Human Pr Writer

Human PR Writer

Write PR descriptions that sound like a person wrote them.

Before Writing

Ask yourself:

1. What's the human impact? Why does this matter to users?
2. What questions will reviewers have? Pre-answer them.
3. What's the best reading order? Guide reviewers through the diff.

Gather Context

Run the context script to extract git information:

uv run scripts/pr_context.py

This outputs JSON with:

• Current branch and base branch
• Changed files with additions/deletions
• Recent commits

For just the diff:

uv run scripts/pr_context.py --diff

PR Structure

Title

[Impact prefix] [Action] [Context]

Keep under 72 chars. Human-readable.

Good:

• "Fix login timeout on slow connections"
• "Add password reset flow"
• "Speed up dashboard load by 40%"

Avoid:

• "Implement comprehensive authentication enhancement"
• "This PR introduces..." (never start with "This PR")

Opening (The Why)

Start with user impact, not technical implementation.

Template: > Users were experiencing [pain]. This change [solution] so that [benefit].

Example: > Users were getting logged out mid-session. This adds a refresh token rotation so sessions stay alive during active use.

Avoid: > This PR implements a comprehensive authentication enhancement to improve user experience.

The Plot

Explain key changes with trade-off explanations.

Template:

## What changed

- [Change 1] because [reason/trade-off]
- [Change 2] - decided against [alternative] because [why]

## What I didn't do

- [Out of scope thing] - filing as #XXX for later

Example:

## What changed

- Switched to refresh tokens instead of extending session cookies. Cookies were getting too large and causing issues with some proxies.
- Added rate limiting to the refresh endpoint. Considered using Redis for distributed rate limiting but stuck with in-memory for now - we only have one instance.

## What I didn't do

- Didn't migrate existing sessions. Users will need to log in again after deploy. Filed #456 to add a migration script later.

For Reviewers

Help reviewers navigate your changes.

Template:

## For reviewers

**Reading order:** [file1] → [file2] → [file3]

**Screenshots:** [prompt for screenshots if UI changes]

**Questions you might have:**
- Why [X]? → [answer]

Example:

## For reviewers

**Reading order:** `auth.py` → `middleware.py` → `tests/test_auth.py`

The diff looks big but most of it is test fixtures. The actual logic is ~50 lines.

**Why not use the existing token util?** It doesn't handle the refresh case correctly - it was generating new tokens on every request instead of reusing them.

Humanizing Checklist

Before finalizing, check against [references/humanizing-pr-patterns.md](references/humanizing-pr-patterns.md):

• [ ] Opens with user impact, not "This PR..."
• [ ] No AI vocabulary (crucial, pivotal, enhance, seamless, etc.)
• [ ] No inline-header bullet lists ("Performance: ...")
• [ ] No em dash overuse
• [ ] States what wasn't done (scope honesty)
• [ ] Pre-answers obvious reviewer questions
• [ ] Has some personality/voice, not just dry facts

Creating the PR

Use gh CLI to create:

gh pr create --title "[title]" --body-file pr_description.md

Or edit inline:

gh pr create --web

Quick Reference

| Robotic | Human | |---------|-------| | "This PR implements..." | "Users were hitting X, so I..." | | "Additionally, Moreover" | Just start the next sentence | | "seamless, robust, intuitive" | Be specific about what changed | | "Category: Description" bullets | Plain prose or simple bullets | | "This should improve..." | "Benchmarks show X% faster..." |

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: dnth
• Source: dnth/agent-skills
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.