A custom AI reviewer for qualitative and mixed-methods rigour

PerfectPaper’s qualitative agent reads an interview, focus group or mixed-methods manuscript against the reporting standards qualitative reviewers apply: how participants were sampled and why, what saturation claim is made and what supports it, how coding was developed and checked, whether the researchers’ own position is addressed, and whether the quotations shown support the themes claimed. Copy the brief below into a custom agent slot.

Qualitative work is the widest gap in the collection. PerfectPaper’s standing team was built for quantitative manuscripts, so for an interview or focus group study a custom agent is not an enhancement but the substance of the review.

When to use this agent

• Your manuscript reports interviews, focus groups, or open-ended survey responses
• You are submitting to a journal requiring COREQ or SRQR
• Your methods say themes emerged and little else
• Coding was done by one person
• Your mixed-methods paper has a thorough quantitative methods section and a thin qualitative one
• Your sample size is justified by the word “saturation” and by nothing else
• A reviewer has previously asked which methodological approach you used
• Your findings section is a list of themes, each followed by a single quotation
• Your qualitative and quantitative results are reported in separate sections that never meet

What qualitative rigour is judged against

Qualitative rigour is judged against reporting standards that specify what the manuscript must state, not against statistical criteria imported from trial reporting. Two standards carry almost all of the weight in health and social science journals, and they differ in scope rather than in strictness.

Neither COREQ nor SRQR is a quality appraisal instrument. A study can answer all 32 COREQ items completely while describing a poorly conceived sampling strategy and an unconvincing analysis, which is why this agent reads the methods and findings text rather than an author-completed form. The COREQ checklist covers the items individually, which items authors most often leave unreported, and where the checklist’s own assumptions fight the method.

The agent’s tenth criterion, integration, has no home in either standard, because both were written for single-strand qualitative studies. Mixed-methods integration is assessed separately below.

The agent brief

The agent brief below is the exact prompt to paste into a custom agent slot; PerfectPaper does not ship a qualitative reviewer as a standing agent, so this text is the whole of it. Suggested settings: work type domain_review, skill level expert, no tools required.

Name: Qualitative and mixed-methods rigour

You are reviewing a qualitative or mixed-methods study against qualitative
reporting standards such as COREQ and SRQR. Apply the criteria of qualitative
research; do not evaluate it as though it were a quantitative study, and do
not treat small sample size as a limitation in itself.

Determine and report:

1. Approach. Determine whether the manuscript names its methodological
   approach: thematic analysis, grounded theory, framework analysis,
   phenomenology, content analysis. Report where none is named, and report
   where the named approach does not match what the methods describe doing.

2. Sampling. Determine the sampling strategy and its rationale: purposive,
   maximum variation, theoretical, convenience, snowball. Report sampling
   described only by its size. Determine whether the strategy fits the
   research question, and report where the recruitment route restricts the
   perspectives that could have been captured.

3. Saturation. Determine whether saturation is claimed and, if so, what
   supports it: when it was assessed, against what, and whether further
   interviews were conducted to confirm it. Report a saturation claim asserted
   without any account of how it was determined. Note where saturation is
   invoked but the approach used does not use the concept.

4. Data collection. Determine whether the interview guide is described or
   provided, who conducted the interviews and their relationship to
   participants, the setting, and whether interviews were recorded and
   transcribed. Report omissions.

5. Coding. Determine how codes were developed, whether inductively or from a
   prior framework, how many people coded, and whether coding was compared
   across coders. Report single-coder analysis presented without any
   discussion of how interpretive consistency was addressed. Do not require an
   inter-rater statistic; many approaches reasonably reject it. Report where a
   statistic is presented as the sole evidence of rigour.

6. Reflexivity. Determine whether the researchers address their own position
   and its influence on data collection and interpretation. Report its
   absence, since it is required by most qualitative reporting standards.

7. Evidence for themes. Determine whether each theme is supported by
   quotations, whether quotations are attributed to distinguishable
   participants, and whether the number of participants contributing to each
   theme can be gauged. Report themes supported by a single quotation, and
   report where all quotations appear to come from a few participants.

8. Analytic depth. Determine whether the analysis moves beyond restating what
   participants said toward interpretation. Report findings sections that
   summarise content without analysis, since this is a common
   substantive weakness in qualitative submissions.

