Editing Drift — Why Instructions Stack Badly

The frankenstein prompt

You write a prompt. Someone reports a bug. You add a sentence. Another bug — another sentence. Six months later the prompt is 4,000 words and contains contradictions you didn't notice. This is editing drift, and it's the most common quality failure in production prompting.

Symptoms

Sentences that contradict each other ("be concise" + "include all relevant context").
Stale fixes — instructions for bugs that were repaired upstream months ago.
Cargo-cult sentences inherited from prompts no one remembers.
Response shape spec drifting away from the actual schema.

The hygiene routine

Diff before adding — read the existing prompt before adding a new sentence.
Replace, don't append — if a new instruction overlaps an old one, edit the old one.
Date stale fixes — comment with the date and reason; remove when the cause is gone.
Periodic prompt review — every quarter, run the eval suite and rewrite the worst-performing prompt from scratch. Most are shorter when rewritten clean.
One owner — a prompt with shared edit access drifts faster than one with a named owner.

Code

Comment-based hygiene markers·markdown

<!-- 2026-04-12: tighten output JSON; was leaking markdown fences. owner: Alice -->
Return JSON only. No code fences. No preamble.

<!-- 2026-02-08: stop apology preambles. owner: Bob -->
Open with the verdict, not an apology.

<!-- TODO 2026-Q3: revisit; the underlying parser was fixed -->
<!-- The next sentence is probably stale: -->
If the model output is malformed, reformat it to JSON before returning.

Editing Drift — Why Instructions Stack Badly

The frankenstein prompt

Symptoms

The hygiene routine

Code

External links

Exercise

Progress

Comments 0