Prompt refinement

A terse or under-specified prompt ("fix that", "add the flag") makes the model guess, which wastes the first few tool rounds. Prompt refinement runs one cheap LLM call before the turn starts that rewrites your prompt into a clearer, self-contained version, then runs the turn on the rewrite.

text

your prompt -> [refine call] -> refined prompt -> normal turn (preamble, tools, goal, judges)

On by default. It is a small pre-turn cost that makes the turn land better. Turn it off with /refine off if you want prompts sent verbatim.

You see the rewrite every turn. When the prompt is changed, the REPL prints it before any work starts:

[refine] -> Add a --dry-run flag to the sync command and add a unit test that exercises it.

The refined text is what the model receives, what /prompt shows as the user message, and what is saved to the session and the memory bridge for that turn - so a later model switch sees exactly what was answered.

It sees the conversation so far. The refine call gets the same prior-turn history the real turn does, so it can resolve what a reference actually means - "change that to use port 465" becomes concrete once the refiner can see the config you were just discussing. Only prior turns are included; the prompt being refined is not yet part of the history at this point.

What it will not do. The rewrite preserves your intent and every concrete detail (names, paths, numbers, constraints). It does not answer the request, add requirements you did not imply, or invent facts. An already-clear prompt comes back essentially unchanged (and no [refine] line is printed).

When it is skipped.

Short or trivial prompts (a one-liner, a bare question) - refining them is not worth a call.
No model is served yet (a local backend that has not been downloaded/started).
The refine call errors, returns nothing, or returns a runaway expansion - the original prompt is used unchanged. Refinement can never block a turn.

Refinement runs before goal decomposition and the judge roster, so the plan and the panel are built from the refined prompt.

From the REPL

/refine            # show whether refinement is on or off
/refine on         # enable (persists across sessions)
/refine off        # disable (persists across sessions)

The choice is written to ~/.kdeps/agent-loop-settings.yaml and restored on the next kdeps run.

Prompt refinement ​

From the REPL ​

See also ​

Prompt refinement

From the REPL

See also