[prompt-analysis] Copilot PR Prompt Analysis - 2026-06-24

[github-actions[bot]]

🤖 Copilot PR Prompt Pattern Analysis — 2026-06-24

Summary

Analysis Period: Last 30 days
Total PRs: 1,000 | Merged: 801 (80.1%) | Closed: 195 (19.9%) | Open: 4 (0.4%)

Prompt Categories and Success Rates

Category	Total	Merged	Closed	Success Rate
Bug Fix	804	640	160	80.0%
Testing	132	108	24	81.8%
Feature	49	40	9	81.6%
Update	7	6	1	85.7%
Documentation	4	3	1	75.0%
Refactor	3	3	0	100%

Bug fix PRs dominate (80.4% of all PRs), reflecting the repository's focus on iterative agentic workflow refinement.

Prompt Analysis

✅ Successful Prompt Patterns

Common characteristics in merged PRs:

• Average prompt length: 186.9 words
• Optimal length bucket: 201–400 words → 84.9% merge rate
• Most correlated keywords: enforce, experiment, explicit, cache, surface, resolve, replace, emit, guidance
• Code references present: 97.3% of merged PRs

Example successful prompts

1. PR [review] Migrate set_issue_type safe output from GraphQL to single REST issues.update call #41284 — Migrate set_issue_type safe output from GraphQL to single REST call → Merged
   Body: Migrate set_issue_type safe output handler: GraphQL → REST issues.update — Summary — Migra...

2. PR feat: omit sudo from generated lock.yml when network-isolation is enabled; bump firewall to v0.27.10 and mcpg to v0.3.30 #41269 — feat: omit sudo from generated lock.yml when network-isolation is enabled → Merged
   Body: When sandbox.agent.network-isolation: true is set, AWF uses container-based egress (not host-iptab...

3. PR fix: UpdateContainerPins wipes containers section on every gh aw update run #41262 — fix: UpdateContainerPins wipes containers section on every gh aw update → Merged
   Body: gh aw update was deleting the entire "containers" section from actions-lock.json on every run....

❌ Unsuccessful Prompt Patterns

Common characteristics in closed PRs:

• Average prompt length: 187.5 words (nearly identical to merged — length alone is not a predictor)
• Most correlated keywords: tool-cache, findings, gh-proxy, claude, proxy, fail, upgrade, wip
• WIP-labeled PRs: 37 total, only 14 merged (37.8%) — less than half the baseline rate
• Very long prompts (400+ words): 71.8% merge rate vs 84.9% for 201–400 words

Example unsuccessful prompts

1. PR Add repo field to link_sub_issue tool schema for cross-repo support #41236 — Add repo field to link_sub_issue tool schema for cross-repo support → Closed
   Body: link_sub_issue already had full cross-repo plumbing (target-repo/allowed-repos in Go config)...

2. PR Add GH_AW_POLICY governance controls for safe-outputs MCP startup #41142 — Add GH_AW_POLICY governance controls for safe-outputs MCP startup → Closed
   Body: This change introduces policy-scoped governance flags for safe-outputs (GH_AW_POLICY_*) and wires...

3. PR chore(deps): bump charmbracelet/x/exp/golden to v0.0.0-20260622092256-25656177ba8e #41232 — chore(deps): bump charmbracelet/x/exp/golden (dependency update) → Closed
   Body: Advances the charmbracelet/x/exp/golden pseudo-version pointer to the June 22 monorepo commit...

Key Insights

• 🎯 WIP labels are a strong failure signal: PRs titled [WIP] or wip merge at only 37.8%, less than half the overall 80.1% rate. Avoid submitting WIP prompts as real requests.
• 📏 Sweet spot is 201–400 words: That bucket achieves 84.9% vs 71.8% for prompts over 400 words. Longer is not better — over-specification correlates with lower success.
• 🔍 Specificity matters at the margin: 97.3% of merged PRs contain code references (file paths, symbols, extensions) vs 96.4% for closed — always cite the relevant symbol or file.
• 🚫 Proxy/infrastructure-path prompts struggle: Keywords proxy, tool-cache, gh-proxy, claude cluster in closed PRs, suggesting infrastructure-routing changes are harder to land and may need upstream coordination.

Recommendations

1. DO write prompts that name the specific function, file, or behavior to change — code references in the body are nearly universal in merged PRs.
2. DO target 201–400 words for the PR body — provide enough context to be unambiguous, but avoid padding.
3. DO use action verbs like enforce, resolve, replace, emit, surface — these correlate with landed work.
4. AVOID [WIP] labels unless the PR is genuinely not ready for review; it cuts merge probability in half.
5. AVOID infrastructure-proxy topics (Claude routing, tool-cache, gh-proxy paths) without prior upstream alignment — these close more often than they land.

Historical Trends

Date	PRs	Success Rate	Top Category
2026-06-18	1,000	79.2%	bug_fix
2026-06-19	1,000	79.5%	feature
2026-06-20	1,000	79.2%	testing
2026-06-21	1,000	79.5%	bug_fix
2026-06-22	1,000	79.7%	bug_fix
2026-06-23	1,000	79.8%	bug_fix
2026-06-24	1,000	80.1%	bug_fix

Trend: Success rate climbed from 79.2% → 80.1% over the past week (+0.9 pp), a slow but consistent improvement. Bug fix remains the dominant category every day.

References:

• §28129824305

Generated by 🔍 Copilot PR Prompt Pattern Analysis · 40.5 AIC · ⌖ 32.3 AIC · ⊞ 7.6K · ◷

• expires on Jun 25, 2026, 1:19 PM UTC-08:00