# Two years of prompt engineering for coding work, what actually moved the needle — retained record

> Retained reddit record from reddit:PromptEngineering.

Canonical: https://highsignal.app/data/records/87676dff80ba055b

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:PromptEngineering)  published 2026-09-14 12:07 ingested 2026-09-15 02:39

[original source ↗](https://www.reddit.com/r/PromptEngineering/comments/1wg1l47/two_years_of_prompt_engineering_for_coding_work/)

## retained text



```
Studied engineering, IT side. Worked through most of it, part time during the day and a five hour night shift on top, and I did most of my actual learning on that night shift with a laptop open when nothing was happening. Been doing this properly for about two years now.

Did two prompt engineering trainings somewhere in there. Neither was bad but what I actually took from them is that the people who are good at this are the ones using models constantly, not the ones who sat through a course.

Things that made a real difference for coding work:

1. **Tell it what not to touch.** My early prompts all described what I wanted built. The failures came from what it changed on the way there. Now I list the files and behaviours it must leave alone before I describe the task at all. Cut my "it worked but broke something else" rate more than anything else I have tried.
2. **Make it restate the plan before writing anything.** One paragraph, in its own words, what it thinks the task is. Half the time the restatement is wrong and I catch it before any code exists instead of after reviewing forty lines.
3. **Static content first, dynamic content last.** Took me embarrassingly long to understand this one. Caching works on prefix matching, so anything that changes near the top invalidates everything after it. I had my working notes at the start of the prompt and the stable project context underneath, which is exactly backwards. Flipping it made a real difference to cost on long sessions and took five minutes.
4. **Do not max the effort setting on everything.** Most models expose some version of this now. Turning it up on a rename or a boilerplate file buys you nothing, and there is reasonable evidence that on some models high effort does slightly worse on coding than medium. Worth testing on your own tasks instead of assuming more is better.
5. **Skip the tricks that stopped working.** Telling it to think step by step, assigning it an expert persona, telling it not to hallucinate. On current reasoning models these do nothing measurable and the step by step one adds latency for it. Few shot examples still outperform all of that.

6. **The same prompt does not land the same way on two models.** I run Opus 5 and glm-5.3 alongside each other and the difference in instruction following is bigger than I expected. One tends to take an instruction literally, the other tries to work out what you actually meant and fills gaps on its own. Neither is wrong, they just went through different post training, so a prompt tuned against one is not automatically tuned for the other. Swapping the end point and keeping the prompt identical does not work. Worth noting the Chinese models are also cheaper per token than most of the western ones, which matters more than it sounds like it should once you are running things in a loop.

None of this is clever. It is mostly being specific about constraints and then not undoing that with clutter.
```



document fetched 2026-09-15 00:41 · canonical https://reddit.com/r/PromptEngineering/comments/1wg1l47/two_years_of_prompt_engineering_for_coding_work
