The Research Operating System · Research with AI
The Grounded
Research Loop
AI can produce an answer in seconds. It cannot decide whether the answer is true, what the evidence means, or whether you can defend it. That work is still yours.
A seven-stage, tool-agnostic workflow for doing that work systematically. Learn it once and it holds across any tool, any field, any future model.
Most AI research fails in one of two ways, and both are quiet enough to survive peer review until they don't.
The first is a fluent answer built on citations that do not exist. The model generates a plausible reference rather than retrieving a real one, and because the sentence reads well, nobody checks. The second is subtler and far more common: a review that is technically accurate and completely inert. Eighty sources, summarized one by one, with no argument holding them together. Author A found X. Author B found Y. That is a list. A list is not a synthesis, and a reviewer will know the difference within a paragraph. Neither failure is a prompting problem — both are workflow problems, and both are what the Loop exists to prevent.
The Grounded Research Loop
Two ideas that change how you research with AI
Grounded beats fluent. A claim that cannot be traced back to a source passage does not exist. The single most important habit in AI research is preferring tools that retrieve real sources over models that generate plausible-looking ones — and verifying every claim before you keep it. That is why the method is called grounded.
Synthesis beats summary. The value of a review is not a pile of summaries; it is an argument organized around ideas, disagreements, and gaps. The category's most common failure — serial summarizing — is what the Synthesize stage exists to prevent.
A list is not a synthesis
Three papers, one finding
Suppose three studies all report that AI assistance improves developer productivity.
Summarizing
Compressing what each source said, one at a time.
“All three studies found that AI assistance improved productivity.”
Accurate. Inert. It adds nothing a reader could not get from the abstracts.
Synthesizing
Comparing sources against each other — agreement, conflict, method, and the strength of the evidence.
“All three report gains, but they are not measuring the same thing. Two measure task completion speed; the third measures merged-code quality and finds a smaller effect. Only one followed teams beyond six months — and there the gain narrows. The consensus is thinner than it looks.”
This is an argument. It is also the sentence a reviewer remembers.
The difference is not effort. It is organization: synthesis is arranged around ideas, summary around authors. Every stage below is built to make the first one possible.
The Loop, stage by stage
Each stage names the failure it prevents and the artifact it produces. Stages four through six are the engine — that is where a pile of sources becomes an argument — so they get the most room.
01
Frame
Turn a topic into an answerable, scoped question. Whatever ambiguity you tolerate here is inherited by every stage that follows — and it compounds. Most bad reviews were lost before a single source was opened.
02
Gather
Find the candidate evidence base by retrieval, not generation. This is where fake citations are born: ask a model to “find me sources” and it will happily invent them — complete with plausible authors, a real-sounding journal, and a DOI that resolves to nothing.
Use tools that search real corpora. Keep the raw sources: you will need to trace back to them in stage six, and you cannot trace back to something that was never there.
03
Triage
Screen and rank. Decide what actually belongs in the review before you invest in reading it. Uploading everything you found is not thoroughness — it is noise, and noise dilutes every stage downstream. Apply explicit inclusion and exclusion criteria, and rank what survives by how much it can change your conclusion.
Output → Priority Map — sources tiered by inclusion criteria, so effort lands where it matters
04
Extract
Pull the same structured facts from every source: question, method, sample, finding, limitation. The temptation is to write loose prose summaries instead. Resist it — without consistent extraction you cannot compare studies against each other, and comparison is the whole point. This is the stage that makes synthesis mechanically possible rather than merely aspirational.
Output → Evidence Matrix — one row per source; columns for question, method, sample, finding, limitation
05
Synthesize
Organize the evidence around ideas, disagreements, and gaps — never around authors. Flip the axis: themes become rows, sources become columns, and each cell records whether a source agrees, contradicts, or is silent. The contradictions and the silences are the interesting part; they are where the unexplored questions live.
Good research is not collecting more information. It is deciding what survives.
This is the stage the entire Loop is built to reach — and the one most people skip.
Output → Insight Matrix + Research Depth Score — a 0–100 measure of how deeply you synthesized rather than summarized
06
Verify
Trace every claim back to a specific source passage, and label it: grounded, weak, or unsupported. Fluent output is persuasive precisely when it should not be — the sentences that read most smoothly are often the ones with nothing underneath them. If you cannot point to the passage, you cannot defend the claim, and you should not ship it.
Output → Confidence Check — every claim rated and traced
07
Cite
Produce traceable, correctly formatted citations. Never auto-generate them: they either do not exist, or they exist and do not say what you claimed. Then loop back — verifying and synthesizing routinely reveal that the question needs reframing or the evidence base needs widening. You are done when another pass changes nothing material.
The Loop on one real question
A scoping review on AI and developer productivity. One screen, start to finish — this is what the seven stages actually cost you.
