Commands Experimental

mp-refine-run

Generate or refine prompts with local and paid strategies, append behavior, and explicit boolean flags.

Use this command to generate prompt packs or refine an existing prompt/tool description. It offers local template generation, online prompt-factory work, and classic recursive meta-prompting—not a single fixed refinement strategy.

Start here

With Ochat installed, write a task description in task.md:

mp-refine-run -task-file task.md -output-file generated-prompt.md

The default strategy may call models and incur charges. To use the local template factory without model calls:

mp-refine-run -task-file task.md -meta-factory true -output-file local-pack.md

Output is appended, not replaced, when -output-file is supplied. Use a fresh file for a separate result. Without that flag, output goes to stdout.

Options

Flag	Default	Meaning
`-task-file FILE`	Absent	Task description. Recommended; despite the help saying required, the parser accepts omission and uses empty task text.
`-input-file FILE`	Absent	Existing prompt to iterate; absence selects creation paths.
`-output-file FILE`	Absent	Append result to this file; otherwise print it.
`-action generate\|update`	`generate`	Action supplied to the recursive flow.
`-prompt-type general\|tool`	`general`	Select general/tool behavior where the selected strategy distinguishes it.
`-meta-factory BOOL`	`false`	Local template factory; takes precedence over the other strategy flags.
`-meta-factory-online BOOL`	`true`	Enable online factory strategy.
`-classic-rmp BOOL`	`false`	Disable online factory mode; does not override `-meta-factory true`.

These booleans take explicit true or false arguments.

Strategy selection

With -meta-factory true, call Prompt_factory.create_pack without an input file or iterate_pack with one. This is local template generation.
Otherwise online mode is enabled only if -meta-factory-online true and -classic-rmp false.
With no input file and online mode enabled, try Prompt_factory_online.create_pack_online with the GPT-5 model constructor. If it returns no result, fall back to the general recursive flow with online factory mode disabled.
Other cases select Mp_flow.first_flow or tool_flow according to prompt type, passing the chosen online setting.

For example, refine a tool description using the classic strategy:

mp-refine-run -task-file task.md -input-file draft-tool.md \
  -prompt-type tool -action update -classic-rmp true \
  -output-file refined-tool.md

The binary does not expose a universal iteration, model, or spending-cap flag. Review the selected library strategy before automating paid runs. Invalid action/type values explicitly exit 1; parsing, file, or provider failures may also exit nonzero. There is no documented fixed 20,000-token truncation contract.

References

Meta-prompting library: evaluators and recursive flows.
Prompt factory: prompt-pack templates.
Provider environment: component-specific settings.
Implementation: exact strategy dispatch.
Command index.