A prompt that worked well on day one rarely stays that way untouched. Staff add a line here to handle an edge case, another there to fix a one-off complaint, and eighteen months later a once-clean 200-word instruction has become a sprawling 1,600-word document full of overlapping, sometimes contradictory rules that nobody fully remembers the reasoning behind. Prompt hygiene is the discipline of preventing that rot, or cleaning it up once it's happened.
How prompts actually rot
Additive patching: a new instruction added for each new edge case, never removed even after the case stops occurring
Contradiction creep: two instructions added months apart that quietly conflict, with the model picking one inconsistently
Context bloat: examples and explanations that made sense when the prompt was simple, now redundant against later additions
Ownerless changes: multiple staff editing the same prompt over time with no record of who changed what or why
A worked example of rot in action
A Melbourne recruitment firm's candidate-screening prompt had grown from a clean 180-word instruction to just under 2,000 words over fourteen months, with three different staff members having added rules at different times to handle specific complaints or edge cases. A review found two directly contradictory instructions about how to handle candidates with employment gaps, added five months apart by two different people who'd each been solving a different specific complaint without checking what was already there. The model had been resolving the contradiction inconsistently, which nobody had noticed until a systematic review compared outputs against the prompt line by line.
A simple maintenance routine
Assign one owner per standing prompt, responsible for reviewing and approving any change rather than allowing ad hoc edits
Set a quarterly review date, reading the full prompt fresh rather than just skimming the most recent additions
Test any proposed new instruction against a handful of past real examples before adding it permanently
Remove instructions for edge cases that haven't recurred in the last six months, rather than keeping every rule indefinitely
Why this is worth the discipline
A rotted prompt doesn't usually fail dramatically, it degrades quietly, producing slightly-off output that staff learn to work around rather than flag as a problem worth fixing. The Melbourne firm's fix, a half-day rewrite reconciling the contradictions and removing three stale edge-case rules, cost roughly $650 in review time and produced a prompt under 400 words that performed measurably more consistently than the 2,000-word version it replaced.
If a standing prompt in your business has grown considerably since it was first written, that's usually worth a review. Get in touch through /contact and we'll help you clean it up.
Building the review into an existing habit
Rather than creating a new standalone process, attach the quarterly prompt review to whatever regular operational review a business already runs, a monthly team meeting, a quarterly ops check-in, so it doesn't compete for calendar space against everything else. The businesses that maintain this discipline long-term are usually the ones that made it a five-minute agenda item on an existing meeting, not a separate task that's easy to deprioritise when things get busy.
It's also worth keeping a simple changelog alongside any standing prompt, even just a dated one-line note of what changed and why. This single habit would have caught the Melbourne firm's contradiction immediately, since a changelog entry noting 'added employment-gap handling rule' from five months earlier would have been visible to the second staff member adding a conflicting rule, rather than the change happening invisibly inside a document nobody was tracking edits to.
A related habit worth adopting: before adding any new instruction to a standing prompt, ask whether an existing instruction could be adjusted instead of adding a new one alongside it. Most rot comes from addition rather than revision, and a team that defaults to editing an existing rule to cover a new case, rather than bolting on a separate one, keeps the prompt materially leaner over the same eighteen-month period.
None of this is about achieving a perfect, permanently clean prompt. Some accumulation is normal and healthy as a workflow genuinely needs to handle more cases over time. The goal is catching the difference between necessary growth and unmanaged sprawl, which a quarterly read-through with a genuinely critical eye reliably distinguishes, even without any formal tooling to support the process.
If it's been more than six months since anyone last read a standing prompt start to finish, that alone is a reasonable trigger for a review, regardless of whether anyone has noticed a specific problem yet.
A prompt that's genuinely lean tends to also be faster and marginally cheaper to run, since every extra word of accumulated instruction is tokens sent with every single call. Cleaning up rot isn't purely a quality exercise, it usually pays for itself partly in reduced ongoing token cost as well.