9. Trustworthiness. Determine whether any strategy is described: member
   checking, negative case analysis, audit trail, peer debriefing. Report the
   absence of all of them.

10. Integration. For mixed-methods studies, determine whether the qualitative
    and quantitative strands are integrated in the analysis or merely reported
    side by side, and report the latter.

Judge the work by qualitative standards throughout.

Writing your own variant of this brief follows the same pattern used across the cluster; how to write a custom reviewer agent covers the structure, and custom AI reviewer agents lists the slots a single review can hold.

Questions about what the agent flags

Can an AI reviewer actually handle qualitative research, or is it only built for quantitative papers?

Yes, but only as a separate agent built for the purpose. A general-purpose reviewer built for quantitative manuscripts reads a small sample as a weakness and an absent p-value as an omission, which is the wrong frame entirely. This agent is built to read interviews and focus groups against the standards qualitative reviewers actually apply: sampling rationale, saturation, coding, reflexivity, and quotation evidence.

What happens if I never named a specific method — I just wrote that themes emerged from the interviews?

The agent reports that the analysis cannot be evaluated and names the specific missing statement, rather than guessing an approach from the vocabulary used. An unnamed approach is itself a finding, because saturation, coding structure and the role of reflexivity all mean different things under grounded theory, framework analysis and reflexive thematic analysis.

Does this just check my paper against the COREQ checklist, or does it actually read what I wrote?

The agent reads the manuscript text, not a completed form. A study can answer all 32 COREQ items while describing a poorly conceived sampling strategy and an unconvincing analysis, so the agent checks whether the interviewer is identified, whether the guide and setting are described, and whether the analysis account actually lets a reader judge the interpretation.

What this agent catches

Sampling: purposive is a strategy, not a number

A sampling strategy is the rule that decided who was approached and why, and a manuscript that reports only how many people were interviewed has not stated one. The agent reads for the rule first: purposive, maximum variation, theoretical, convenience, snowball, or a criterion-based frame, each with the reason it fits the research question.

The published alternative to counting is Malterud, Siersma and Guassora’s “Sample Size in Qualitative Interview Studies: Guided by Information Power” (Qualitative Health Research 2016;26(13):1753–1760), which proposes that the more information a sample holds relevant to the study, the fewer participants are needed, and names five dimensions that determine it: the aim of the study, sample specificity, use of established theory, quality of dialogue, and analysis strategy. A sample-size paragraph written against those five dimensions answers the reviewer’s question directly, and it can be written before recruitment rather than defended after it.

A reviewer trained on saturation-based expectations does not always accept an information-power justification on its own terms, even when it is methodologically sound. “We stopped at eight interviews because the sample held enough information for this narrow aim” can still draw a request for more data from a reviewer checking against a remembered number rather than against the five dimensions. The remedy is not a larger sample; it is naming Malterud’s framework directly in the methods section, so the reviewer is evaluating the approach actually used instead of substituting the one they expected.

Recruitment route is the second half of the criterion, and it is the half authors omit. Where every participant was recruited through one clinic, one professional network or one patient organisation, the perspectives that route excludes are a property of the findings, not a footnote — the same structure that makes selection bias a design problem in quantitative work, differently named. Snowball recruitment compounds it, because participants nominate people like themselves. State the route, state what it excluded, and the objection is closed before it is raised.

Common questions about sample size and sampling

I only interviewed 12 people for my study — is a reviewer going to reject that as too small a sample?

Not on the number alone. Qualitative samples are sized for depth and range of perspective, not to estimate a population value, and the agent is briefed not to treat a small N as a limitation by itself. What draws an objection is a sample described only by its size: twelve interviews with no stated rationale reads as arbitrary, while twelve justified against Malterud’s information-power dimensions reads as a decision.

Is convenience sampling something reviewers automatically flag, or is it fine if I explain why I used it?

Convenience sampling on its own is not the defect reviewers flag; an unexamined recruitment route is. Where every participant came through one clinic, one professional network or one list, the agent reports what that route excludes, because the perspectives it could not reach are a property of the findings, not a footnote. State the route and what it excluded, and the objection is largely closed.

Saturation: what the claim has to carry

A saturation claim has to state when saturation was assessed, against what, and whether further data collection was conducted to confirm it. Where the word appears exactly once, in a sentence of the form “data saturation was reached”, the claim is asserted rather than reported, and this is among the most common findings in qualitative review.

