From prompt chaos to one-click text editing: A tool for recurring tasks#

Text editing with AI usually requires complex prompts. We have developed a tool that works with simple buttons instead – for translations, corrections and analyses without prompt dialogue.

What was the problem?#

Formulating prompts takes time.

For everyday tasks such as translating or correcting texts, the classic chat approach is often too cumbersome. You have to rephrase what you want every time. We wanted to find out: Is it possible without prompts?

What can the tool do?#

It offers 12 predefined functions (e.g. correction, translation, summarisation, style analysis) that work at the touch of a button:

• Input-output principle: enter text on the left, select a function, get the result on the right
• Custom commands possible: Buttons can be expanded with additional instructions
• History function: Stores up to 15 editing steps with automatically generated titles

Initial tests show that the tool is being used productively for translations, corrections and stylistic adjustments.

How was it developed?#

We first created a detailed description of the requirements:

• Create specification (90 minutes): Describe functions, discuss implementation options, select systematically
• Develop prompts (30 minutes): Optimise each button so that only the edited text is returned without comments
• Implementation (60 minutes): Modular structure according to specifications
• Total effort: 3 hours over two days
• Result: 2,000 lines of code, divided into application logic, export functions and prompt definitions

Why did it go so well?#

Because we put a lot of time into preparation.

A clear description of the requirements prevented many errors later on. Instead of programming straight away, we spent 90 minutes defining exactly what the tool should be able to do. The specification went through 2-3 iterations until everything was clear.

This greatly simplified the actual programming.

Important insights#

1. A good description saves time later on#

We wrote 14 pages of specifications before a single line of code was written. This significantly accelerated the implementation. Changes were only minor tweaks, not fundamental corrections.

The specification was the foundation for everything else.

2. Status documents enable flexible working#

We introduced an ‘Implementation_Status’ document. It was updated after each step. This allowed us to jump back in at any time without losing track of the big picture.

The document always contained the complete state of affairs.

3. Split code into modules#

Each file remained under 1,000 lines (maximum 1,500 lines). This simplified maintenance. A clear separation of tasks between application logic, export and prompts made changes easy.

This kept the code clear.

4. Clear prompts sometimes need post-filtering#

Even well-structured prompts occasionally generate meta-comments. A combination of precise instructions and downstream filtering worked best. Not everything can be solved by prompts alone.

Heuristics remain useful.

What can others learn from this?#

• Specification before implementation: A detailed description of the requirements significantly reduces later problems.
• Use status documents: A systematically maintained document on the implementation status enables contextualised re-entry at any time.
• Keep files under 1,000 lines: Clear modularisation simplifies maintenance and changes.
• Combine prompt engineering with post-filtering: Precise instructions plus technical filtering work better than prompts alone
• Build tools for specific use cases: Artefact-centred approaches work well for recurring tasks

Conclusion#

✔ Text editing without prompt dialogue is practical and accepted

✔ Systematic preparation reduces implementation effort to 3 hours

✔ Specification-driven approach with status documents has proven successful

This is part of a series on AI-assisted development. The focus is on what can be learned from such projects – not just on the results.