A user said, "I stopped using it because I couldn't tell what it just did." Your AI summary recorded, "users had concerns about feedback clarity." Both describe the same moment. Only one of them will make a stakeholder lean forward. The first is a person quitting your product mid-task in frustration. The second is a bullet point that dies in a slide. This lesson is about the quote-pull problem, the specific, predictable way AI summaries sand the most important phrase out of your research, and the verbatim-pull pattern that gets it back. You will leave with a re-usable Claude prompt that forces direct quotation and a research-quote library with provenance, so the phrase that changes minds survives the trip from interview to readout.

The Phrase That Died in the Summary

Summarization is compression, and compression is loss. That is fine for most content; you do not need the exact wording of a logistics update. But user research is the one place where the exact wording is the value, because the specific phrase a user chose carries the emotion, the severity, and the texture that a paraphrase erases. "I couldn't tell what it just did" tells you the failure was about feedback and visibility of system status, that it happened mid-action, and that it was bad enough to make someone quit. "Concerns about feedback clarity" tells you none of that. It tells you a category. Categories do not move roadmaps; quitting users do.

The quote-pull problem is that AI summarizers are trained to produce smooth, generalized prose, so they reflexively convert vivid, specific utterances into bland, abstract categories. The model is not malfunctioning. It is doing exactly what summarization rewards: shorter, cleaner, more general. The trouble is that "shorter, cleaner, more general" is the precise opposite of what a research quote needs to be. The verbatim is the asset, and the default summary throws it away while looking helpful.

In user research, the paraphrase is the bug and the verbatim is the feature. A summary that loses the user's exact words has lost the only thing the research was for.

Why the Best Phrase Is the First to Go

There is a cruel asymmetry in how summarizers compress. The phrases most likely to be paraphrased away are the vivid, idiosyncratic, emotionally loaded ones, because those are the phrases that look least like the generic prose the model is pulled toward. "It felt like the app was gaslighting me" gets smoothed to "user reported a frustrating experience" precisely because the original is too specific and too colorful to survive the averaging. The blandest content survives summarization intact; the sharpest content is exactly what gets sanded off. So the model does not lose random phrases. It systematically loses the best ones, the quotable ones, the ones you would have built your readout around.

The Severity Signal Vanishes Too

Beyond the words, summaries flatten severity. A user who says "this is mildly annoying" and a user who says "I will cancel over this" both become "users expressed dissatisfaction" in a generic summary. The intensity gradient, which is one of the most important things research tells you, collapses into a flat category. When you later prioritize, you cannot tell the cancel-threat from the mild gripe, because the summary erased the difference. Recovering the verbatim recovers the severity, because the user's own words carry the heat the category dropped.

The Verbatim-Pull Pattern: Force the Quote

The fix is to stop asking the model to summarize and start asking it to extract. A summary prompt invites paraphrase. An extraction prompt forbids it. The difference is in the instructions, and the instructions are the whole technique.

Here is the re-usable Claude prompt structure. "From this transcript, do not summarize. Extract direct quotes only. For each notable moment, output the participant's exact words inside quotation marks, copied character for character from the transcript, with nothing added, nothing smoothed, and nothing generalized. Include the timestamp or line reference. Do not paraphrase. Do not combine multiple statements into one. Do not replace the participant's words with category labels. If a statement is vivid, emotional, or specific, that is exactly the kind you must capture verbatim. Then, separately and clearly labeled, you may add your own one-line interpretation, but the quote must always appear first and unaltered."

That prompt does three things. It forbids the default summarization behavior explicitly. It tells the model that vividness is a signal to capture, not smooth. And it separates the verbatim from the interpretation so you can always see the user's actual words before the model's gloss. The separation is critical: when interpretation and quote are merged, you cannot tell where the user stopped and the model started.

The Verbatim Test: Could the User Sue You for It?

Here is a sharp test for whether you have a real verbatim or a smoothed one. Ask: if the participant read this quote attributed to them, would they say "yes, those are my exact words," or would they say "that's not quite what I said"? A true verbatim passes the first test. A paraphrase the model dressed up in quotation marks fails it. Apply this test to every quote that will appear in a readout, because a quotation mark around a paraphrase is worse than no quote at all; it claims authenticity it does not have, and a sharp stakeholder will catch it.

