The prompt and the context kept in git and released like code, in twelve slides: the context window of one call, prompt and context engineering, the prompt file, a prompt change in a pull request, tool definitions, the release of prompt and model together, the prompt pipeline, a prompt registry and the code repository, context rot, trimming and summarising the history, and prompt caching.
The slides
The tool is lesson 3 of
LLMOps for AI Agents, in twelve slides:
- What is a prompt?: one sentence of a prompt changed in the running service, and the same change made like a code change.
- The context window of one call: the five parts our code assembles, and where each comes from.
- Prompt engineering and context engineering: the instructions, and the code that chooses everything else.
- The prompt file in git: the on-call assistant's prompt, with its version, model, settings and templates.
- A prompt change reviewed in a pull request: the diff, and the eval checks beside it.
- Tool definitions: prompts the model reads: the name, the description and the schema of
get_pods. - The release: prompt, model version and settings: why a prompt and its model change together.
- The prompt pipeline: pull request, evals, deploy: the gate, and the rollback by revert.
- A prompt registry and the code repository: where the running service gets its prompt.
- Context rot: quality falling as the window fills: why a lean window beats a full one.
- Trimming and summarising the history: a long conversation compacted.
- Prompt caching: a stable prefix reused: the order of the parts, and the call that misses the cache.
The buttons on the left and right edges move between slides. So do the arrow keys and a swipe.
The lab
The token counts, the eval scores and the checks in the pull request are examples made up for the drawings. The
context-rot curve is illustrative: it has the shape the studies in the sources report, not numbers measured on one
model. The totals of the window and of the compacted history are computed by the slides.
Sources