The Literature Review Playbook

Start here

Why this guide exists

It is week six. You have 340 papers in a folder, roughly forty of which you have actually opened. Your notes live in three places and agree with none of them. You have read enough to feel behind and not enough to feel oriented. Then your supervisor reads your first draft chapter and writes the sentence every researcher gets at least once: this is descriptive; where is your critique?

That comment is not really about writing. It is about process. A review reads as descriptive when it was built by collecting papers and summarising them one at a time, which is what almost everyone does the first time, because nobody teaches the alternative until you have already got it wrong.

This guide teaches the alternative. It walks through seven stages, from framing a question to writing the final chapter, and at each stage it tells you what artefact you should end up holding. If you produce those artefacts as you go, the review largely writes itself, your methods section is already documented, and the gap you claim at the end is one you can actually defend.

What this guide does differently

• It distinguishes between review types. A narrative review and a systematic review are not the same task and do not carry the same obligations. Stage 0 helps you work out which one you are doing before you commit six months to it.
• It treats AI as a component, not a theme. No stage of a rigorous review is removed by AI. What changes is how much of each stage is mechanical effort and how much is judgement, and the gain is concentrated in searching, screening and extraction, where the same operation repeats across hundreds of records.
• It is built around artefacts, not advice. Nine templates ship with this guide as editable Word and Excel files. The guide is the instructions; the templates are the work.
• It shows the work. One worked example, on mobile health interventions and medication adherence, runs through every stage so you can see each artefact filled in rather than described.

Who this is for, and who it is not

How to use it

Read it once end to end before you start searching, then come back to one stage at a time as you reach it. If you are already mid-review, the process map below will tell you which stage you are actually at. Two checklists come with this guide: the one-page Quick Start Checklist in the toolkit, which gets you moving in ten minutes, and the fuller master checklist near the end of this playbook, designed to be printed and ticked as you work.

Two things are worth knowing before you begin. This process is front-loaded, so the decisions that determine whether your review holds up are made before you have read a paper properly. And the stages are iterative even though they are numbered: you will revise criteria after piloting and revise your search after screening. That is normal. What causes problems is revising silently.

The shape of the whole thing

The review on one page

Most reviews go wrong not because a stage is done badly but because a stage is skipped, and the cost only becomes visible months later. This is the full sequence, and what each stage has to hand over to the next.

Before anything else

Which review are you actually doing?

This determines everything downstream: how many databases you search, whether you need a protocol, whether two people must screen, and how long the work will take. Getting it wrong is expensive, because a systematic review cannot be assembled retrospectively out of a narrative one.

Three questions that settle it

• Is your question answerable from existing studies as stated? If yes, a systematic review is possible. If your question is really about mapping a territory, it is not, and forcing it will produce a weak review.
• Are you working alone? Full systematic reviews assume two screeners. Solo researchers usually agree on a documented compromise with a supervisor, such as dual screening a ten percent sample and reporting the agreement level. Agree on it in advance.

Stage 0

Framing the question

A vague question produces an unmanageable result set and a review with no spine. Your question needs a structure that names its component concepts, because those concepts become your search blocks, your inclusion criteria, and eventually your themes.

Use a framework, not instinct

Write the criteria down before you search

Criteria written after screening has begun are not criteria; they are rationalisations. Decide them now and record why. Date and language limits are the two most frequently challenged decisions in any review, so if one exists for capacity reasons, say so and list it as a limitation. An examiner will accept a declared constraint far more readily than an unexplained one.

Keeping scope, and the four questions that do it

Scope creep is rarely a single decision. It is twenty small ones, each defensible alone, that turn a six-week chapter into a six-month one. The defence is writing the boundary down before the pressure to move it arrives.

• What is the unit of analysis? Studies, interventions, participants, or documents? Reviews drift when this is left implicit, because a single paper reporting three interventions gets counted differently on different days.
• What counts as the outcome? Name the constructs you will accept and the ones you will not. If you accept anything adjacent, your synthesis will compare things that are not comparable.
• Where does the field end? Every topic has a neighbouring literature that is interesting and not yours. Write down which adjacent field you are deliberately not reviewing, and why. This sentence goes in your limitations and saves you an examiner question.
• What is the earliest date that makes sense? If there is a real inflection point, a technology, a policy, a landmark study, anchor to it and say so. A date chosen because it produces a manageable number is a capacity decision, which is fine, but must be declared as one.

Turning criteria into screening rules

Criteria and screening rules are not the same thing. A criterion states a principle; a screening rule tells you what to do with the record in front of you at eleven at night when it half fits. Write the second kind.

Stage 1

Building a search that holds up

A search has two jobs that pull against each other: find everything relevant, and stay small enough to screen. The way you resolve that tension has to be visible to someone else, which is why the search is built in blocks rather than typed as a sentence.

