What makes a good AI prompt for evidence synthesis β and why a sequence of five focused prompts will always outperform a single elaborate one.
The single-prompt trap
Most people approach AI tools with what feels like the most efficient strategy: one comprehensive, carefully worded prompt containing everything they need. The result is usually a long, generic, over-hedged response that requires another hour of manual filtering to be useful. The intuition that more detail upfront produces better output is wrong. It produces longer output β which is not the same thing.
The reason is structural. A single prompt asks the AI to simultaneously orient to a topic, retrieve relevant evidence, filter for quality, apply geographic and equity constraints, and format a usable synthesis. That is five distinct cognitive operations compressed into one instruction. Each one degrades the others. The AI satisfies the surface form of the request without doing any of them well.
The fix is sequential prompting: a chain of focused exchanges where each prompt builds on what came before, each response narrows the search space, and the conversation converges on the answer you actually need. This is not a workaround β it is how good analysis works, whether the analyst is human or machine.
Why this is the same loop that runs everything
There is a deeper principle here worth naming. Every iterative improvement system β from gradient descent in machine learning to the way a skilled researcher refines a literature search over multiple sessions β follows the same closed loop:
ΞΈ
Your current understanding
what you know before the prompt
β
L(ΞΈ)
The gap
what the AI response reveals you're missing
β
βL
The refinement signal
which direction to push the next prompt
β
ΞΈβ²
Updated understanding
the new state that becomes the next prompt's starting point
βΊ each response updates the prior β the conversation converges
Each prompt you send is a parameter update. Each response is a loss signal telling you where your understanding is incomplete. The next prompt corrects in the direction of that signal. A single-shot prompt skips this loop entirely β it asks for the answer before you know what you don't know. Sequential prompting is the loop, run deliberately.
The biological constraint is real: your prefrontal cortex can hold roughly four to seven active threads in working memory at once. A complex UHC literature review β juggling populations, interventions, equity dimensions, geographic filters, and methodological quality simultaneously β exceeds that limit quickly. You feel it as cognitive fatigue. Sequential prompting externalises that working memory into the conversation thread, letting the AI hold the prior while you focus on the next refinement signal.
The anatomy of a well-structured prompt
Before building a sequence, each individual prompt needs these five elements. Missing any one of them degrades the response quality predictably.
Role
Tell the AI what kind of expert it is
AI tools perform significantly better when given an explicit role that defines the register, expertise level, and frame of reference for the response. A generic prompt gets a generic answer.
"You are a health economics researcher supporting WHO India, specialising in health financing and UHC in LMIC contexts."
Context
State what you already know
Give the AI your current understanding β the findings you have already, the gaps you are trying to fill, the decision the brief will inform. This is the ΞΈ in the loop: your prior state. Without it the AI starts from zero on every prompt.
"I have already reviewed 12 studies on PM-JAY inpatient coverage. I need evidence specifically on outpatient financial protection, which those studies did not cover."
Task
One specific job, precisely stated
The most common prompting error is asking for too many things at once. One task per prompt. "Find evidence, summarise it, assess quality, flag gaps, and format it for a brief" is five prompts compressed into one β and it produces five mediocre outputs instead of one good one.
"Find and summarise evidence on community health insurance and outpatient OOP in South Asia, published after 2015."
Constraints
Geography, methodology, equity, date
Explicit constraints are where Session 1 (bias) and Session 3 (PECO-F) do their work at the prompt level. Without them the AI defaults to its WEIRD-skewed training distribution. Every health financing prompt for WHO India work should name the geographic and equity constraints explicitly.
"Prioritise India and South Asia. Include only studies reporting financial risk protection outcomes disaggregated by income quintile or gender. Exclude high-income country evidence."
Format
Tell it how to structure the output
An AI given no format instruction will choose one for you β usually a long flowing paragraph that buries the findings you need. Specify the structure: numbered list, comparison table, one sentence per finding, or a specific template. The brief anatomy from Session 5 is a valid format instruction.
"Return a numbered list. Each item: intervention name, population, outcome measure, effect size if reported, study quality in one word."
A five-prompt sequence for a UHC comparison
This is a worked example of sequential prompting for a health financing question: comparing tax-funded versus contributory insurance schemes for informal sector financial protection in India. Each prompt builds on the previous response β the conversation is the protocol.
01
Orient β establish the landscape
"You are a health economics expert focused on LMIC health financing. What are the main mechanisms by which tax-funded and contributory health insurance schemes differ in their financial protection outcomes for informal sector workers? Give me a conceptual framework β no citations yet, just the analytical structure I should use to compare them."
This prompt does not ask for evidence. It asks for the framework that will make the evidence search coherent. The response tells you what dimensions to look for in studies β and reveals gaps in your own framing before you search.
02
Retrieve β targeted evidence search
"Using the framework above, find systematic reviews and economic evaluations published after 2015 comparing tax-funded versus contributory health insurance on catastrophic expenditure and utilisation among informal sector households. Restrict to South Asia and sub-Saharan Africa. Return a numbered list: scheme type, country, population, outcome, effect direction, study quality."
Now that the framework is established, the retrieval is precise. The structured output format forces the AI to commit to specifics β and makes gaps immediately visible as empty cells.
03
Probe the equity dimension
"From the studies above, which ones report outcomes disaggregated by income quintile, gender, or caste? For those that do, what do the disaggregated findings show β does financial protection accrue equally across subgroups, or does it concentrate in upper BPL quintiles?"
This prompt uses the previous response as its input β it only makes sense in sequence. It targets the F dimension of PECO-F that generic searches miss, and forces the AI to distinguish aggregate from disaggregated findings.
04
Surface the India-specific gap
"Based on what you have found, what is the most significant gap in the direct India evidence? Specifically: is there evidence on PM-JAY's financial protection impact for informal sector workers that is comparable in quality to the LMIC studies above? If not, what is the best available proxy evidence?"
This prompt converts the evidence gap from an embarrassment into an explicit, documented finding. The answer becomes the 'Evidence Gaps' section of the policy brief β drafted by the AI from its own retrieval limitations.
05
Synthesise for the brief
"Now synthesise the above into three policy-ready findings and one evidence gap statement, formatted for a WHO India evidence brief directed at MoHFW. Plain language, no citations in the findings text, no hedging. Each finding: one sentence stating direction of effect, population, setting, and quality caveat."
The final prompt does not repeat the retrieval work β it transforms what has already been established into the specific output format needed. This is the Session 5 brief generator, run manually as the closing step of the sequence.
Practice the loop
π Live prompt lab β sequential conversation
This is a persistent conversation β each message you send carries the full history forward. Use it to practice the five-prompt sequence above, or run your own health financing question through an iterative chain. Watch how each response changes the quality of your next prompt.
Start with an orient prompt (no evidence yet β just framework). Then retrieve. Then probe equity. Then surface gaps. Then synthesise.
System
You are in a persistent health economics research session. I am a WHO India health financing specialist. Each message carries our full conversation forward β use what we have established in every response. I will build up to a policy brief through a sequence of focused prompts. Begin when ready.
Conversation history is held in this session only β it resets when you reload the page. For a real research chain, copy your prompt sequence and responses into a document as you go.
π― Key takeaway
A single elaborate prompt is a single parameter update on a cold model. A five-prompt sequence is an iterative convergence β each exchange narrows the search space and updates the prior. The loop is: orient β retrieve β probe equity β surface gaps β synthesise. Your prefrontal cortex has a working memory limit; the conversation thread does not. Use the thread as external memory and let each response tell you where to push next. Session 3 applies this to the hardest prompting task in health economics: critical appraisal of cost-effectiveness studies.