# Review Prompt Templates URL: https://research-memex.org/docs/case-studies/systematic-reviews/review-prompt-templates Description: Prompt templates for the systematic-review course that mark what only the researcher can supply, and stop for review before the model goes further. *A template can carry your judgment to the model, or quietly hand it over.* These are the prompt templates for the systematic-review course. Session 2 uses the IMO prompt and the initial worksheet. Session 3 uses the refined worksheet. Session 4 uses the workflow design and the team brief. [Cognitive Blueprints](/docs/implementation/core-references/cognitive-blueprint-prompts) keeps the general examples. The convention is the one Cognitive Blueprints sets out under [active and passive prompting](/docs/implementation/core-references/cognitive-blueprint-prompts#active-and-passive-prompting). Square brackets such as `[topic]` are slots for your material. A bracket that begins with `insert:` marks something only you can supply: your question, your current account, your criteria, what a judgment word means in your field. Every template ends with the same guard line, so that a model given a half-filled template asks instead of guessing: `If anything in [insert: …] is still unfilled, stop and ask me before starting.` Each template also assumes the benchmark from [Session 1](/docs/case-studies/systematic-reviews/session-1-review-by-hand): the files in `notes/benchmark/` that you made by hand, on the same question and often the same papers. Run a template on that material first. Then every output has a hand-made twin, and the gap between them tells you something about the prompt as well as the model. ## IMO template for systematic reviews [#imo-template-for-systematic-reviews] The [IMO structure](/docs/implementation/core-references/cognitive-blueprint-prompts#understanding-the-imo-structure) moves from hypothesis to verification to refinement. In the original paper the model forms its own hypothesis. In a review, the hypothesis worth testing is yours: the account you hold after Session 1, written down before the prompt runs. Left to form its own, a model can settle on the most general statement the papers will bear, and then verify that. Three things make this version active. You supply the hypothesis. The verification quotes its evidence with a paper ID and a page, and lists disconfirming evidence first. And the prompt stops before refinement, so you judge what the verification found before the model revises anything. ```markdown You are helping me with a systematic review. The judgments in it are mine; your job is to test them against the texts. ## Context Review question: [insert: your question, from 01-question.md] Synthesis logic: [insert: aggregative, interpretive or explanatory, and the reason, from 02-protocol.md] Papers: [the set, with IDs] ## Hypothesis (mine) My current account of what these papers say: [insert: two or three sentences, written before this prompt runs] Do not replace it with your own. ## Verification Test my hypothesis against the texts. For every point, quote the passage and give the paper ID and page. 1. Evidence that contradicts it or that it cannot explain. By a disconfirming case I mean [insert: what counts in your field] 2. Evidence that supports it. By support I mean [insert: what counts, for example a finding, not an author's framing] 3. Papers my hypothesis says nothing about ## Stop Stop here. Show me the verification and wait. Do not refine the hypothesis until I reply. ## Refinement (after I reply) Using only the passages I have confirmed, propose changes to my hypothesis. For each change, name the passage that requires it. Test these alternative explanations: [insert: the rivals you want ruled in or out] ## Final synthesis (after I approve the refinement) Write it up as [insert: the format, for example a conceptual model or a summary of debates], with every claim tied to a locator. If anything in [insert: …] is still unfilled, stop and ask me before starting. ``` What you supply: the question, the synthesis logic, your hypothesis, what counts as support and as a disconfirming case, and the rival explanations. What you check: every quoted passage against the paper, before you reply to the stop. ## Initial prompt worksheet [#initial-prompt-worksheet] Fill this in before the first run. The benchmark line points at the file your output will be compared with; the account line is the version of your thinking the output will be held against. ```markdown # My synthesis prompt My benchmark: [insert: link to your Session 1 extraction table or memo] My question: [insert: your review question] My current account, written before the AI reads anything: [insert: two or three sentences] ## My chosen method [insert: the synthesis approach you are adapting, and why it fits your question and the traditions of your papers] ## My IMO-inspired prompt [Paste your prompt here. Check that it names the method, defines each judgment word, asks for locators, stops for your review, and ends with the guard line.] ## Notes on my approach - Why I chose this method: - How I adapted the IMO structure: - Judgment words I defined, and how: - What I expect this to reveal: ``` ## Refined prompt worksheet [#refined-prompt-worksheet] Fill this in after you have set the output beside your benchmark. The divergence line comes first because it is the evidence for every change below it. Where the output was generic, too smooth or oddly confident, run the [Output Mirror](/docs/implementation/roles/output-mirror) on the passage before deciding what to change: the answer may be in the prompt rather than the model. ```markdown # Refined prompt My benchmark: [insert: link to your Session 1 extraction table or memo] Where the AI output diverged from my benchmark: [insert] ## What the divergence says about my prompt - - ## My improved prompt [Updated version here, ending with the guard line] ## Changes made and why - - ``` ## Agentic workflow design [#agentic-workflow-design] A workflow design is the plan an agent team runs from. Three rules carry most of the weight. Each agent gets a defined input and a defined output file, so its work can be checked on its own. A human gate sits after extraction and after synthesis, the two points where an error spreads into everything downstream. And a contradictions log keeps disagreements between papers visible instead