rk-ef-pr-description

name: rk:ef-pr-description description: "Generate full PR descriptions in English (default) from code changes + Jira card link. Produces Everfit-style markdown (Summary, Ticket, What changed, Test plan, Rollout). Pass --lang=vi for Vietnamese. Triggers on: 'pr description', 'tạo description PR', 'write PR body', 'generate PR description', 'commit description from jira'." argument-hint: " [--scope=] [--lang=en|vi] [--draft]" metadata: author: rock288 version: "1.1.0"

PR Description Generator

Produce a full PR description in the team's house style (see references/template.md) given:

Code changes — the actual diff in the working tree (or a specific commit range)
Jira card link — to extract ticket ID, title, type, parent epic

Inputs

Arg	Required	Default	Notes
`<jira-link-or-card-id>`	yes	—	Full URL (`https://everfit.atlassian.net/browse/UP-71331`) or just the ID (`UP-71331`)
`--scope=<scope>`	no	inferred from changed paths	Conventional-commit scope, e.g. `video-workout`, `assignment`, `payment`
`--lang=<en\|vi>`	no	`en`	Output language for the PR body. `en` matches the team default (see §Language); `vi` for VN-only reviewer audiences.
`--draft`	no	off	Skip the "you must run tests" gate — produce a draft body even if no test results are available

Workflow

1. Resolve the Jira card

Try, in order:

Atlassian MCP: call mcp__claude_ai_Atlassian__getJiraIssue with the parsed card ID. Extract summary, issuetype.name, parent.key (if any), customfield_* for sprint if present.
Fallback — ask the user: if MCP is unavailable or returns 404, use AskUserQuestion to collect:
- summary (one-line title)
- issuetype (Story / Task / Bug / Tech Debt / Spike)
- parent epic key (optional)

Save the resolved fields into local memory for the session so later skills (e.g. rk:ef-branch-name) can reuse without re-fetching.

2. Gather code change context

Pick one source of diff, in this order of preference:

If the user passes a commit range like HEAD~3..HEAD, use that.
Else git diff origin/main...HEAD if branch is ahead of main.
Else git diff --cached (staged) and git diff (unstaged) combined.

Collect:

File list (git diff --name-only)
Per-file diff stat (git diff --stat)
Full diff (capped — see safety below)
Existing commit messages on this branch (git log --format=%B origin/main..HEAD) — these are the strongest signal about intent

Safety cap: if the full diff > 200KB, don't pipe it raw into the prompt. Instead summarise per-file: top-level symbols added/removed, then ask the user which files to expand.

3. Infer scope (if `--scope` not given)

If all changed files share a modules/<X>/ or everfit/controller/<X>/ prefix → scope = <X>.
Else pick the directory with the most changed lines.
Else ask the user.

4. Render the template

Use references/template.md verbatim as structure. Fill in:

Title: <type>(<scope>): <CARD-ID> <one-line summary> where <type> comes from the issue-type map below
Ticket line: **Ticket:** <CARD-ID> (add (parent epic: <PARENT>) if a parent epic was resolved)
Summary: 1–3 sentences. Lead with the verb (Replaces / Adds / Fixes / Removes). State what changes about behavior, not what files moved.
What changed: group by subsystem (Schema / Helpers / Services / Endpoints / Middleware). Use bullet lists, link files with `path/to/file` (backticks, not raw paths). For multi-file changes use a Markdown table like the PR-16944 endpoints table.
Test plan: checkbox list. [x] for verified, [ ] for deferred. Pull test counts from the commits' messages if mentioned (e.g. "523 passing").
Rollout: state feature flag status (none / behind flag X) and rollback strategy. If unsure, ask.

Issue-type → commit-type map

Jira issue type	Commit type	Branch type
Story	`feat`	`feat`
Task	`feat`	`feat`
New Feature	`feat`	`feat`
Improvement	`feat`	`feat`
Bug	`fix`	`fix`
Defect	`fix`	`fix`
Tech Debt	`refactor`	`refactor`
Refactor	`refactor`	`refactor`
Spike / Research	`chore`	`chore`
Documentation	`docs`	`docs`
Performance	`perf`	`perf`

If the Jira type is anything else, fall back to chore and warn.

5. Output

Write the rendered description to stdout (the conversation) — do not write it to a file unless the user passes --save=<path>.

End the response with a one-line summary: which sections were filled vs deferred, and whether any field had to be guessed (mark with ⚠️).

Language

Default: English. PR titles and bodies are written in English regardless of the chat language. Matches PR 16944 and the recent commit history (feat(challenge): UP-71209 Update response structure ...).

Override with --lang=vi when the PR audience is VN-only (e.g. internal QA / Ops-only changes). Even then:

Card ID, file paths, symbol names, error codes, commit-type prefix (feat / fix / …), and CLI commands stay in English / verbatim.
Section headings (## Summary, ## What changed, ## Test plan, ## Rollout) stay in English so search / tooling that greps them keeps working.
Only the prose inside each section flips to Vietnamese.

If the user passes the trigger in Vietnamese (tạo description PR) without --lang, the output is still English — the trigger language is not the body language. Mention this in the trailing summary if unsure.

Style rules

Match PR 16944's tone: terse, technical, factual. No marketing language.
Code references in body: use backticks, not bare paths. For files in the repo, prefer GitHub blob links when the user passes a branch name.
Tables for multi-row data (endpoints, files). Bullets for ≤ 5 items.
Never invent test counts — only quote numbers found in commits or user input.
Keep "What changed" granular but skip rote details (e.g. don't list every new test file).

When NOT to use this skill

Single-commit message body (use plain git commit flow).
Releases / changelogs (different format).
Internal design docs (those live under docs/modules/<X>/).

[[rk:ef-branch-name]] — generates the matching branch name from the same Jira card
[[rk:git]] — handles the actual commit/push/PR creation flow