Reusable prompt snippets the team pastes into Claude / Codex / other LLM sessions to get consistent output quality. One file per prompt, one prompt per file. Plain markdown. Copy-paste or reference by URL.
Current prompts
| File | Purpose |
|---|---|
research-quality.md | ”Do research at the oxFlow house quality level.” Paste into any Claude / Codex session when you want a research deliverable matching the existing research notes. |
How to use a prompt
Three ways:
- Copy-paste. Open the file on github.com or locally, copy the whole body, paste into the target chat.
- Reference by URL. In a Claude Code / app session, say: “read
https://github.com/361-coders-nz/oxflow-hub/blob/main/prompts/research-quality.mdand follow it for the next task.” Claude will fetch it. - @-import (Claude Code only). Drop
@prompts/research-quality.mdinto your message; Claude Code resolves the path from the repo root.
Adding a prompt
Keep it short. Good prompts are:
- Self-contained. They work in a session that doesn’t have the repo cloned.
- Specific. Name the deliverable, the structure, the voice, the quality bar.
- Exemplar-linked. Point at at least one committed example so the reader can see “this is what good looks like.”
- Versionable. Change it over time as the team’s quality bar evolves. Git history is the versioning.
Filename convention: lowercase, hyphenated, verb-led where possible (research-quality.md, spec-a-feature.md, review-a-pr.md).
What does NOT go here
- Full system prompts for agents — those live in
.claude/agents/(per the shared project memory note recommendation). - Slash commands — those live in
.claude/commands/. - One-off session content — belongs in
status/sessions/or a research note. - NDA-covered or client-confidential material.
Related
- Shared project memory research — the broader memory-layer context.
- Agentic coding + PR review research — the review workflow these prompts feed into.
_template-research-output.md— the agent-facing version of the research quality bar (duplicate content intentional — this folder is human-facing, that one is agent-facing).