Anatomy of a search string

Every concept from your framework becomes one block. Inside a block, you use OR to catch every phrasing authors might have used. Between blocks, you use AND to require all concepts.

Find the terms you would not have thought of

Missed literature is almost always a vocabulary problem rather than an effort problem. Papers you need sit under terminology from a neighbouring field, or under a term the field has abandoned. Start by harvesting terms from three papers you already know are relevant: if your string does not retrieve those three, the string is wrong, not the database. Then close the rest of the gap two ways.

• Use controlled vocabulary. Many databases index articles against a fixed thesaurus: MeSH in PubMed, Emtree in Embase, the ERIC thesaurus in education. A thesaurus term catches papers regardless of the words the authors chose. Combine thesaurus and free text terms rather than choosing between them.
• Ask a language model for the terminology map, then verify it. One of the safest AI tasks in the process, because every suggestion is checked the moment you run it. Ask for synonyms, superseded terms, adjacent field terms, and abbreviations, then test each and keep what returns relevant results.

Databases are not interchangeable

Overlap between databases is lower than people assume, and the papers you miss are the ones indexed somewhere you did not look. Two or three well-chosen sources plus supplementary searching beats five chosen at random. Pick one broad multidisciplinary source such as Scopus or Web of Science, one specialist source for your field, and at least one grey literature route. Expect to rewrite your string for each.

The string will not survive the move between databases

Search syntax is not standardised. A string that works in one platform will silently return the wrong thing in another rather than throwing an error, which is the dangerous part. Check these four before you trust a result count.

The searching that happens after the search

A database search alone is not a comprehensive search, and reviewers know it. Four supplementary methods close most of the gap, and each takes a fraction of the time the main search took.

• Backward chaining. Read the reference lists of your most relevant included papers and any existing review on the topic.
• Forward chaining. Find everything that has cited those key papers since publication. This catches recent work your keywords missed.
• Grey literature. Theses, preprints, government and NGO reports, trial registries. Studies with null results are systematically less likely to be published, so a review built only on journal articles inherits publication bias. Search at least one grey source and say which.
• Hand searching. Scan the last two years of the two or three journals that publish most in your area.

A real search log: the date searched, the query as entered, the query as executed against each source with counts, and how duplicates were removed.

Stage 1, continued

Documenting the search

A review is judged partly on whether someone else could repeat it. That requirement explains most of the record-keeping in this stage, and it shapes how AI-powered search tools can be used.

The reproducibility problem with AI search

A Boolean string is reportable. Someone else can paste it into the same database and get the same result. Semantic retrieval works differently: results depend on a model, and models are updated. The same query next year may not return the same papers, and there is no string to publish.

This does not make AI-assisted tools unusable in a systematic review. It means the identification stage specifically has to stay deterministic: a fixed string, a corpus query, a rule-based deduplication, and a count anyone could reproduce. Where a tool applies AI later, at screening or extraction, that is a different matter, because those stages are decisions you review one at a time rather than a retrieval you have to reproduce.

The search log

One row per database per attempt, kept as you go. It becomes your methods section almost verbatim, and it is the artefact people most regret not keeping.

Set up citation management now

Zotero, Mendeley or EndNote, chosen on day one and used consistently, will save you a week. Export results directly from each database rather than adding records by hand, use collections that mirror your screening stages, and attach the PDF when you retrieve it. Deduplicate inside the manager and record how many duplicates were removed, because that number goes in the flow diagram.

Stage 2

Screening without losing the thread

Screening is where a review is most often quietly compromised, because it is repetitive, it happens under fatigue, and every decision is invisible unless you record it. The process itself is simple. The discipline is in the logging.

Two stages, not one

At title and abstract, assess every deduplicated record against your criteria. The decision is deliberately crude: could this plausibly meet the criteria, yes or no. You need only a total for exclusions at this stage. At full text, retrieve and read every surviving record against the same criteria, and here every exclusion needs a recorded reason, because those reasons are reported. This is also where you discover that abstracts routinely misrepresent what a paper did.

Deduplicate before you screen, not during

Records overlap heavily across sources. Deduplicating before screening prevents the same paper from being screened twice with different outcomes. Automatic matching on DOI and title handles most of it; the residue needs a manual pass, because two failure cases are predictable.

• Same record, different metadata. Missing DOIs, abbreviated journal titles and author initials defeat exact matching. Sort by title and scan.
• Companion papers. Trials often publish protocol, primary outcome and secondary analysis separately. These are one study, and treating them as three inflates your evidence.

