Five Changes to Reflacto's Prompt, Six Models: What Really Helps

In the model test, the biggest jump came from my own prompt. So I tested five more changes, on their own and together, with six local models and 9,348 answers.

Daniel Schmeiß

With a local AI model, Reflacto helps you at every step of a reflection, for example with “Help me say it” or “Give me ideas”. Behind each of these eight helps is a prompt, an instruction to the model. In the answers from the model test, I found five mistakes the models kept making. I wrote one change against each, three for a single help and two for all eight, and tested them with six models.

An example: a teammate has missed several handover deadlines. The step asks “How can you phrase feedback constructively and solution-focused?”, the field is still empty, and the person picks “Help me say it”.

Previous prompt

Since you haven't written your phrasing yet, I cannot provide a draft for you to use. Please describe how you would like to open the conversation …

With the change

“Felix, the handover for the search indexing work was delivered on [Tuesday] instead of [Friday], and I didn't receive a message beforehand. Because of this, Priya could not start the indexing work as planned and had to move two of her own tasks to the next sprint. …”
Gemma 4 26B on the same step, excerpts from both answers. With the change, the model builds a draft from the earlier answers and leaves gaps where it lacks something.

The short version

  • From now on, Reflacto uses all five changes. That is exactly the version that ran in the test, and it meets every rule I set beforehand.
  • Three changes measurably work. “Help me say it” no longer refuses, “Give me ideas” delivers the kind of ideas the step asks for, and small models stop copying templates.
  • What works names a specific mistake. A general rule like “don't ask for something that is already there” changed almost nothing.
  • The smallest model gets no better with all five. Below a certain size, a better prompt barely helps.
Averaged over six models. Below each value, the range from the weakest to the strongest model.

The five changes

The thinking partner on a step that already has an answer: it offers three helps. Along with it, all eight helps; marked are the three with a change of their own.
  • Facts or interpretation? Observable actions such as “interrupted me” no longer count as judgments.
  • Help me say it. With an empty field, the model drafts from earlier answers instead of refusing. It makes up no numbers, dates or names.
  • Give me ideas. When the step asks for downsides, it gets downsides, not ways to avoid them.
  • Don't ask again. No questions about what is already written, no yes/no questions.
  • Shape in words. The answer form is described instead of shown as a template with <placeholders>.

What got better

Share of answers with the mistake each change was written against. Fixed check means: measured by a check in TypeScript, without AI.

With an empty field, “Help me say it” refused in one answer in eight before, now never. Ideas of the wrong kind drop from 19% to 11%, template leftovers from 3.3% to 0.7%. For “Facts or interpretation?”, false alarms drop from 75% to 45%, but by my rules that falls just short of proof. The models still ask about what is already written as often as before.

Every model reacts differently

Rated blind from 1 to 10.

Gemma 4 26B gains 2.3 points with “Help me say it”. This is the model that refused with an empty field in the model test, so the prompt was the cause. Gemma 4 E4B, my recommendation for laptops, gains 1.2 points with “Give me ideas”. For LFM3-1.7B, all five together bring nothing.

If you use a local model, you get the new prompts automatically. The local AI guide shows how to set one up.

How it was measured

I ran 47 situations in German and English through six models on my PC (GeForce RTX 5070 Ti, 16 GB), each change on its own and all five together against the previous prompt. Claude rated the answers blind from 1 to 10. A second Claude instance gave almost the same scores for 645 answers (correlation 0.93). I decided before the first run when a change counts as better and did not change that afterwards.

An AI did the rating, not a person. The situations come from the same AI that proposed the changes, so I froze them before the run. Every request ran only once.

Try it yourself

Reflacto is free, needs no account and runs in your browser. Your reflection stays on your device.

Start a reflection