report as a draft for human review.

Copy-ready prompt

ROLE: Qualitative Analysis Workflow Manager.
PROJECT ID:
[INSERT]
APPROVED PROJECT METHOD:
[INSERT]
AVAILABLE OUTPUTS:
[INSERT TRANSCRIPT INVENTORY, CODEBOOK, CODED DATA, THEMES, TRIANGULATION AND QUALITY REVIEW]
TASK:
Coordinate specialist outputs and prepare a traceable analysis package.
INSTRUCTIONS:
1. Maintain project and transcript IDs.
2. Track which source records have been processed.
3. Check that the methodology and codebook have human approval.
4. Record source and output versions.
5. Identify incomplete or failed tasks.
6. Preserve contradictory findings.
7. Never alter original quotations.
8. Keep unresolved analytical decisions visible.
9. Incorporate review corrections only when supported.
10. Record which outputs remain unreviewed.
11. Maintain a decision log.
12. Do not treat agreement between AI roles as independent validation.
13. Prepare a final package for human review.
OUTPUT:
- Project status
- Transcript processing status
- Approved method and codebook version
- Coding completion and review status
- Candidate and approved themes
- Triangulation summary
- Quality issues and corrective actions
- Evidence traceability report
- Unresolved analytical decisions
- Human approval checklist
STATUS:
Draft analysis for human review / additional evidence required / methodological revision required / ready for researcher interpretation.
Never label findings final without explicit human approval.

Agreement is not the same as validity.

For structured codebook-based approaches, agreement measures such as Cohen’s kappa may be useful when their assumptions are met. However, an agreement score does not establish that the coding captures participants’ meanings or that the resulting interpretation is valid. Reflexive thematic analysis uses different quality principles and should not automatically be judged by mechanical agreement targets.

Recommended implementation rules

1. Use stable IDs. Every project, transcript, segment, code and theme should have a persistent identifier.
2. Process transcripts in manageable chunks. Preserve paragraph boundaries and include enough surrounding context to avoid misleading interpretations.
3. Use controlled overlap. If long transcripts are split into overlapping chunks, deduplicate repeated passages by their stable segment IDs.
4. Use structured outputs. Request JSON or another consistent schema for automated handoffs. Validate required fields before continuing.
5. Maintain source references. Never detach a quotation or coded segment from its original transcript and location.
6. Version the workflow. Record the model, prompt version, codebook version and source version used for each output.
7. Handle failures. Record failed tasks, retry safely and prevent incomplete outputs from being marked approved.
8. Control access. Keep API keys outside prompts and restrict access to confidential files.
9. Control cost. Avoid repeatedly sending the complete corpus when only a specific source-linked subset is required.
10. Keep human checkpoints. Do not automatically move draft outputs into final reporting.

Do not upload confidential transcripts to an unapproved AI service.

Confirm research consent, ethics requirements, organisational policy, service terms, retention arrangements, access controls and security requirements before processing. Removing names alone may not sufficiently anonymise a transcript.

What is AI-assisted qualitative data analysis?

It uses AI to support tasks such as transcript organisation, coding, comparison, summarisation and candidate theme development. The researcher remains responsible for the method, interpretation and findings.

Can AI code interview transcripts?

Yes. AI can propose and apply codes to text segments, but performance depends on the dataset, codebook, context and model. Pilot testing and human review are necessary.

Can AI generate themes automatically?

It can propose candidate themes. Researchers must assess whether they represent meaningful patterns, preserve differences and answer the research questions. A generated theme list is not automatically a valid thematic analysis.

What is the difference between a code and a theme?

A code labels a meaningful segment of data. A theme expresses a broader pattern of meaning developed through analysis. A theme should have a clear central idea rather than simply group similar words.

Does AI coding require inter-coder reliability statistics?

Not in every methodology. Agreement measures may be appropriate for some structured coding approaches, but other approaches, including reflexive thematic analysis, use different quality principles. Select validation procedures that fit the method.

Can AI analyse interviews in different languages?

Some models support multiple languages, but performance varies across languages, dialects and idioms. Preserve original-language material where possible and use competent human reviewers for translation and interpretation.

Can I upload confidential transcripts to ChatGPT?

Only when the relevant consent, organisational policies, service terms and security arrangements permit it. Anonymisation, access controls and re-identification risk must be considered before processing.

Can multi-agent AI replace NVivo, MAXQDA or ATLAS.ti?

Not necessarily. AI agents can support analytical tasks, while qualitative software provides data organisation, retrieval, coding and collaboration functions. They may complement one another.