Calibrate first, then guard against drift

Pilot on the first fifty records, then check whether your decisions were actually driven by what you wrote down. Criteria almost always need one revision after contact with real records, which is fine provided you revise before the main screen and record that you did. Revising silently halfway through is the problem. With a second screener, both screen the same fifty independently and report an agreement statistic such as Cohen’s kappa. After that, screening becomes a fatigue problem: the decisions at record 400 are not made with the care of those at record 40, and the drift is invisible from the inside.

• Screen in blocks of about fifty, and stop when you notice you are reading the title only.
• Keep your criteria visible on screen. Recalling them from memory is where drift starts.
• When unsure at title and abstract, carry the record forward rather than excluding it. An uncertain pile you settle deliberately is fine; an over-excluded record is gone silently.

Report the numbers as a flow diagram

Every count must be recorded at the moment it happens. This example uses the worked question from Stage 0.

Illustrative numbers only. Record every count as you go, because reconstructing them afterwards is close to impossible and your methods section depends on them.

Stage 3

Appraising what you found

This is the stage most often skipped, and skipping it is what turns a review into an annotated bibliography. Appraisal is the difference between reporting that three studies found an effect and being able to say that two of them were at high risk of bias and the third was not.

Match the tool to the design

The instruments look intimidating and are not. Each is a short structured checklist with published guidance, and after three or four papers you will apply one in fifteen minutes. Whichever you use, the underlying questions are the same: could the groups have differed at the start, could the measurement have produced the effect, who dropped out, was the analysis decided before the results were seen, and is the claim bigger than the design supports.

What to do with the ratings

• Record a rating for every included study in your evidence table, never as a separate document, you will forget to reconcile.
• Weight your synthesis by quality. A finding supported by three strong studies and contradicted by one weak one is not a contested finding.
• Consider a sensitivity check: would your conclusions change if the high-risk studies were removed? If yes, that is a finding in itself and belongs in your discussion.
• Do not exclude studies for low quality unless your protocol said you would. Appraise them, include them, and weight them accordingly.

Stage 4

Extraction and synthesis

The single biggest upgrade available to most reviews is replacing free-form notes with one consistent extraction table covering every included study. Free-form notes force you to re-read everything when you sit down to write. A table lets you read across studies, which is where synthesis actually happens.

The evidence table

Extract the same fields, in the same order, for every study, on the day you read it. Six columns are shown here for space; the full template has twelve, including theoretical framework, effect size, funding source and page references for quotable passages.

Read across the table, not down it

Once fifteen to twenty rows are filled, stop reading and start comparing. Themes come from patterns across rows, and those patterns are usually invisible while you are still reading papers one at a time. Four questions surface most of them.

• Which findings contradict, and does that track a difference in population, measurement or design?
• Which method or measure dominates the field, and what would a different one likely reveal?
• What appears in the limitations column repeatedly across unrelated studies?

Build a synthesis matrix, then turn themes into claims

Turn the themes into columns and the studies into rows, then mark which study speaks to which theme. Sparse columns are candidate gaps, dense columns are your main sections, and each column becomes one or two paragraphs of the written review.

Then take the step people skip: write each theme as a claim that could, in principle, be wrong. Mobile reminders and adherence is a theme and tells the reader nothing. Mobile reminders produce a short-term effect; the evidence cannot show it is sustained is a claim, and the section becomes the evidence for and against it.

Name your synthesis method

Handling contradictions

A contradiction between studies is not a problem to smooth over. It is usually the most interesting thing in the review, and the sentence findings were mixed is where a promising chapter goes to die. Mixed findings are a question, not a conclusion, and the question has a small number of answers.

If the studies really do disagree and nothing accounts for it, say so explicitly. An unexplained contradiction you have actively tested is a legitimate finding, and frequently the most defensible gap available to you.

Stage 5

Finding and testing the gap

Most reviews describe what has been studied clearly, then arrive at what is missing almost by accident in the last fortnight of writing. Gap finding is its own stage, run against your evidence table rather than against memory.

Gaps come in types

Test the gap before you build on it

A gap that survives ten minutes of scrutiny is worth a chapter. One that does not will be found by your examiner. Run all four checks.

• Search specifically for the gap. Use the terms that a study filling it would use. If three papers appear, the gap is your search, not the field.
• Check that it is not a gap for a reason. Some things are unstudied because they are unmeasurable, unfundable or unethical. Establish which applies.
• Make it narrow enough to justify one study. If your gap statement could support twenty different dissertations, it is a topic, not a gap.
• Say why it matters. A gap with no consequence is a hole, not an opportunity. Name what remains unresolvable because it exists, and name the design feature your own study has that the reviewed studies lacked.

Stage 6

Writing the review

The structural mistake that generates the descriptive comment is organising by paper. Smith found X, then Jones found Y, then Patel found Z. Each sentence is accurate, and the paragraph argues nothing. Strong reviews are organised by claim, with studies cited as evidence inside them.