The Research-Quote Library With Provenance

The artifact this lesson produces is a research-quote library, a durable, reusable store of verbatim quotes that any designer or PM on your team can pull from for a readout, a persona, a JTBD statement, or a pitch. Build it as a table, in Notion or a shared sheet, with one row per quote and these columns: the verbatim quote in quotation marks, the participant identifier, the source (which interview or survey) with a timestamp or line reference, the theme it supports, the severity (the user's own intensity, captured from their words), and a verified flag.

Provenance is the point. Each quote carries enough metadata that anyone can trace it back to the exact moment a real human said it. This turns the library into an evidence base rather than a collection of floating sentences. When a designer six weeks later needs to justify a design decision, they pull a verified verbatim with full provenance and the decision is grounded in a real user voice, not in a half-remembered paraphrase. The library compounds: every study adds rows, and the team's pile of defensible, traceable user evidence grows.

Severity as a First-Class Column

Make severity a column you fill from the user's own words, not a number you guess. "I will cancel over this" is high severity, in the user's voice. "It would be nice if" is low. Capturing severity from the verbatim, rather than re-introducing the flattening the summary caused, means the library preserves the intensity gradient that prioritization depends on. When you later sort the library by theme, you can see at a glance which themes carry cancel-threats and which carry nice-to-haves, because you kept the heat in the user's words.

The Three Quote-Pull Failures to Catch

Across many AI research summaries, three quote failures recur. Catch them by name.

Failure One: The Category Swap

The model replaces the user's words with a category label, as in "feedback clarity" for "I couldn't tell what it just did." Catch it with the verbatim test: if the user would not recognize the words as theirs, you have a category, not a quote. Demand the original.

Failure Two: The Frankenquote

The model stitches phrases from two separate moments into one tidy quotation that the user never actually said in that form. It reads beautifully and is fabricated. Catch it by confirming every quoted sentence appears as a contiguous utterance in the transcript, not assembled from parts. A frankenquote is a fabrication wearing quotation marks.

Failure Three: The Severity Flatten

The model strips the intensity, turning "I will cancel" into "expressed dissatisfaction." Catch it by checking whether the captured quote still carries the heat of the original. If the user's words conveyed a threat, a quitting, or a strong emotion, and the captured version reads neutral, the severity was flattened and you re-pull the verbatim that holds it.

When a Summary Is Fine, and When It Is Not

This lesson is not anti-summary. Summaries are useful for orientation: a quick paragraph telling you what an interview covered helps you decide where to dig. The rule is simple. Use the summary to navigate; never use it as the evidence. The moment a phrase will appear in a readout, justify a design decision, anchor a persona, or shape a roadmap, it must be a verified verbatim from the quote library, not a line lifted from the summary. Summaries are the map; verbatims are the territory. You read the map to find your way, but you build on the territory.

This is the same L2 discipline in a new shape. The model gives you the fast orientation; you supply the verified evidence. The quote library is where your verification lives, and the verbatim-pull prompt is how you stop the model from sanding off the phrase that matters before you ever see it.

Putting It to Work This Week

Save the verbatim-pull prompt somewhere you will reuse it, and run it on your next transcript instead of asking for a summary. Build the research-quote library with provenance and severity columns, and add every verified verbatim you pull. Apply the verbatim test to every quote before it enters a readout. When you find a category swap, a frankenquote, or a severity flatten, re-pull the original.

You will know it is working the first time a stakeholder leans forward at a quote, the way nobody ever leaned forward at "concerns about feedback clarity." That lean is the phrase doing its job. The model will always pull toward the bland category, because that is what summarization rewards. Your job is to pull back toward the user's exact words, because that is where the research lives, and to keep those words in a library with provenance so the phrase that changes minds is always one verified row away.