There is now empirical guidance to write against. Hennink and Kaiser reviewed 23 articles that tested saturation on real data or by statistical modelling, in “Sample sizes for saturation in qualitative research: A systematic review of empirical tests” (Social Science & Medicine 2022;292:114523), and reported that studies using empirical data reached saturation within a narrow range: 9 to 17 interviews, or 4 to 8 focus group discussions, particularly in studies with relatively homogeneous populations and narrowly defined objectives. Multi-country research and meta-theme analysis needed more.

The number that circulates as a rule of thumb — twelve interviews — traces to an earlier, frequently cited experiment: Guest, Bunce and Johnson’s “How Many Interviews Are Enough? An Experiment with Data Saturation and Variability” (Field Methods 2006;18(1):59–82), which analysed sixty in-depth interviews and reported that saturation occurred within the first twelve, with the basic elements for metathemes present as early as the sixth. The number travels less well than the finding: those sixty interviews were with women in two West African countries on one narrowly defined topic, which is about as homogeneous a study population as an interview study gets. A heterogeneous sample, a broad research question or an interpretively demanding analysis is not entitled to borrow twelve as a target, and the agent does not treat it as one.

The distinction that decides most manuscripts is which saturation is claimed. Hennink, Kaiser and Marconi examined 25 in-depth interviews in “Code Saturation Versus Meaning Saturation: How Many Interviews Are Enough?” (Qualitative Health Research 2017;27(4):591–608) and found code saturation — the point at which the range of thematic issues had been identified — at nine interviews, while meaning saturation, a richly textured understanding of those issues, took 16 to 24. Their formulation is the one to borrow: code saturation indicates when researchers have heard it all, meaning saturation when they understand it all.

The agent also reports saturation invoked where the approach does not use the concept. Saturation originates in grounded theory, where it describes theoretical categories rather than a count of new codes. Braun and Clarke argue the case against transferring it in “To saturate or not to saturate? Questioning data saturation as a useful concept for thematic analysis and sample-size rationales” (Qualitative Research in Sport, Exercise and Health 2021;13(2):201–216). Under reflexive thematic analysis, “saturation was not sought, because the analytic approach does not use the concept, and the sample size rationale was X” is the correct answer, not an omission.

Worked example: a saturation claim that does not survive reading

Consider a manuscript reporting 19 semi-structured interviews with clinicians about a new referral pathway. The methods state that interviews continued until data saturation was reached. The findings present five themes. The paper is well written and the quotations are vivid.

Three things in that manuscript contradict each other, and none is visible in any single sentence. The recruitment paragraph says all 19 participants were recruited at two sites through a clinical lead who also co-authored the paper. The findings attribute quotations to “P3”, “P7”, “P11” and “P14” and to nobody else across four of the five themes. And the analysis section says coding was inductive, conducted by the first author, and that themes were then agreed in discussion with the team.

A reviewer reads those three facts together and asks a single question: if saturation was assessed against new codes, and the codes came from one coder reading a sample recruited by a co-author, what exactly stopped appearing? The saturation claim is not false. It is unevaluable, which in review is worse, because the author cannot answer it without data that were never recorded.

The agent reads the same three artefacts against each other — the recruitment route, the distribution of participant identifiers across themes, and the coding account — because no one of them is a finding on its own. Fixing this at revision takes one paragraph: state that saturation was assessed as code saturation, name the point in the interview sequence at which no new codes were added, state that two further interviews were conducted to confirm it, and state that recruitment through a clinical lead may have narrowed the range of views. Fixing it after rejection takes a resubmission — see why papers get rejected.

Coding: one coder, two coders, and the kappa question

Single-coder analysis is not a defect, and an inter-rater statistic is not a certificate. What draws objections is single-coder analysis presented with no account at all of how interpretive consistency was addressed, and the agent is briefed to report exactly that gap rather than to demand a second coder.

Inter-rater reliability is rarer in practice than manuscripts imply. McDonald, Schoenebeck and Forte analysed a representative sample of CSCW and HCI papers from 2016 to 2018 in “Reliability and Inter-rater Reliability in Qualitative Research” (Proceedings of the ACM on Human-Computer Interaction 2019;3(CSCW)) and found inter-rater reliability reported in around one in nine qualitative papers. Their guidance is the useful part: whether to calculate agreement depends on the epistemology of the approach, and reporting it where the approach treats coding as an evolving interpretive process misrepresents what was done.