What descriptive and critical actually look like

What changed

• The paragraph opens with a claim, so the reader knows what the evidence is being used to argue.
• Studies are grouped by what they show rather than listed in the order they were read.
• The disagreement is explained by pointing at a methodological difference, not just noted.
• It closes by naming what the literature cannot currently establish, which is the gap arriving in its natural place.

A structure that works for a chapter

• Open with the problem, not the literature. One or two paragraphs on why this question matters, before any citation.
• Declare scope and method. What you searched, when, against what criteria, and what you excluded. This is the search log written into prose.
• One section per theme. Each opens with a claim, presents agreeing evidence, presents contradicting evidence, and explains the difference.
• A synthesis section that crosses themes. What the field collectively knows, and how confident it is entitled to be given the quality of the evidence.
• The gap, then your study. Stated specifically, connected to what you are proposing to do about it.

Paragraph structure that produces critical prose

Most descriptive writing is a structural problem rather than a thinking problem. This shape forces the argument to the front and is worth applying mechanically until it becomes habit.

Reporting verbs carry judgement

The verb you attach to a study is an appraisal, and inconsistent verbs make a review read as though it has no view.

Four checks before you call it done

Proportions matter more than word counts. The failure mode is always the same: the descriptive middle swells and the synthesis gets whatever space is left. If your themes consume three quarters of the chapter, the review is still summarising rather than arguing. Then read the first sentence of every paragraph as a single list. If that list reads as an argument, the chapter works.

• Paragraphs that begin with an author name. Almost always a sign the paragraph is organised around a paper rather than an idea.
• A citation appearing once and never returning. Either it supports a claim that matters, or it is padding.
• Hedging everywhere. If every finding is qualified equally, you have not appraised anything. Reserve hedges for the studies that earned them.

Working method

Using AI without compromising the review

The difference between safe and unsafe AI use here is not about capability. It is about whether the output can be verified cheaply. Terminology suggestions are checked the second you run them. A fabricated citation may not be caught until your viva.

What verification actually means

Verify is easy to write and easy to skip. For a citation, resolve the DOI before it enters the draft; if it cannot be found, it does not exist, whatever the formatting suggests. For an extracted field, locate the sentence it came from and record the page number. For a screening summary, read the abstract before excluding anything borderline. For edited prose, confirm that no fact, number or citation has appeared that you did not write.

Ground every prompt in text you supply

Most fabrication happens when a model is asked to recall rather than to process. A prompt that includes the abstract, paragraph or table you are working with is far safer than one asking a general question, and the output is more useful.

Declare what you used

Policies on AI use vary from prohibited to expected and change faster than departmental handbooks. Check your institution’s current policy, keep a record of which tools you used at which stage, and disclose it in the form asked for. A brief methods sentence is normally sufficient.

Quick reference

The master checklist

Print this. Tick as you go. Anything still unticked when you reach the writing stage is something you will have to reconstruct under time pressure.

Take these with you

Templates and further reading

The nine templates

Each is referenced at the stage where it is used and designed to be filled in as you work. Templates 1 to 3 are in the companion Worksheets document; 4 to 9 are sheets in the companion workbook, where the flow counts calculate themselves from your screening decisions and check that the numbers reconcile.

Sources worth owning

• Booth, A. et al. Systematic Approaches to a Successful Literature Review. The most practical single volume.
• Page, M. J. et al. The PRISMA 2020 statement. BMJ. The reporting standard itself, and shorter than expected.
• Higgins, J. P. T. et al., eds. Cochrane Handbook for Systematic Reviews of Interventions.

Running this as a pipeline

Every stage here can be done with a reference manager and a spreadsheet, and the templates are built for that. ResearchCollab’s systematic review workflow runs the same sequence as a guided pipeline: a PICO intake, a locked search string and criteria, a deduplicated identified count, criteria applied to titles and abstracts with a reason per decision, a comparable evidence table, and sections drafted only from what the earlier steps recorded.

It ends with what reviewers ask for first: a PRISMA flow whose counts reconcile, the PRISMA 2020 checklist, a decision log giving the reason and criterion for every include and exclude, an evidence table, risk of bias figures, and a draft manuscript with numbered references. The draft is a scaffold to argue with, not a finished review.

A PRISMA 2020 flow diagram whose counts reconcile, built from the screening decisions rather than typed in afterwards.
* Records identified from databases and registers combined. ** Records excluded during screening against the predefined inclusion and exclusion criteria.

A shorter walkthrough of the same seven stages, with the templates offer, is on the Literature Review Toolkit page.

Version 2.1, August 2026. This guide describes general research practice and is not a substitute for your institution’s requirements. Check departmental and journal policies on review methodology and AI use before applying anything here to submitted work.