If you build with machine learning for a living, you have likely been told at some point to “scrap it and rebuild, it will be tidier”. Sometimes that is right. More often it is how a team ends up retraining a classifier that already worked, rewriting an evaluation harness that nobody trusted less, or re-deriving a prompt that a colleague had tuned months ago. Reusing what already works is a legitimate engineering habit in AI and AIOps, and it is surrounded by myths. Let us walk through six of them, in a friendly, no-lecture way, and see what actually holds up.
Myth 1: Reusing a model component means inheriting its flaws forever
People imagine reuse as a one-way door. You pick up a feature pipeline from last quarter’s project and suddenly its quirks are yours for life. In practice a reused component is a starting point you can inspect. Say your anomaly-detection service already has a well-behaved windowing step that smooths noisy metrics. Pulling that step into a new service does not freeze it. You read it, write a test that pins its behaviour, and adapt the parameters for the new data. The flaws you find are flaws you would have met anyway, only now you meet them with a head start.
The better question is not “is this perfect” but “has this survived real traffic”. A step that has handled a few months of messy production input has a track record that a fresh rewrite cannot have yet.
Myth 2: Good AI teams always start from a blank notebook
There is a romantic story in which strong practitioners open an empty file and conjure a solution. Look at how mature teams actually work and you will see the opposite. They keep a small shelf of proven pieces: a data-cleaning function that has caught the same three kinds of bad rows for years, an evaluation script that everybody understands, a prompt skeleton with known failure cases written beside it. Starting from those is not laziness. It is how experienced people reserve their attention for the part of the problem that is genuinely new.
Next time you reach for a blank notebook, ask what in your last two projects would still apply. If the answer is “most of the plumbing”, start there and spend your creativity elsewhere.
Myth 3: Reusing a prompt is the same as copying an answer
With language models the confusion gets sharper. A tuned prompt feels like a finished artefact, so people assume reusing it means getting last time’s output again. It does not. A prompt is closer to a recipe base than a finished dish: the instructions, the format constraints, the examples and the guardrails are the stable parts, and the inputs change each time.
A support team, for example, might keep one carefully tested prompt for summarising incident tickets. When a new product line arrives, they do not write a new prompt. They swap the example tickets, add one line about the new terminology and re-run the same checks. The structure that survived testing is preserved; only the ingredients change.
- Keep the instruction block and output format together as one unit.
- Keep two or three “known tricky inputs” next to it, so a change can be checked quickly.
- Note in one line what the prompt is bad at. That sentence saves the next person an afternoon.
Myth 4: Reuse slows you down because you have to find things
This one has a kernel of truth: reuse that depends on remembering where you put something is fragile. But the cure is not to abandon reuse; it is to give your proven pieces a findable home. A single folder in the team repository, with a short readme that lists each piece in a sentence, beats any amount of heroic memory. If a component turns up within a minute of searching, pulling it in beats writing it fresh in nearly every case.
One quick test: ask a teammate to find your evaluation helper without your help. If they succeed in under a minute, your shelf works. If they do not, rename things and add a line to the readme before you move on.
Myth 5: Reuse is only for engineers, not for decision-makers
Decision-makers reuse too, usually without naming it. A manager who keeps a trusted set of questions for reviewing any model proposal is reusing a proven artefact. The same applies to the way organisations think about AI-assisted choices, a theme explored in a piece on AI and everyday business decisions. The tooling changes quickly, but a stable set of review questions, such as what data went in, what the output is used for and what happens when it is wrong, carries over from one project to the next.
If you lead a team, write those questions down once, store them where people can reach them and refine them after each review. Over a year they become one of the most valuable things your group owns.
Myth 6: If it worked in one domain, it must transfer unchanged
This final myth leans in the opposite direction and deserves equal honesty. Reuse does not mean copy-and-pray. A feature-engineering routine tuned on one kind of data may encode assumptions that quietly fail elsewhere. An education team and an operations team can share the shape of a workflow without sharing its thresholds, which is a point worth keeping in mind when reading about AI in learning settings and comparing it to your own context.
The habit that guards against this is a short “adaptation note” attached to each reused piece: what did we change, and why. If you cannot write that note, you probably have not understood the component well enough to reuse it yet, and that is useful information.
What a healthy shelf looks like after a few months
Picture the goal state before you start collecting. A working shelf is small, usually a dozen or so items rather than hundreds, because every piece on it has earned its place by being used. Each item has a one-line purpose, a note on its limits and a saved example. Old items are retired when nobody has touched them for a long stretch, so the shelf does not turn into a junk drawer. Newcomers on the team can be pointed to it on day one, and they tend to ask sharper questions because they can see what the team already trusts. If your shelf grows past what anyone can scan in a few minutes, that is a signal to prune, not to add a search engine.
Notice also that the shelf has a cultural effect. When people see that a tidy, well-labelled contribution gets reused by others, they start writing their own work in a way that is easier to lift out. Reuse becomes a quiet incentive for clearer code and clearer prompts.
How to start a reuse shelf this week
You do not need a platform or a procurement process. Try this modest sequence:
- Pick the last three things that went smoothly in your AI or ops work: a pipeline step, a monitoring rule, a prompt.
- Move each into one shared place, with a sentence about what it does and what it should not be used for.
- Add a tiny test or a saved example input and its expected result.
- The next time a similar task appears, begin from the shelf and record what you adapted.
After a month, look at which items were used again. Those are your keepers. The rest can be tidied away without guilt.
A glance at where people compare notes
Teams that share a shelf often want to see how others organise theirs. Beyond blogs and repositories, there are community spaces for that kind of exchange. One example is RanchiOrbit, an events website whose sections cover event listings and venues as well as a forum, blog and stories, which is a reminder that meetups and workshops still sit alongside online reading. Check any organiser’s own page for details before you plan around one.
What to remember
Underneath all of these sits one assumption, that reuse and quality pull against each other. They seldom do. A proven component that is tested, labelled and adapted with care tends to be more reliable than a new one that has seen no traffic. The real discipline is in the labelling, testing and adapting, not in the rebuilding.
So the next time someone suggests starting over, ask two questions. What already works here, and what would we lose by throwing it away? You may still decide to rebuild. But you will be doing it for a reason, not out of habit, and your shelf of reusable pieces will keep getting better with each project. That is how AI and AIOps teams quietly get faster over time: not through heroic rewrites, but through remembering what worked and being able to find it.