- Frame “Does AI coding assistance improve developer productivity, and by what measure?” 1 question
- Gather Retrieved from real corpora — not generated. 120 papers
- Triage Priority Map: peer-reviewed, measured outcomes, since 2022. The other 85 were opinion, duplicates, or unmeasured. 35 papers
- Extract Evidence Matrix: one row per paper — method, sample, outcome measure, effect size, limitation. 35 rows
- Synthesize Insight Matrix: themes as rows. The papers agree on speed and split on quality — and only four followed teams past six months. 6 themes
- Verify Confidence Check: every claim traced to a passage. Two claims dropped — the sources didn't support them. 2 dropped
- Cite Every reference resolves to a real paper. 33 cited
The conclusion the Loop produced: AI assistance reliably improves speed; its effect on code quality is contested and thinly evidenced beyond six months. That sentence is defensible — and it is not one you could have written from 120 abstracts.
Go deeper
Research OS Prompts
Every stage of the Grounded Research Loop as a tested prompt — Frame through Cite, plus the Evidence Matrix, Insight Matrix and Confidence Check.
Explore the Research Prompt PackageOne-time · Lifetime access · No subscription
Five core artifacts
Each stage leaves something behind. Together they form the evidence chain that makes a review auditable — and the shared vocabulary the rest of this site is written in.
- Priority Map
- Candidate sources tiered against explicit inclusion and exclusion criteria, so reading time lands where it can change the answer. 120 papers → 35, once “peer-reviewed, measured outcome, since 2022” was applied.
- Evidence Matrix
- One row per source; columns for question, method, sample, finding, limitation. The structured base that makes cross-study comparison possible. Row: Peng 2023 · RCT · n=95 · 55% faster on one task · single task, no quality measure.
- Insight Matrix
- The axis flipped: themes as rows, sources as columns, each cell marked agree, contradict, or silent. This is where extraction becomes argument. Row “long-term effect”: 31 sources silent, 4 measured it. The silence was the finding.
- Research Depth Score
- A 0–100 measure of how deeply a review synthesizes rather than summarizes — across cross-source integration, handling of contradictions, gap identification, and grounding. A review that lists sources one by one scores under 30, no matter how well written.
- Confidence Check
- Every claim traced to a source passage and labelled grounded, weak, or unsupported. What you run before anything leaves your desk. “AI improves code quality” → unsupported. Two sources measured it; they disagreed.
Start here — before you open any tool
This is the Synthesize stage compressed into a single prompt. Paste it into NotebookLM, Claude, ChatGPT, or Perplexity — the method comes before the tool.
This is the Synthesize stage in one prompt. The full Research Prompt Pack includes a tested prompt for every stage of the Loop — Frame, Gather, Triage, Extract, Synthesize, Verify, Cite — plus the Confidence Check and Depth Score prompts.
Eight ways this goes wrong
Print this. Before a review leaves your desk, look for these — they are the failures the Loop is built to catch, and every one of them survives a fluent draft.
- Any citation you have not opened and confirmed exists. Fake
- A review organized by author instead of by theme. List
- Sources that agree — with no account of the ones that don't. No conflict
- A claim you cannot trace to a specific passage. Ungrounded
- Confident language on a thinly evidenced point. Overclaim
- Everything you found, uploaded — with no triage. Noise
- A question so broad that any source seems relevant. Unscoped
- A first draft you never looped back through. One pass
Questions researchers ask
How do you do research with AI without hallucinated citations?
Use retrieval-based tools that answer only from sources you supply, and verify every citation against the original passage before you keep it. The Verify stage exists for exactly this: a claim you cannot trace is a claim you cannot defend. Never let a model generate a reference list.
What is the best AI for a literature review?
There is no single best tool — different stages of the Loop reward different tools. Retrieval-grounded tools are strongest for Extract and Verify; broad models help with Frame and Synthesize. The method is what carries across all of them.
Can AI write my literature review for me?
It can accelerate every stage and it cannot own any of them. The argument, the interpretation, and the critical appraisal remain yours — AI output is a starting scaffold, not a finished review. It also cannot defend a claim in front of a reviewer. You can.
What is the difference between summarizing and synthesizing research?
Serial summarizing lists sources one by one with no overall argument. Synthesis organizes evidence around ideas, agreements, contradictions, and gaps — anchoring paragraphs to concepts rather than authors. Synthesis is what makes a review worth reading.
Is it acceptable to use AI for academic research?
Most institutions accept AI assistance when it is disclosed and the outputs are verified — similar to using statistical software. Follow your institution's policy, keep a human in the loop for judgment, and confirm every citation is real.
The page teaches the method. The Prompt Pack simply implements it.
The Research Prompt Pack
A tested prompt for every stage of the Loop — Frame, Gather, Triage, Extract, Synthesize, Verify, Cite — plus the Confidence Check and Depth Score prompts.
Get the Research Prompt Pack →
Already know the method? The full prompt library covers research, content, studio and multi-AI workflows.
Free Download
Get the Free Starter Kit
Ten complete AI research workflows — research, writing, study, slides and multi-AI. Free, delivered instantly, works on NotebookLM’s free tier.
Free · No credit card · Delivered instantly