Research Workflow

Build the mental model

Research is not "search the internet until something sounds right." It is a repeatable pipeline. Most bad research comes from skipping a step, usually evaluate and synthesize, jumping straight from search to output.

"AI" is a topic; you cannot research a topic. "How does Local RAG reduce cloud API dependency?" is a real question, specific enough to search for, narrow enough to answer, testable against evidence.

1. Define the question

Write one full-sentence question, not a topic word. Not "AI" but "How does Local RAG reduce dependency on cloud APIs?"

2. Search

Prioritize primary sources (the original paper, official docs, the dataset itself) and use secondary sources only for orientation.

3. Collect sources

Gather at least three to five sources that directly address the question. Never decide from a single source.

4. Evaluate

Check every source against five criteria: author, date, evidence, publisher, and references, plus known bias.

5. Take notes

Record notes in a fixed structure: claim, evidence, source, interpretation, not free-form paragraphs.

6. Synthesize and produce output

Compare notes, surface contradictions, form a conclusion, then write the article, report, or lesson.

Evaluation is the step most people skip entirely. A source that fails several checks might still contain something true, but it earns a lower weight in the final synthesis.

Primary Source: A source that comes directly from the original creator, researcher, or data: the paper itself, official documentation, or the raw dataset.

Secondary Source: A source that summarizes or interprets primary work, such as a blog post, news article, or review. Useful for orientation but weighted lower than a primary source.

Synthesis: The step of comparing multiple notes, connecting related claims, surfacing contradictions, and forming a conclusion — not copy-pasting several summaries together.

text

RESEARCH WORKFLOW PIPELINE
--------------------------
RESEARCH WORKFLOW PIPELINE
---------------------------
[question] -> [search] -> [collect sources] -> [evaluate]
                                                      |
                                                      v
    [output] <- [synthesize] <- [structured notes] <-+

Connect it to a real scenario

Before opening a single tab, write your real research question as one full sentence, not a topic word.

Search for the primary source first: the paper, spec, repository, or original announcement
Treat blog posts and social threads as leads toward the primary source, not the destination
Do not draw a conclusion until you have collected at least three to five sources

Check	What to look for
Author	Who wrote it, and do they have relevant credentials
Date	When it was written, and whether it is still current
Evidence	Whether claims are backed by data or citations
Publisher	Who published it, and their reputation
References	Whether it cites its own sources

Take notes in a fixed structure: claim, evidence, source, interpretation. This structure is what makes synthesis possible.

In synthesis, compare claims, flag contradictions explicitly, decide what you actually believe given the evidence, and only then write the output.

Research is a pipeline that starts from a real question, not a topic
Skipping evaluation means treating bad sources at the same trust level as good ones
Synthesis means comparing and concluding, not copy-pasting several summaries together

Try the working example

javascript

function scoreSource(source) {
  let score = 0;
  const reasons = [];

  if (source.hasAuthor) { score += 1; reasons.push("+1 has a named author"); }
  else { reasons.push("+0 no named author"); }

  if (source.hasDate) { score += 1; reasons.push("+1 has a publish/update date"); }
  else { reasons.push("+0 no date given"); }

  if (source.hasEvidence) { score += 1; reasons.push("+1 makes claims backed by evidence"); }
  else { reasons.push("+0 no evidence for claims"); }

  if (source.hasReferences) { score += 1; reasons.push("+1 cites its own sources"); }
  else { reasons.push("+0 cites nothing"); }

  if (source.knownBias) { score -= 2; reasons.push("-2 publisher is on the known-bias list"); }

  let verdict;
  if (score >= 3) verdict = "evaluate"; // use with confidence
  else if (score >= 1) verdict = "caution"; // usable but weight it lower
  else verdict = "reject"; // do not rely on this alone

  return { score, verdict, reasons };
}

const sources = [
  {
    name: "Official RAG whitepaper (vendor docs)",
    hasAuthor: true,
    hasDate: true,
    hasEvidence: true,
    hasReferences: true,
    knownBias: false,
  },
  {
    name: "Anonymous blog post, no date",
    hasAuthor: false,
    hasDate: false,
    hasEvidence: true,
    hasReferences: false,
    knownBias: false,
  },
  {
    name: "Sponsored content on a known-biased outlet",
    hasAuthor: true,
    hasDate: true,
    hasEvidence: false,
    hasReferences: false,
    knownBias: true,
  },
];

for (const s of sources) {
  const result = scoreSource(s);
  console.log(s.name, "->", result.verdict, "(score:", result.score + ")");
}

You should see

Official RAG whitepaper (vendor docs) -> evaluate (score: 4)
Anonymous blog post, no date -> caution (score: 1)
Sponsored content on a known-biased outlet -> reject (score: 0)

A score of 3 or higher means "evaluate" (usable with confidence), 1-2 means "caution" (usable but weighted lower), and 0 or below means "reject."

5-minute try-it

Collect three real sources for a research question of your own and run scoreSource() on them. For any source that lands on "caution", explain how you would weight it lower when writing your synthesis.

One important caution

Mistaking a topic for a question — "AI" cannot actually be researched

Skipping evaluation and treating every source at the same trust level

Wikipedia: Source evaluation — Productivity Systems