The agent therefore reports two symmetrical failures. The first is a lone coder with no audit trail, no peer debriefing and no reflexive account — nothing at all standing where consistency should be discussed. The second is a kappa presented as the whole of the rigour argument, with no description of how the codebook was developed, how disagreements were resolved, or what the coders were disagreeing about.

A number attached to coding agreement is also frequently the wrong statistic for the design that produced it. Cohen’s kappa is defined for exactly two raters coding the same nominal categories; a manuscript describing three or more coders and reporting one kappa value has either dropped raters from the calculation without saying so, or has run Fleiss’ kappa — the version built for more than two raters — and mislabelled it. The agent does not require either statistic, but where one is reported, it is read against the coding design that supposedly produced it, not taken at face value.

Naming the approach is what makes either judgement possible. Braun and Clarke’s “Using thematic analysis in psychology” (Qualitative Research in Psychology 2006;3(2):77–101) sets out a six-phase procedure and objects to the language of themes “emerging” from the data, because that account is passive and conceals the researcher’s active role in identifying and selecting them. A methods section that says themes emerged has, in one verb, declined to describe the analysis — which is the shape of most methods objections in qualitative review.

Questions about saturation and coding

My reviewer wrote that data saturation was not clearly demonstrated — what does that actually mean, and how do I fix it before I resubmit?

A comment like that means the manuscript asserted saturation without saying how it was assessed, against what, or whether further data collection confirmed it — a bare sentence like “data saturation was reached” reads as a claim, not evidence. The fix is one paragraph: name the point in the interview sequence where new codes stopped appearing, and state whether code saturation or meaning saturation was being tracked.

Is it a problem that only one person coded all of my interview transcripts?

Not by itself. A lone coder is not a defect the agent is briefed to flag. What draws an objection is a lone coder with no account at all of how interpretive consistency was addressed — no audit trail, no peer debriefing, no reflexive record of how coding decisions were made. Naming that process, even without a second coder, answers the objection reviewers actually raise.

Do I need to report a kappa score to prove my coding was reliable?

No. A kappa statistic is not required, and reporting one is not, by itself, evidence of rigour either — it is defined for two raters, so a three-coder team reporting a single value has usually mishandled it. What the agent looks for is a description of how the codebook was built and disagreements resolved, statistic or not.

Reflexivity: the domain with the worst compliance

Reflexivity is the COREQ domain authors most often skip and reviewers most often use as a diagnostic. De Jong and colleagues extracted 2,775 COREQ appraisals of individual qualitative studies from 139 reviews, in “A meta-review demonstrates improved reporting quality of qualitative reviews following the publication of COREQ- and ENTREQ-checklists, regardless of modest uptake” (BMC Medical Research Methodology 2021;21:184), and reported a mean domain 1 score of 2.86 out of 8 for studies published after 2008, against 8.51 out of 15 for study design and 6.20 out of 9 for analysis and findings.

The item-level picture is where the agent aims. In that same post-2008 sample, participant knowledge of the interviewer was reported in 16% of studies, a relationship established before the study in 18%, interviewer characteristics in 23%, and experience and training in qualitative methods in 26%, while the identity of the interviewer or facilitator was given in 58%. Every one of those items concerns the person who asked the questions, on the premise that in interview research the researcher is an instrument whose properties belong in the methods section the way an assay’s properties would.

Reflexivity is often read as one undifferentiated item, but it separates into two components a reviewer weighs differently. Personal reflexivity discloses who the researcher is, what relationship they had to the topic or participants, and how that shaped the interaction — this is what COREQ’s domain 1 items test, and it is the half de Jong’s item-level figures describe. Epistemological reflexivity addresses how the research question, the interview guide and the analytic lens were themselves shaped by the assumptions built into the chosen methodology; it is not itemised on any checklist and is instead judged from how the discussion section handles alternative readings of the same data. A manuscript can satisfy every domain 1 item and still read as reflexively thin at the second, harder-to-fake level.

