Source-grounded research workflow
Produce a cited synthesis where every claim traces to a quote from a source you supplied, then audit it for unsupported sentences.
1Label your sources
Paste each source with a short ID header: [S1] Title, author, date, URL, followed by the full text. Labelling up front is what makes every later citation checkable. Drop sources you would not be comfortable citing to your reader.
4 steps and 3 ready-to-use prompts are in the full version
Get this resource with editable variables, examples and future updates as part of a download pack.
Get Everything Bundle — $49Example input
Question: Does a four-day week reduce burnout in knowledge-work teams? Sources: [S1] a national pilot report, [S2] an academic review, [S3] a company blog post about its own trial, [S4] a critical newspaper analysis.
Example output
Synthesis (312 words), confidence: Medium Organisations in the pilot reported lower self-reported burnout after six months [C2, C4], although participants volunteered, which likely inflates the effect [C9]. The academic review finds the effect is strongest when hours are actually cut rather than compressed into four longer days [C11, C12]... Gaps: no source measures burnout beyond 12 months; none covers client-facing support teams. Audit excerpt Sentence 3 | C6 | Partly supported | C6 reports productivity "held steady", not "increased". Corrected: "Self-reported productivity held steady [C6]."
Troubleshooting
- Quotes don't match the source exactly: ask the model to re-copy quotes character for character and spot-check two or three with your browser's find function.
- Sources are too long for the context window: run step 2 one source at a time and merge the tables, then do steps 3-4 on the merged table.
- The synthesis sneaks in background knowledge: re-run step 3 with 'If a sentence has no claim ID, delete it.'
Known limitations
- Models can still mis-quote or misattribute. The audit step reduces errors but does not replace reading the key sources yourself.
- The workflow is only as balanced as the sources you choose; it does not find missing perspectives.
- Very long PDFs may be truncated silently by some chat interfaces; check that the last pages were read.