of letting a synthesis agent resolve them quietly. Before you add a role, the [Arrangement Auditor](/docs/implementation/roles/arrangement-auditor) asks whether a simpler arrangement would do; "add nothing" is a real answer. ```markdown # My agentic workflow design Question: [insert: your review question] Benchmark each output is checked against: [insert: the Session 1 files] Simpler workflow I am comparing this with: [insert: for example one prompt in one conversation] ## Agent 1: Extraction Input: the included full texts Output: extraction.md, one row per paper, fields: [insert: your extraction fields, from 06-extraction.md] Prompt: "Fill each field with the paper's own words and a page locator. Write 'not reported' when the paper is silent. Do not infer." ## Gate 1 (me) Check [insert: how many] extraction rows against the PDFs, every field and every locator, before anything continues. ## Agent 2: Appraisal Input: extraction.md as corrected at Gate 1, and the full texts Output: appraisal.md Prompt: "Apply [insert: CASP, MMAT or your checklist] item by item. Quote the passage behind each rating. Where the checklist does not fit the study design, say so rather than rating." ## Agent 3: Synthesis Input: extraction.md as corrected at Gate 1, appraisal.md, and the full texts, so that every quotation comes from a paper and not from a summary Output: synthesis.md and contradictions.md Prompt: "Follow [insert: your synthesis logic and its steps]. By a theme I mean [insert: your definition]. Log every disagreement between papers in contradictions.md, quoted with locators, and do not resolve it." ## Agent 4: Critique Input: synthesis.md, contradictions.md and the full texts, not the earlier summaries Output: critique.md Prompt: "For each claim in synthesis.md, check the cited passage. List claims the passage does not support, papers the synthesis never mentions, and contradictions it smoothed over." ## Gate 2 (me) Compare synthesis.md with my benchmark memo, read critique.md, and decide what stays. If anything in [insert: …] is still unfilled, stop and ask me before starting. ``` ## Team brief [#team-brief] A workflow design can also be handed to a single agent as one prompt. In the module's second run, a prompt that described the review's phases was enough for a Cherry Studio agent to create a sub-agent for each phase and run them; [Session 4](/docs/case-studies/systematic-reviews/session-4-agentic-workflows#one-prompt-a-team) walks through that run. The brief below can be pasted into a Cherry Studio agent or into Claude Code started in the review folder. It stops after extraction, which is the first point where you can check the team's work against the papers. Point it at the same records and papers as your benchmark: the 50 abstracts from `04-screening.md` and the 5 full texts from `06-extraction.md`. ```markdown # Team brief: [review title] ## The review Question: [insert: your review question, from 01-question.md] Protocol: [insert: path to your protocol]. Read it before anything else. Synthesis logic: [insert: aggregative, interpretive or explanatory] Records: [path to the abstracts file] Papers: [path to the folder of full texts, one file per paper] Write only inside [output folder]. Never edit the papers or anything in notes/benchmark/. ## The team Before creating anything, list the sub-agents you will create, what each receives, which file each writes, and where you will stop. Wait for my approval. Then create one sub-agent per phase. Give each only the inputs listed, and have it write its output to the named file. 1. Screening. Input: the records and the criteria in the protocol. Output: screening.md, one row per record: ID, decision, the criterion that decided it, confidence. Mark a record UNCERTAIN when [insert: your rule for borderline cases]. 2. Extraction. Input: the full texts of the included papers. Output: extraction.md, one row per paper, fields: [insert: your extraction fields], each entry with a page locator. Write "not reported" rather than infer. ## Gate Stop after extraction. Show me screening.md and extraction.md, list the sub-agents you created and the inputs each received, and wait. Do not start the next phase until I reply "continue". ## After the gate 3. Synthesis. Input: extraction.md as I corrected it, and the full texts, so that every quotation comes from a paper. Output: synthesis.md, following the synthesis logic above. By a theme I mean [insert: what a theme must have to count]. Keep every disagreement between papers in contradictions.md, quoted with locators; do not resolve it. 4. Critique. Input: synthesis.md, contradictions.md and the full texts, not a summary of them. Output: critique.md, listing each claim its passage does not support, and each quotation in contradictions.md that its locator does not bear out. Stop again after critique and wait for me. ## Reporting At each stop, say what you did that this brief did not ask for. If anything in [insert: …] is still unfilled, stop and ask me before starting. ``` What you supply: the question, the protocol, the borderline rule, the extraction fields and what counts as a theme. What you check at the gate: two extraction rows against the PDFs, field by field. What you check at the end: the synthesis against `08-synthesis-memo.md`. ## Checking what comes back [#checking-what-comes-back] The templates reduce what the model has to guess; they do not check its work. After each run, three checks are worth the few minutes they take. * Benchmark: set the output beside the Session 1 file it corresponds to, and write down the first place they diverge. * Output Mirror: run the [Output Mirror](/docs/implementation/roles/output-mirror) on one paragraph that reads as generic or too smooth, and name the [Failure Museum](/docs/implementation/core-references/failure-museum) exhibit it resembles, if any. * Arrangement Auditor: before adding an agent, a second model or a scheduled run, ask the [Arrangement Auditor](/docs/implementation/roles/arrangement-auditor) whether the arrangement earns it. *** Navigation: [Session 2](/docs/case-studies/systematic-reviews/session-2-ai-powered-practice) • [Session 3](/docs/case-studies/systematic-reviews/session-3-human-vs-ai-synthesis) • [Session 4](/docs/case-studies/systematic-reviews/session-4-agentic-workflows) • [Return to Case Study Overview](/docs/case-studies/systematic-reviews/systematic-review-syllabus)