A reflexivity paragraph that satisfies a reviewer has four parts and runs to about 120 words: who conducted the interviews and their professional role; what relationship, if any, existed with participants before the study; what the researchers’ prior assumptions about the topic were; and one concrete instance of how that position shaped a decision — a question added to the guide, a line of enquiry not pursued, a reading of a transcript revised in discussion.

The last part is what cannot be manufactured at revision. A concrete instance of position shaping a decision is a record of fieldwork, and written afterwards to satisfy a form it reads exactly as retrospective. Working the domain into the first draft is a matter of academic writing practice rather than compliance.

Evidence for themes: quotations, attribution, and how many participants

A theme is a claim about a pattern across participants, so the evidence a reader needs is how many participants contributed to it. Quotations were present in 92% of the post-2008 studies de Jong and colleagues appraised, which makes their presence close to uninformative; their attribution is what carries the argument, and the agent counts distinguishable participant identifiers rather than counting quotations.

Three patterns trigger a finding. A theme supported by exactly one quotation is a claim about a pattern evidenced by a single instance. A findings section presenting twelve quotations under four distinguishable identifiers is formally compliant and substantively hollow, and a reviewer notices before any checklist does. And quotations with no identifiers at all — no “P7”, no role label, no line reference — make the distribution unrecoverable, which converts every theme in the paper into an assertion.

The remedy is cheap and belongs in the findings rather than the methods. Attach a stable identifier to every quotation. State, in one clause per theme or in a single summary table, how many participants contributed to it. Where a theme rests on a minority view, say so and say why it was retained — a negative case reported deliberately is a strength, and the same instance found by a reviewer is a defect.

A wide distribution can still mislead in a different way. Quotations often skew toward the most articulate or most extreme participants in the sample, because those are simply the easiest passages to select and the most quotable — not because they are the most representative. A theme evidenced by identifiers spread across twelve of nineteen participants looks well supported until a reader checks whether those twelve say materially different things from the other seven, or whether the seven were just harder to quote. The agent flags a skewed distribution the same way it flags a narrow one; a wide spread of citations is not, on its own, evidence that the quotations were representative rather than merely convenient.

The failure this prevents is an overreach in the abstract rather than in the findings. A theme evidenced by four of nineteen participants and described in the abstract as characterising clinicians’ views is an overclaim, and it is the sentence editors quote back.

Analytic depth: description presented as analysis

A common substantive weakness in qualitative submissions is a findings section that summarises what participants said without interpreting it, and it is a weakness no reporting checklist detects. A paper can satisfy every COREQ item, attribute every quotation, and still present themes that are topic labels — “communication”, “time pressure”, “training needs” — followed by paraphrase and a quotation.

The diagnostic the agent applies is whether a theme makes a claim. “Time pressure” names a topic; “clinicians described the referral pathway as safe in principle but treated it as optional under time pressure, which is where the two accounts of the same protocol diverge” makes a claim that the quotations can support or fail to support. A theme stated as a claim can be argued with. A theme stated as a topic cannot, and a reviewer who cannot argue with your themes cannot endorse them.

Three tells separate the two in a manuscript. Themes that could be swapped between any two papers on the topic. A findings section whose structure mirrors the interview guide, section for section, which means the analysis reproduced the researcher’s questions rather than the participants’ account. And a discussion that repeats the findings in different words rather than relating them to anything.

Depth is not length. It is a stated relationship between what participants said and what the researcher concluded from it, and it survives the objection that any competent reader could have written the findings from the transcripts alone.

Trustworthiness: what member checking can and cannot establish

Trustworthiness is demonstrated by a described strategy, not by a named one, and the agent reports the absence of all of them rather than the absence of any particular one. Lincoln and Guba’s four criteria in Naturalistic Inquiry (1985) — credibility, transferability, dependability and confirmability — remain the vocabulary most journals expect, and the strategies that serve them are member checking, negative case analysis, an audit trail and peer debriefing.

Member checking is the one most often listed and least often examined. Birt, Scott, Cavers, Campbell and Walter set out the critique in “Member Checking: A Tool to Enhance Trustworthiness or Merely a Nod to Validation?” (Qualitative Health Research 2016;26(13):1802–1811), observing that member checking is often mentioned as one item in a list of validation techniques, which neither acknowledges the value of the method nor addresses its uneasy fit with the interpretative stance of qualitative research. Their alternative, Synthesized Member Checking, returns interview data together with the interpretation, months later, so participants engage with and add to the analysis rather than confirming a transcript.

