一键导入
pr-create
Proactively draft and create pull requests by inferring title, refs, and body from git context, then confirm and open via gh CLI.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
菜单
Proactively draft and create pull requests by inferring title, refs, and body from git context, then confirm and open via gh CLI.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | pr-create |
| description | Proactively draft and create pull requests by inferring title, refs, and body from git context, then confirm and open via gh CLI. |
| license | MIT |
| compatibility | claude cursor opencode |
| metadata | {"version":"1.1.0","languages":"all","audience":"developers","workflow":"automation","since":"2026-02-25"} |
I take ownership of PR authoring end-to-end:
gh pr create after explicit user approval.The default outcome is an opened GitHub PR (not just a draft message), unless you choose to edit before creation.
Supporting files in this skill:
assets/pr-description.template — Markdown structure used when composing the final bodyUse when you want to:
Stage: Authoring and opening a PR (before review starts).
pr-self-review when the PR draft is ready and you want an author checklist before requesting review.pr-split when the main question is PR size/scope and whether to split.github-pull-request-handling when review threads already exist and you need to resolve/reply to comments."Help me write a PR for this feature" "Create the PR from this branch, you choose title/body" "Open a PR and infer ticket refs from commits" "Draft everything, show me once, then create with gh if I approve"
Here is the default execution flow (agent-led, low-friction):
Gather context without asking the user for manual inputs first:
main/master or configured default)Parse issue identifiers from branch names and commit messages. Prefer existing refs over asking the user.
Common patterns to detect:
[A-Z][A-Z0-9]+-\d+ # Jira style, e.g. PROJ-123
#\d+ # GitHub issue shorthand, e.g. #456
(?:fixes|closes|refs)\s+#?\d+ # commit trailer style refs
Examples:
PROJ-123-feature-name -> PROJ-123fix(auth): handle timeout (refs #812) -> #812closes #991 -> #991If no reference is found, ask one focused question:
Generate one best-guess semantic title using:
<type>(<scope>): <subject>
Where:
feat, fix, chore, ci, deploy, docs, improvement, refactor, testauth, api, ui)Produce a ready-to-submit body with:
## Summary (1-3 outcome-focused bullets)## Changes (grouped by area or subsystem)## Test plan (actual commands run and/or manual verification steps)## Risk / rollout (risks, migration notes, rollback hints when relevant)Refs: ... line when an issue reference is availableDo not ask the user to write these sections first. Draft them from repository context, then let the user edit/approve.
Show the proposed title + body and ask:
gh pr create?"Only ask additional questions if required information is truly missing (for example, no issue ref found and project requires one).
gh on explicit approvalAfter clear approval:
gh pr create --title ... --body ... (use HEREDOC for body)If the user does not approve, stop after sharing the draft and wait for edits.
Use this structure for the generated body:
## Summary
- (1–3 bullets describing what changed and why)
## Changes
- (group by area: `python/`, `infra/`, `service/`, etc.)
## Test plan
- [ ] (exact commands you ran or manual steps you completed)
## Risk / rollout
- (optional) Risks, staging notes, rollback instructions
Always include Refs: <ticket> when possible (Jira, Linear, GitHub issue).
If there are breaking changes, include migration notes in Risk / rollout.
ALWAYS use this skill whenever you would otherwise ask the user a question in free-form chat -- for clarifications, confirmations (especially destructive actions), missing parameters, multiple-choice decisions, or structured form input. Elicitations are routed through the Unique AI Platform UI via `unique-cli elicit create` + `elicit wait` (or `elicit ask` outside an agent harness) so the user gets a proper structured prompt and you get a structured answer back. Do NOT ask the user in plain chat when you can use this skill instead.
Read Agentic Table (magic table / due-diligence) sheets through the unique-cli agentic-table command. Use when the user or task involves inspecting an Agentic Table: a sheet's state and metadata, a specific cell's value or lock state, a cell's edit/answer history, or the sheet's generated export artifacts (reports, question exports). These are read-only (Tier 0) commands — they never modify the sheet and never require confirmation. Access to each sheet is enforced by the platform; a denial is reported as `agentic-table: permission denied`.
Call MCP (Model Context Protocol) server tools on the Unique AI Platform using the unique-cli mcp command. Use when the user asks to invoke, call, or execute an MCP tool, or when they need to send a JSON payload to an MCP server through the CLI. The JSON payload is forwarded 1:1 to the platform's MCP call-tool API.
Invoke connected Unique spaces/subagents through the unique-cli subagent command. Use when the workspace exposes connected-space tools and you need to delegate a question or task to one of those configured assistants.
Search the Unique AI Platform knowledge base using the unique-cli search command, with automatic per-turn citation tracking so cited facts render as clickable reference chips and `<sup>N</sup>` footnotes on the Unique platform. Use whenever the user asks to find, search, or query documents or content on Unique, including filtering by folder or metadata. Also covers `unique-cli read <cont_id>` for reading the full indexed text of a document when its content ID is already known. NOTE: This search uses combined vector + full-text indexing. Excel (.xlsx/.xls), CSV (.csv), and image files are NOT full-text indexed, so they will not appear in search results. To locate these file types, use the unique-cli-file-management skill instead (browse folders with `unique-cli ls` to find them by name).
Search the documents uploaded for the CURRENT task/row (e.g. an Agentic Table row's attached files) via the `unique-cli uploaded-search "<query>"` command, with the same per-turn citation tracking as `unique-cli search` so cited facts render as `<sup>N</sup>` footnotes and clickable reference chips on the Unique platform. ALWAYS use this skill when the user refers to documents they uploaded/attached to this task, or when you need facts from the task's own attached files. These uploaded files are scoped to the chat, NOT to a knowledge-base folder, so they will NEVER appear in `unique-cli search` results no matter the folder or metadata filters. Use `unique-cli search` for the knowledge base and this command for the task's uploaded documents; the two are complementary and citation numbering is shared across them within a turn.