# Prose

> Use when the user mentions OpenProse or prose workflows, wants explicit multi-agent research and synthesis, or wants a reusable orchestration pattern. In Codex, emulate OpenClaw OpenProse using spawn_agent, structured decomposition, and workspace artifacts.

- Skill: `hoanganhduc/prose` (Agent Skill)
- Install (CLI): `npx skillmds@latest add hoanganhduc/prose`
- Raw SKILL.md: https://api.skillmd.com/api/skills/hoanganhduc/prose/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: hoanganhduc (https://skillmd.com/u/hoanganhduc)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/hoanganhduc/prose

---


# Prose

This is a Codex adaptation of OpenClaw OpenProse, not the OpenProse VM itself.

## Concept mapping

- OpenProse `agent` / `session` -> Codex `spawn_agent`
- OpenProse `parallel` blocks -> multiple independent spawned agents
- OpenProse file state -> normal workspace files and notes

## When to use

- The user explicitly mentions `prose` or `OpenProse`
- The task is a research pipeline with separate roles
- The task benefits from explicit parallel work and final synthesis

## Workflow

1. Break the task into independent tracks.
2. Keep the immediate blocking step local.
3. Spawn agents only for bounded, non-overlapping subtasks.
4. Ask each spawned agent for concrete output, not vague exploration.
5. Integrate results locally into the final answer or file.
6. Before final writing, load `writing-style-settings.md`, select any relevant
   overlay such as `math-manuscript-style.md`, and record the active style
   profile in the workflow manifest or final synthesis artifact. Final writing
   records should include `style_profile_ref`, `policy_hash`,
   `active_overlays`, `active_requirement_ids`, and `style_applied`.

## Good patterns

- Researcher + writer
- Comparator A + comparator B
- Source gathering + synthesis
- Evidence collection + verification

## Constraints

- Do not spawn agents just to duplicate your own immediate next step.
- Reuse or wait on sub-agents only when their result is actually needed.
- Keep ownership clear if multiple agents may write files.
- Do not present writing as final when the active style profile was not loaded
  or recorded for a writing-producing deliverable.
- Do not treat a bare `style_applied: true` assertion as sufficient evidence
  unless the workflow also records the loaded policy and selected requirement
  IDs.

## User expectation

If the user gives an actual `.prose` file, read it and translate its intent into Codex-native orchestration rather than pretending Codex can run the VM directly.