Two distinctions the agent enforces follow from that. Returning transcripts for correction and returning findings for comment are different acts serving different purposes, and manuscripts conflate them constantly. And participant agreement with an interpretation is not validation of it, because a participant is not positioned to assess a cross-case reading of nineteen interviews they did not see.

Neither COREQ nor SRQR requires you to have done member checking. COREQ asks directly whether participants gave feedback on the findings, and SRQR asks which techniques were used to enhance trustworthiness and why, which means “member checking was not undertaken, because the analysis is a cross-case interpretation participants could not assess” is a complete answer.

Mixed methods: integration is the item that is missing

Integration is what makes a study mixed methods rather than two studies in one manuscript, and it is the component most often absent. O’Cathain, Murphy and Nicholl assessed 118 mixed-methods studies funded by the Department of Health in England between 1994 and 2004, obtaining proposals or final reports for 75, and reported in the Journal of Health Services Research & Policy (2008;13(2):92–98) that researchers mainly ignored the mixed-methods design and described only the separate components, and that integration of data across components was rare. Transparency was worse for the qualitative component than the quantitative one: 42% of proposals lacked a clear exposition of the qualitative data collection and analysis, against 18% for the quantitative component.

Fetters, Curry and Creswell give the vocabulary a reviewer expects, in “Achieving integration in mixed methods designs — principles and practices” (Health Services Research 2013;48(6 Pt 2):2134–2156). Integration operates at three levels. At the design level it runs through three basic designs — exploratory sequential, explanatory sequential and convergent — plus four advanced frameworks. At the methods level it takes four forms: connecting, where one database links to the other through sampling; building, where one informs the data collection of the other; merging, where the two are brought together for analysis; and embedding, where collection and analysis link at multiple points. At the interpretation and reporting level it appears as narrative, data transformation, or a joint display.

Design notation makes the priority question explicit in a single string, following the convention Morse set out in “Approaches to Qualitative-Quantitative Methodological Triangulation” (Nursing Research 1991;40(2):120–123): capitals mark the strand with primacy, a plus sign marks concurrent implementation, an arrow marks sequential. QUAL+quan is therefore a concurrent design in which the qualitative strand is weighted more heavily, qual→QUAN a sequential design in which a supplementary qualitative phase feeds a dominant quantitative one. A manuscript that never states its own notation has usually not decided which strand the study privileges, and that indecision is exactly where integration goes missing — connecting, building, merging and embedding all require knowing which database is doing the informing.

A joint display is the single simplest repair available to a mixed-methods manuscript: one table whose rows are the constructs and whose columns are the quantitative result, the qualitative finding, and a statement of fit — confirmation, expansion or discordance. It converts “the two strands were integrated” into something a reviewer can check.

Frequently asked questions

My advisor keeps telling me to address my positionality in the write-up, but I’m not totally sure what that means or how much I’m supposed to say — where do I even start?

Start with four things: who conducted the interviews and their professional role, what relationship existed with participants beforehand, your own assumptions about the topic going in, and one concrete instance where that position shaped a decision — a question added to the guide, a line of enquiry not pursued, a reading revised in discussion. That last part is what a reviewer checks for, because it cannot be manufactured afterward.

Can PerfectPaper review a mixed-methods manuscript, or only pure qualitative studies?

Yes — this agent covers the qualitative strand while PerfectPaper’s standing quantitative team reads the other, and its tenth criterion is specifically whether the two strands are integrated in the analysis or only reported side by side. Run both agents on the same manuscript; a joint display table is the simplest evidence that integration actually happened.

What reviewers write when qualitative rigour is mishandled

Reviewers rarely write “the qualitative methods are weak” first. They write the specific version, and these are the comments that decide the paper.

“The authors state that data saturation was achieved but provide no account of how this was determined.” “The methods do not name a methodological approach; thematic analysis is asserted without reference to a specific formulation.” “Coding was performed by a single author with no description of how interpretive consistency was addressed.” “No information is given about the researchers’ background, training or prior assumptions; reflexivity is not addressed.” “It is not clear who conducted the interviews or what relationship they had with participants.” “Please state how many individuals were approached and how many declined.” “The quotations presented under theme 3 appear to come from the same two participants; please indicate how many participants contributed to each theme.” “The findings section summarises participant accounts but does not interpret them; the themes read as topic headings.” “The qualitative and quantitative results are presented separately with no integration; please state what the qualitative strand added to the quantitative finding.”

