What can we learn from these three creatine studies, and which conclusions change when the population, protocol and endpoint change?
The audience is an R&D or scientific marketing team. AI organises the supplied documents, aligns fields for comparison and prepares a preliminary synthesis for professional review.
01Identify the document and version
02Extract data with a reference
03Check comparisons and uncertainty
04Deliver a synthesis and open questions
02 / The studies
One ingredient. Different questions.
Open a study and follow the data to its source. These are editorial summaries assisted by AI; the references allow you to check them.
S01 · PMID 10449017
Performance and muscle fiber adaptations to creatine supplementation and heavy resistance training
Neither value is below 0.05. P-values do not measure the size of a benefit. They are therefore not displayed as efficacy bars.
03 / The matrix
One row per study. No invented average.
The two muscle trials and the cognition trial measure different outcomes. This matrix places them side by side without pooling results, scoring efficacy or treating them as a meta-analysis.
Comparative matrix of the teaching corpus: extraction from public abstracts
Two deliberately constructed errors show what review needs to catch. The initial sentences are simulations, not an experiment log or output attributed to a specific model.
Synthesis to correct · simulation / S02
Creatine adds 3.3 kg of muscle compared with placebo.
Corrected version / S02
In the trial of older men undertaking training, lean tissue increased 3.3 kg with creatine and 1.3 kg with placebo: a 2.0 kg difference between changes.
Review criterion: Correct the comparison, retain the population and use the name of the measured outcome.Open the original abstract
Synthesis to correct · simulation / S03
5 g per day significantly improves memory and reasoning.
Corrected version / S03
At 5 g/day for 6 weeks, neither main test meets p<0.05. A possible small effect appears in the Bayesian analysis and needs further investigation.
Review criterion: Separate statistical approaches and retain the uncertainty expressed by the authors.Open the original abstract
What makes it into the final dossier?
Retain
Findings, units, comparator, context and source.
Separate
Observed data, editorial calculations and interpretations.
Keep open
Product transferability, safety and regulatory assessment.
05 / The toolkit
Take the method into your next project.
Actual materials from this case, freely downloadable. The PDF combines the synthesis, evidence and checks; the CSV retains the fields; the prompt lets you repeat the process with other documents.
ROLE
Act as a scientific extraction assistant for an R&D team. Prepare a preliminary ingredient dossier from the supplied documents only.
BRIEF
Ingredient: [name]. Question: [population, intervention, comparator, outcomes]. Corpus: [files or readable texts, version and date].
SOURCE CONTROL
Before analysis, list the documents actually read. Distinguish full text from abstract, inaccessible file and secondary source. Do not infer missing information or pretend to have read a PDF.
EXTRACTION
One row per study: ID, author/year, DOI/PMID/URL, document level, design, population, sample and groups, intervention, dose and unit, duration, comparator, endpoint, result, uncertainty, adverse events, limitations, exact source section/table/page. Use NOT REPORTED when a field is absent; use NOT VERIFIED when you cannot check it.
COMPARISON
Separate within-group change from between-group difference. Retain the original outcome name and units. Label any calculation explicitly, showing its inputs. Do not pool heterogeneous endpoints. Separate frequentist and Bayesian conclusions. Statistical significance does not establish clinical relevance.
DELIVERABLES
1. Comparative matrix. 2. Executive synthesis of at most 200 words with study IDs for every factual claim. 3. Data, interpretation and open questions in separate blocks. 4. Decision log: initial sentence, corrected sentence, reason and source. 5. Checklist for human review.
REVIEW
Recheck every number, population, comparator, DOI and locator against the supplied source. Keep uncertainty and negative findings. Do not produce dosing advice, product efficacy promises or regulatory authorisation. Record unresolved issues; do not invent a human reviewer or completed validation.
CASE CORPUS
S01: PMID 10449017 - public abstract.
S02: PMID 11740307 - public abstract.
S03: PMID 37968687 - public abstract; open-access full text available.
Historical teaching selection, not an exhaustive literature review.
The effects of creatine supplementation on cognitive performance—a randomised controlled study
DOI 10.1186/s12916-023-03146-5 · PMID 37968687
Editorial selection of three trials to illustrate the workflow; sources checked on 2 October 2026. The historical studies explain comparisons and do not, by themselves, represent the current state of evidence. AI assisted organisation and drafting: no measured time savings or performance claims are made. This case does not recommend doses, establish commercial claims or judge finished-product efficacy.
AI advisory
Need to make a dossier easier to read?
We start with your company’s documents and define an extraction, comparison and review workflow.
GMA