Review Story

PR Story Generator

You are generating a structured narrative that tells the story of a pull request. The output must be parseable for later visualization.

The user will provide a PR number or URL. Use it in the gh commands below.

Step 1: Gather PR metadata

Run these commands to collect context:

gh pr view <pr> --json title,body,author,baseRefName,headRefName,commits,files,additions,deletions

Extract the PR number, branch name, base branch, title, and description. Pay attention to the PR description (body) — it often explains the author's intent, motivation, and context that aren't visible from the code alone. Use this to inform your narrative.

Step 2: Get PR comments and review feedback

Fetch conversation comments and review comments to understand the discussion around the PR.

Important: In interactive Bash shells, the history expansion character can break jq expressions that use the "not equal" operator. Avoid using that operator in jq; use length > 0 instead. Also avoid jq string interpolation \(...) inside arguments — use string concatenation with + instead.

gh pr view <pr> --json comments --jq '.comments[] | "**" + .author.login + "**: " + .body'

For review-level comments:

gh pr view <pr> --json reviews --jq '.reviews[] | select(.body | length > 0) | "**" + .author.login + "** (" + .state + "): " + .body'

For inline code-level review comments, first get the repo from the PR URL, then use the API:

gh pr view <pr> --json url --jq '.url'
gh api repos/<owner>/<repo>/pulls/<pr-number>/comments --paginate --jq '.[] | "**" + .user.login + "** on `" + .path + "`:\n" + .body + "\n---"'

These comments provide valuable context — they may reveal why certain decisions were made, what was revised, or what trade-offs were considered. Weave relevant insights from the discussion into your narrative where appropriate.

Step 3: Get the commit history

gh pr view <pr> --json commits --jq '.commits[] | .oid + " " + .messageHeadline'

If there are multiple commits, this is important — you will narrate the evolution of changes commit by commit.

Step 4: Get the full diff

gh pr diff <pr>

If the diff is very large (over ~2000 lines), focus on the most structurally significant changes. Skip trivial changes like import reordering, whitespace, lock files, and auto-generated code.

Step 5: If multiple commits exist, examine evolution

For each significant commit, look at what specifically changed. Use the GitHub API so commits are available even when the local clone doesn't have the objects (e.g. fork-based PRs):

gh api repos/<owner>/<repo>/commits/<commit-sha> --jq '.files[] | .filename + " | " + .status + " | +" + (.additions | tostring) + " -" + (.deletions | tostring)'
gh api repos/<owner>/<repo>/commits/<commit-sha> --jq '.files[] | select(.filename == "<specific-file>") | .patch'

This helps you explain HOW the code evolved across commits, not just the final state.

Step 6: Write the story

Output format

Write a Markdown document. The narrative is prose text explaining what the author did and why, interspersed with code snippets.

Every code snippet must be wrapped in a fenced block with the story-diff info string and a JSON metadata comment on the first line:

Some narrative explaining what happens next...

```story-diff
<!-- {"file": "src/api/users.ts", "commit": "a1b2c3d", "type": "addition", "description": "New user validation endpoint"} -->
+export async function validateUser(req: Request) {
+  const { email, password } = req.body;
+  if (!email || !password) {
+    throw new ValidationError('Missing credentials');
+  }
+}
```

Then the author moved on to...

Metadata fields:

• file — the file path
• commit — short SHA (if attributable to a specific commit, otherwise omit)
• type — one of: addition, deletion, modification, rename, context
• description — brief label for what this snippet represents

Line references

When the prose needs to point the reader to specific lines in the immediately following snippet, use line references:

• In prose, use **[N]** — visible bold markers that readers can follow
• In code, use <!--ref:N--> appended to the referenced line
• Reference numbers are unique throughout the entire story (1, 2, 3, ...) — never restart numbering
• The prose anchor [N] must always appear before the code snippet containing its matching <!--ref:N-->
• Only use references when they genuinely aid comprehension — not every snippet needs them

The author initializes the connection pool **[1]** but defers cleanup
to a shutdown hook **[2]**, which ensures graceful teardown even on
SIGTERM.

```story-diff
<!-- {"file": "src/db/pool.ts", "commit": "f4e5d6c", "type": "addition", "description": "Connection pool with graceful shutdown"} -->
+const pool = new Pool({ <!--ref:1-->
+  max: 20,
+  idleTimeoutMillis: 30000,
+});
+
+process.on('SIGTERM', async () => {
+  await pool.end(); <!--ref:2-->
+  process.exit(0);
+});
```

Code snippet rules:

• Use unified diff format with +/- prefixes for added/removed lines
• Include a few lines of unchanged context (prefix with space) where it helps understanding
• Keep snippets focused: 5–40 lines each. Trim aggressively.
• For large structural additions, show the signature/shape, not every line

Narrative rules

• Tell a story, not a list. Write flowing prose: "First, the author introduced...", "To support this, they added...", "This required updating..."
• Follow logical order, not file order. Group by intent: "Here's the new API endpoint, and here's the frontend form that calls it."
• If there were multiple commits, narrate the evolution. "In the first commit, the author scaffolded the migration. In the second, they refined the error handling after..."
• Be opinionated about what matters. Skip boilerplate. Highlight architectural decisions, tricky logic, and non-obvious choices.
• Incorporate discussion context. If PR comments or reviews reveal why a decision was made or what was debated, mention it naturally: "After feedback on the initial approach, the author switched to..." or "As noted in the review discussion, this trade-off was intentional because..."
• Note potential issues or interesting patterns if you spot them, but keep it brief.
• Start with a one-paragraph summary of what the PR accomplishes before diving into the detailed story.
• End with a brief conclusion summarizing the overall change.

Length guidelines

• Small PR (<200 lines diff): 3–8 code snippets, ~500–1000 words of prose
• Medium PR (200–1000 lines): 5–15 code snippets, ~1000–2000 words of prose
• Large PR (1000+ lines): 8–20 code snippets, ~1500–3000 words of prose. Be highly selective.

Step 7: Save the output

Save the story as pr-story-<pr-number>.md in the current directory.