The last two are the ones authors find hardest to answer, because neither has a data fix. An interpretation cannot be added at revision without reanalysis, and integration cannot be retrofitted onto two strands that were designed to run in parallel. Both can be answered honestly by narrowing the claim, which is the substance of how to write a response to reviewers.

When the obvious remedy is unavailable

Re-interviewing participants is frequently impossible, and the honest alternative is a narrowed claim with the limitation named, not a retrospective account of work that was not done. When ethical approval has expired, participants have withdrawn consent, the fieldworker has left the project, or no reflexive notes were kept, the manuscript still has four moves.

Reanalyse rather than recollect. A second coder can be added to existing transcripts, a negative case search can be run over data already held, and a codebook can be documented after the fact provided the manuscript says it was documented after the fact. None of these requires new participants.

Report the distribution you have. Where quotations cluster on a few participants, publishing the count per theme converts a hidden weakness into a disclosed one. Reviewers respond very differently to a distribution they were shown than to one they discovered.

Answer the item with the reason it does not apply. “Saturation was not sought, because the analytic approach does not use the concept” and “member checking was not undertaken, because the analysis is a cross-case interpretation” are compliant answers. Blank, “N/A”, or a claim to have done something you did not, is worse than either.

Write the limitation with a direction. State which perspectives the recruitment route excluded and how that would shape the themes, rather than writing that findings should be interpreted with caution. A limitation that names the mechanism and the likely direction reads as an analysis; a generic acknowledgement reads as an admission. How to write a limitations section sets out the form, and note that the qualitative frame is transferability and reflexivity, not bias and generalisability.

How to report qualitative methods so a reviewer can check them

Report the qualitative methods as a sequence of named decisions with named agents, because every criterion above is answered by one of them. Five sentences carry most of the value, and they can be drafted before the analysis is finished.

Name the approach and the formulation: thematic analysis in a specific formulation, grounded theory in a specific tradition, framework analysis, phenomenology, or content analysis. Name the sampling strategy, the recruitment route, the number approached and the number who declined. Name who conducted the interviews, their role, their relationship to participants, and what participants knew about them. Name how codes were developed, how many people coded, and how disagreements were handled. Name the sample-size rationale, whether that is code saturation with the point at which it was assessed, meaning saturation, or information power against its five dimensions.

In the findings, attach an identifier to every quotation and state the number of participants contributing to each theme. In a mixed-methods paper, add a joint display. Where a journal requires a completed COREQ form, complete it against the final manuscript with a page and paragraph reference per item, after the last revision rather than before, because page references drift with every edit — and check whether your target journal names a guideline at all, which is a journal selection question rather than a formatting one.

How this differs from the built-in review

The standing team covers survey and instrument design, methodology, and logical consistency, all built for quantitative manuscripts. Applied to an interview study, that team will read the paper as though a small sample were a weakness and a lack of statistical testing an omission, which is the wrong frame. This agent supplies the correct one, and the brief instructs it explicitly not to treat sample size as a limitation in itself.

The difference is not tone but criteria. A quantitative reader asks whether the analysis supports an inference to a population; a qualitative reader asks whether the account of the analysis lets a reader judge the interpretation. Those questions have different evidence, and the second is the one that decides qualitative papers — which is why reviewer says the sample size is too small is usually the wrong objection to prepare for here.

If your paper is mixed methods, this agent covers the qualitative strand and the standing team covers the quantitative. Full set: AI peer review for epidemiology.

Review my manuscript

Related agents

The agents below share a review with this one; you can run three at a time. See AI peer review for epidemiology for the full set of five population-science agents.

• Health equity reporting — use it for how race, ethnicity and socioeconomic position are measured and interpreted.
• Competing risks and time-related bias — use it when death from other causes is common, or exposure is defined after follow-up starts.
• Registry and claims data validity — use it for SEER, NCDB and claims analyses that may outrun what the source records.
• Causal language discipline — use it when an observational finding is written with causal verbs.
• Ethics and data sharing — use it because COREQ contains no ethics item, so consent, approval and data availability are not covered by the checklist your journal asked for.