Your Feature Backlog Is Buried in the Research Literature
Ask where a product’s next feature should come from, and the standard answers are three: trending repositories, the arXiv feed, and the competitor’s changelog. All three are useful. None of them answers the question a backlog actually asks, which is what the person using the product is missing.
There is a fourth source, and software largely does not read it: the experimental literature on how people generate ideas, lose them, set them down, and come back to them. It is a century of controlled experiments. Most products that help someone think, write, plan, or remember are hypotheses about cognition, and most of those hypotheses have already been tested.
Ideas come from a funnel, not a feed
Feature ideas have four sources, and they are not interchangeable tabs on the same page.
| Tier | Source | What it gives you | Yield when you have no users |
|---|---|---|---|
| 1 | Your users’ friction: support threads, drop-off logs, session abandonment | Features people will actually use | Empty. You have no users yet. |
| 2 | Adjacent products’ gaps: competitor changelogs, one-star reviews, “build me an X” threads | Features people already pay to solve badly | Spent. The competitive analysis mined it. |
| 3 | The science of thinking: cognition and creativity journals | Features no competitor can claim | High. And uncrowded. |
| 4 | The capability frontier: arXiv, GitHub trends, model releases | What just became possible | Low for what to build. Right for how. |
Tier 4 is where most teams point the firehose, and it answers the wrong question. Trending repositories tell you what developers signal approval of. arXiv tells you what a technique can do a few years before it is productizable. Both answer what is possible. A backlog is a list of what is missing, and abandoned roadmaps sit in the gap between them. Keep tier 4 for feasibility checks. Do not let it set the backlog.
The order flips the day you have users. Tier 1 is the best source in the world, and it is permanently first once it exists. The science lane is where the backlog comes from before product-market fit; after that it becomes an accelerator. It is not a substitute for listening to users. It is the thing that keeps you building while the room to listen to is still empty.
Why the literature lane is uncrowded
Every competitor in a category reads the same three feeds. Reading a fourth source nobody else reads is an arbitrage: the ideas cost nothing to read, they arrive screened by peer review, and they are invisible to every roadmap review that sources from the status quo. Almost nothing in software is built from evidence about the actual person at the keyboard.
The audit that produced this post ran against an idea-capture tool, and it kept turning up the same pattern: the research contradicts the default the industry ships.
Example 1: the open brainstorm is a worse version of solo work
The controlled comparison has been run since 1958: the same number of people, the same problem, the same time, half generating together in a group and half generating alone with their output pooled afterward. The pooled solo work wins, on quantity and on quality (Mullen, Johnson & Salas, 1991). The loss is not in dispute. The cause is argued, but the mechanism named most often is production blocking: one person holds the floor while everyone else waits, and the act of waiting suppresses the ideas being held, alongside evaluation apprehension and free riding (Diehl & Stroebe, 1987). Follow-up experiments showed the damage is procedural, not personal, and can be reduced by changing the procedure (Diehl & Stroebe, 1991).
The design consequence: generation and evaluation have to be two passes. A tool that lets you rate, tag, and triage while you are still producing ideas is asking you to evaluate and generate at the same time, which is the exact setup the literature says loses. And when collaboration matters, let everyone generate silently first, then pool: the written round sidesteps the blocking the spoken round is built from. The product I ran this audit against ships the wrong default, a live canvas where half-formed ideas are visible to everyone and the rating buttons sit on the same screen as the input box. Separating the two modes is a UI change, not an engineering project.
Example 2: stepping away is a scheduled event, not a mood
Setting a problem aside genuinely raises solution rates. The meta-analytic review found a positive incubation effect overall, stronger for divergent tasks, the kind with many acceptable answers, than for insight puzzles; larger when the work beforehand was longer; and weaker when the break was filled with demanding work (Sio & Ormerod, 2009). The practical reading: prepare hard, leave it, come back, and do not spend the break on something harder.
Software treats the opposite as true. “Unfinished” is a state to badge, queue, and nag about, and the archive is where ideas go to die. The design consequence: parking should be a first-class event. A session closes with a return date, the tool re-surfaces it with a prompt, and the break counts as part of the work instead of a lapse of commitment. Most products never ship this, because their model of an unfinished task is a red dot on a list, not a scheduled return.
Example 3: the famous effect about unfinished work is the wrong one
Every retention deck in the industry invokes the Zeigarnik effect: interrupted tasks are remembered better, so unfinished work is your hook. The name comes from Bluma Zeigarnik’s 1927 study, product design has been quoting it ever since, and the replication record does not support it. A 2025 meta-analysis covering 59 publications found no such memory advantage: excluding Zeigarnik’s own data, the recall ratio of interrupted to completed tasks was 0.99 across 37 studies (including the original, the same), with a small pooled effect size (dz = 0.15), and interrupted tasks made up roughly 49% of recalled tasks in the studies that measured it (Ghibellini & Meier, 2025). The memory theory behind all those reminder surfaces does not replicate.
What does hold is the effect that gets less press. Maria Ovsiankina’s 1928 study found people pick interrupted tasks back up, and the same meta-analysis confirms it: when given the chance, people resume the unfinished task about 67% of the time, against a 50% chance baseline (same analysis). The pull is toward finishing, not remembering.
The design consequence cuts against every “you have unfinished items” surface: build resumption, not reminders. A reminder says “you did not finish X,” which is a nag built on a memory model that does not exist. Resumption gives back the state: “you were two ideas into the pricing session, here is exactly where you left off.” The measured tendency is toward completion, so deliver the thing people are already pulling toward, instead of a shame notification about the thing they forgot.
The filter is the work
A literature scan over-produces, because every finding reads like a feature. The scan is the easy half. The filter is what separates a backlog from a list of papers.
A candidate has to close an outcome the product already promises, or a gap users have already named. Anything else is noise, regardless of how interesting the study is. And the literature tells you what to build, not whether it will move your metric: it gives you the mechanic, and the prototype decides the numbers.
This is the same argument as the last post
The previous post argued that a general model is fluent everywhere and reliable nowhere in particular, and that domain knowledge is the conversion layer between the two. The product analogy is exact. Trending repos and preprint feeds are the general model: they show you the surface of what a field looks like. The cognition literature is the domain knowledge: the part that tells you what the field actually checks. A backlog drawn from the first source without the second inherits all of the fluency and none of the discipline.
The audit this post came from ran the same way that argument does: pick the one uncrowded source that knows the user’s mind, then filter everything through outcomes the product is already on the hook for.
When to run it
Instrument the funnel first, so tier 1 starts talking the moment users arrive. Then make the science lane a monthly ritual: one pass through the four columns, three to five candidate mechanics from the literature, filtered by the outcome rule, and build the smallest one. The literature will still be on the shelf a year from now, and so will the users. Only the ordering changes.
The limit matters too. The literature is a source, not a backlog generator: the filter is the work, and the shipping part still asks for everything the product discipline normally demands. The ideas are published, peer-reviewed, and ninety years old in places, waiting on a shelf nobody else in the category reads.
Stellabyte LLC is an AI, RF systems, and data engineering firm based in California. If your product could use a feature roadmap sourced from evidence instead of feeds, get in touch.