- Author
- Mario Zechner
- Published
- 2025-06-02
- Kind
- Personal blog post with one worked example from the author's own project
- Original
- mariozechner.at/posts/2025-06-02-prompts-are-code/
- Subject
- File-based agent memory
- Themes
- Saying it exactly, What lasts
What it is
Mario Zechner is a developer who maintains the Spine runtimes. The post sets out the mental model first. It then walks through a prompt file that ports Java changes to other languages.
What it says
- Zechner says coding agents struggle on big codebases. Causes include missing context, weak execution-flow tracking and lack of taste.
- The post says things fall apart around 100k tokens, regardless of benchmark claims.
- Zechner proposes a mapping. The prompt is the program, and tool calls and user input are I/O.
- State is serialized to disk, as JSON or Markdown. Work can then resume with fresh context.
- The example is a port.md prompt. A script pre-generates a JSON plan of changed Java types.
- Zechner reports that work which took 2-3 weeks now takes 2-3 days. Human checkpoints stay in the loop.
What evidence it gives
The evidence is one real project. The post shows the prompt file, the JSON state format and the jq queries. A video recording is described. The time saving is the author's own report with no measurement.
Interests and limits
Zechner maintains the codebase and designed the workflow, so the result is self-reported. Only one task and one coding tool were used. The post says the LLM cannot do the full port alone. It also says this is not a panacea. The long code listings and middle workflow steps were skimmed, not read in full.
Connections
These links are this wiki's own reading, not claims the sources make.
- On the foolishness of "natural language programming" (EWD 667). Zechner calls a natural language prompt a program. Yet the precise work moves into a generated JSON plan, built by a script. Dijkstra's prediction shows up inside the workflow that seems to refute it.
- Effective harnesses for long-running agents. Zechner and Young's team separately reach for JSON on disk. Each new session starts cold, so progress must live in a file. Zechner's plan comes from a script, and Young's from an agent. Which author of the plan can you trust more?