Picture a data scientist whose desk drawer is labelled “the museum of other people’s good ideas.” It holds printed diagrams, a crumpled page of someone’s tidy evaluation table, a screenshot of a particularly honest model write-up. None of it is ever copied. The drawer simply gets opened whenever a project stalls, the way a carpenter might glance at a joint someone else had cut well. That drawer is, in essence, a swipe file, and across artificial intelligence work the habit is becoming steadier and more deliberate. What follows is a short survey of how common the practice has become and which way it is moving.
What a swipe file means in AI practice
In its simplest form, a swipe file is a personal collection of strong examples gathered from other people’s work, kept in one place so you can draw on them when you begin something of your own. In the world of AI and AIOps, those examples are rarely slogans. They are shapes: how a clear model card is laid out, how a team explains a failure case without defensiveness, how an architecture diagram separates data flow from decision flow, how a post-incident note describes an automation that misfired.
The collection is not a library of finished products to reuse. It is a reference shelf for craft. Your own work still has to stand on its own facts, data and constraints.
Why the habit is spreading
Three pressures push practitioners toward it. First, the field moves quickly, and nobody can read everything. A curated shelf of examples is a way of keeping standards high without keeping up with every release. Second, AI work is explained to mixed audiences: engineers, managers, customers. Explaining well is a learned skill, and learning it from a handful of excellent examples is faster than learning it from scratch. Third, many teams are small. Without a senior editor on hand, a swipe file acts as a quiet mentor that does not need a calendar slot.
Nothing here is novel. Designers, copywriters and filmmakers kept reference boards long before software teams did, and the instinct is identical.
A small story from an operations team
Consider a hypothetical operations group running anomaly detection across a fleet of services. Each time the group wrote up a false-alarm review, the document came out differently: sometimes terse, sometimes rambling, often missing the one fact a reader needed. Someone suggested a shared folder where members dropped examples of reviews they had found genuinely clear, from anywhere, with a one-line note on what made each one work.
Within a couple of months the group’s own write-ups began to share a recognisable backbone: what was observed, what the system concluded, what a person decided, what changed afterwards. Nobody had imposed a template. The pattern emerged because everyone had been looking at the same handful of good examples. This is an invented scenario meant only to show the mechanism, but teams that try the habit often report something similar.
What practitioners tend to collect
When people describe their files, certain categories come up repeatedly.
- Explanations for non-specialists. Passages that make a technical idea land with a reader outside the field, such as a plain account of how a classifier reaches a label.
- Honest limitation statements. Examples where a team says plainly what a system should not be used for.
- Layout and structure. Report outlines, dashboards of a type they admire, ways of ordering a long document.
- Naming and vocabulary. Terms that a team coined and that stayed clear over time.
- Decision write-ups. Short notes that record why one approach was chosen over another.
It is telling that very little of this is code. The shelf mostly holds communication and judgment, since those are what tend to be hardest to teach and easiest to forget.
Where examples come from
The sources are varied. Internal documents that colleagues were proud of, public engineering write-ups, conference slides, and plain-spoken blog posts all qualify. Your own site can contribute too: an article like the one on AI helping managers make sounder calls demonstrates one way to frame benefits for a managerial reader, while a simple walkthrough of how image classification works shows a way of explaining a technical mechanism with no jargon. Saving pointers to such pieces, with a note on the specific technique, is more helpful than saving the pieces themselves.
Increasingly, people also look sideways at fields with their own explanatory traditions. One case is a website styled like an aviation and space academy, which arranges its material into course sections, licence sections and a blog; an AI team might glance at how a staged curriculum is laid out and borrow the idea of sequencing for its own onboarding guide.
How the habit shows up in AIOps work
Operations teams that lean on automation have their own reasons to keep examples. Runbooks, alert descriptions and incident summaries are all pieces of writing that someone has to read at an awkward hour. A team that has collected a few unusually readable incident summaries tends to write its own in a similar spirit: the headline states the effect on users, the body separates what the tooling suggested from what a person did, and the ending says what will be different next time. Nothing about that is tied to a specific product; it is simply careful craft, noticed and kept.
The trend toward lightweight, personal systems
A few years ago, the fashion was elaborate knowledge bases with taxonomies and approval flows. Many of those withered, because maintaining them felt like a second job. What seems to last is smaller: a single folder, a single notes page, a short tagging scheme. The winning feature is low friction. If saving an example costs more than a minute, it will not happen on a Thursday afternoon when a deadline is looming.
Some teams now keep both a personal file and a lightly shared one. The personal one captures taste. The shared one catches the handful of examples everyone agrees are exemplary, and it gets trimmed regularly so it stays small.
Risks and limits worth keeping in mind
The habit has real hazards, and a fair assessment names them.
The first is imitation without understanding. Copying the look of a strong report while lacking its substance produces a polished document that says nothing. The second is homogenisation: if every team studies the same few celebrated examples, outputs begin to look alike, and the differences that mattered quietly disappear. The third is borrowing wording. Reusing someone else’s phrasing is an ethical problem, and in public writing it can also hurt search visibility as repeated passages are discounted. Save the method, write the sentences yourself.
A fourth, subtler risk is confusing admiration with evidence. A beautifully written model summary is still just a summary. It tells you nothing about whether the model works. Treat examples as lessons in presentation and keep your standards for correctness entirely separate.
One more caution concerns freshness. Examples age, and a layout that felt modern two years ago can read as dated now. A quick yearly look through the file, discarding what no longer feels instructive, keeps the shelf honest. It also gives you a small, pleasant record of how your own preferences have shifted.
Starting small, without a project plan
- Pick one container: a folder, a page, a notebook.
- Decide on a single purpose, such as “clear ways to describe model limits.”
- Whenever you read something that does that well, save a pointer and one sentence about why.
- Once a month, spend twenty minutes pruning. Keep the examples you actually reopened.
- Before a hard writing task, pull two or three entries, close them, then write from your own material.
That routine is enough. The point is consistency rather than volume.
What to watch next
It is reasonable to expect the practice to become more routine in AI and AIOps teams as documentation expectations rise. People will likely become more careful about annotating why an example works, since an unexplained saved link has little value after a month. Expect, too, a greater emphasis on collecting counter-examples: pieces that illustrate how an explanation can mislead, so that the same mistake is easier to spot in one’s own draft.
The museum drawer from the opening has long since turned into a tidy digital folder, but the principle has not changed. A good swipe file is modest, selective and used. It does not make anyone’s work original, yet it keeps the bar visible, and in a field that changes as fast as this one, a visible bar is worth having.