Why the Verbatim Is a Political Instrument, Not Just a Data Point

There is a dimension of the quote-pull problem that has nothing to do with accuracy and everything to do with power, and senior designers learn it the hard way. A verbatim quote is the most persuasive instrument a researcher carries into a room full of competing priorities. When you say "users had concerns about feedback clarity," you are offering an opinion that any stakeholder can counter with their own opinion, and in a roadmap meeting opinions get traded until the loudest one wins. When you say "a user told us, in their own words, I stopped using it because I couldn't tell what it just did," you are no longer offering an opinion. You are putting a real human being in the room, and it is much harder to argue with a person who quit your product than with a researcher's category. The verbatim is how design wins arguments it would otherwise lose on volume and seniority.

This is why the summary's flattening is not a neutral loss of fidelity; it is a quiet disarmament of the research function. Every time a vivid quote gets sanded into a category, the readout gets a little easier to dismiss, a little easier to override with "well, I think users actually want." The designers who consistently lose roadmap arguments are very often the ones presenting AI-summarized findings, because the summary stripped out exactly the ammunition that would have made their case unarguable. Keeping the verbatim is therefore not pedantry about wording. It is preserving the single most effective rhetorical tool design has, the unmediated voice of the user, against a summarization process that is structurally biased toward throwing it away. When you fight to keep the exact phrase, you are fighting to keep design's seat at the table.

The Quote Wall as a Team Ritual

One durable way to make the verbatim library matter is to give it a physical or visible home: a quote wall, a FigJam board, or a Slack channel where the rawest, sharpest verified user quotes live where the whole team passes them. The reason this works is that a quote in a research repository is evidence, but a quote on the wall is culture. When an engineer reads "I couldn't tell what it just did" every morning, the team starts designing against that sentence without anyone having to cite a study. The AI verbatim-pull pattern feeds this wall efficiently, because it surfaces the quotable lines from hours of transcript that you would otherwise never have time to find. The tool does the volume; you do the curation; the wall does the persuasion. That division of labor is the whole L2 thesis in one ritual.

Building the Verbatim-Pull Into Your Standing Workflow

A prompt you have to remember to use is a prompt you will eventually forget under deadline, so the real win is making verbatim extraction the default path rather than a discipline you summon. The practical move is to save the verbatim-pull prompt as a reusable template wherever you do synthesis: a saved Claude project with the instructions baked into its custom instructions, a Notion AI prompt block, or a snippet your whole team shares. The goal is that running synthesis and forbidding summarization become the same action, so that nobody on the team ever again types "summarize these interviews" out of muscle memory. The failure mode is not that designers do not know better; it is that the lazy prompt is one word shorter and Friday is one deadline closer.

Pair the saved prompt with a one-line review gate that anyone can apply: no quote enters a deliverable unless it has a participant identifier and a source reference attached. That single rule does enormous work, because a fabricated frankenquote or a smoothed category cannot easily satisfy it; the moment you go looking for the exact line and timestamp to attach, the fabrication reveals itself. The rule is cheap to enforce and expensive to violate, which is exactly the property you want in a verification gate. When the verbatim-pull prompt is the default and the provenance gate is the habit, you get the speed of AI synthesis with almost none of the quote-pull risk, and you stop relying on any individual designer remembering to be careful on the worst day of the sprint. The system carries the discipline so the human does not have to carry all of it alone.

Key Takeaways

• The quote-pull problem is that AI summarizers convert vivid, specific utterances into bland, abstract categories, because summarization rewards shorter, cleaner, more general prose, which is the opposite of what a research quote needs to be.
• The best phrase is the first to go: the vivid, emotional, idiosyncratic utterances least like generic prose are exactly the ones smoothed away, so the model systematically loses your most quotable, mind-changing lines.
• Summaries also flatten severity, collapsing "I will cancel" and "mildly annoying" into "expressed dissatisfaction" and erasing the intensity gradient prioritization depends on.
• Use the verbatim-pull pattern: a Claude prompt that forbids summarizing, forces character-for-character quotation, treats vividness as a signal to capture, and separates the verbatim from any interpretation so you can always see the user's actual words first.
• The three quote-pull failures to name: the category swap (user words replaced by a label), the frankenquote (phrases from separate moments stitched into one fabricated quotation), and the severity flatten (intensity stripped to a neutral category).
• Ship a research-quote library with provenance: verbatim, participant, source with timestamp, theme, severity in the user's own words, and a verified flag. Use summaries to navigate, never as evidence; the moment a phrase enters a readout it must be a verified verbatim.