[prompt-analysis] Copilot PR Prompt Analysis - June 20, 2026

[github-actions[bot]]

🤖 Copilot PR Prompt Pattern Analysis — June 20, 2026

Summary

Analysis Period: Jun 1 – Jun 20, 2026 (last 30 days)
Total PRs: 1,000 | Merged: 792 (79.2%) | Closed: 204 (20.4%) | Open: 4 (0.4%)

Overall merge rate is stable at ~79%, consistent with the prior 7-day range (78.6%–79.7%).

---

Prompt Categories and Success Rates

Category	Total	Merged	Closed	Success Rate
Refactoring	18	17	1	94.4% ✅
Documentation	34	28	6	82.4% ✅
Testing	708	569	139	80.4%
Feature Addition	30	23	7	76.7%
Bug Fix	200	149	51	74.5% ⚠️

Note: Category classification is keyword-based. "Testing" dominates because automated CI fix prompts trigger on fail/test/integration keywords.

---

Prompt Analysis

✅ Successful Prompt Patterns

Common characteristics in merged PRs:

• Average prompt length: 185.9 words
• Code references (backticks): 97.1% of merged PRs include inline code references
• Keywords correlated with success: permission, allow, allowlist, normalized, expanded, completing
• Prompts clearly state a specific before/after: "X was doing Y, this changes it to Z"

Example successful prompts:

3 merged PR examples

1. PR Improve max-tool-denials report with recent shell call context #40506 — Improve max-tool-denials report with recent shell call context

   "The max-tool-denials report showed the last five calls as bare tool names (for example, repeated bash), which made the denial chain hard to diagnose. This change adds lightweight command context so the report shows what was attempted, not just which tool was invoked." → Merged

2. PR Reduce Daily Safe Output Integrator tool-denial guardrail trips #40503 — Reduce Daily Safe Output Integrator tool-denial guardrail trips

   "Daily Safe Output Integrator was failing early after hitting Copilot SDK's tool-denial threshold (guard.tool_denials_exceeded, 5/5), preventing the workflow from completing its integration pass." → Merged

3. PR /help: use ### heading, link commands to source workflows #40500 — /help: use ### heading, link commands to source workflows

   "The builtin /help response used a level-2 heading and rendered commands as plain code spans with no way to navigate to the underlying workflow." → Merged

---

❌ Unsuccessful Prompt Patterns

Common characteristics in closed PRs:

• Average prompt length: 195.5 words (slightly longer — verbosity alone doesn't help)
• Keywords correlated with closure: progress, started, parameters, queue, assign
• [WIP] titled PRs: 27 closed vs 16 merged — 37.2% merge rate (vs 80.7% for non-WIP)
• Boilerplate-body PRs: 18 closed PRs contained only "Thanks for asking me to work on this..." with no actual description — 100% closed

Example unsuccessful prompts:

3 closed PR examples

1. PR fix(assign-to-agent): remove invalid requiresOneOf constraint, add GraphQL query logging #40488 — fix(assign-to-agent): remove invalid requiresOneOf constraint

   "The assign_to_agent validator rejected messages without issue_number or pull_number, breaking triggering-context auto-resolution..." — technically specific but the change was superseded → Closed

2. PR [WIP] Fix failing GitHub Actions job Integration: Workflow Features #40007 — [WIP] Fix failing GitHub Actions job Integration: Workflow Features

   "Thanks for asking me to work on this. I will get started on it and keep this PR's description up to date as I form a plan and make progress." → Closed (never completed)

3. PR [WIP] Extract embedded issue template to shared partial #36743 — [WIP] Extract embedded issue template to shared partial

   "Thanks for asking me to work on this. I will get started on it and keep this PR's description up to date as I form a plan and make progress." → Closed (never completed)

---

Key Insights

• Refactoring prompts outperform all categories at 94.4% — when scope is tightly bounded ("simplify X by doing Y"), implementation complexity is lower and review is faster.
• [WIP] PRs are 2× more likely to be closed — prompts that produce placeholder bodies indicate the agent could not complete the task. These represent 37.2% merge rate vs 80.7% overall.
• Prompt length is not a success predictor — closed PRs average 195.5 words vs 186 for merged. The key differentiator is specificity of the problem statement, not length. Include a concrete failing scenario or exact API/symbol name.
• Code references are near-universal — 97.1% of merged PRs vs 93.6% of closed PRs use backtick code references. Not having them at all is a weak signal for failure.

---

Recommendations

Based on today's analysis:

1. DO: State the broken behavior precisely — reference the exact function, flag, or YAML key involved. E.g., "--timeout flag in DownloadWorkflowLogs only checks elapsed time at loop top" rather than "timeout doesn't work".
2. DO: Include a brief root cause in the prompt body. Prompts with "because..." or "this breaks..." context tend to produce narrower, more reviewable diffs.
3. DO: Prefer refactoring prompts over broad feature prompts — they have the highest merge rate (94.4%) when scope is explicit.
4. AVOID: Submitting [WIP] PRs with boilerplate bodies — these are almost never merged and should be closed early or re-prompted.
5. AVOID: Prompts with vague scope like "improve X" or "refactor the whole Y module" — they generate larger, harder-to-review diffs.

---

Historical Trends (Last 7 Days)

Date	PRs	Merged	Closed	Success Rate
Jun 20	1000	792	204	79.2%
Jun 19	1000	795	201	79.5%
Jun 18	1000	792	196	79.2%
Jun 16	1000	797	196	79.7%
Jun 14	1000	791	202	79.1%
Jun 13	1000	792	204	79.2%
Jun 12	1000	786	206	78.6%

Trend: Stable. Success rate has held in the 78.6%–79.7% band for the past 9 days. No significant drift detected. Week-3 (Jun 16–20) shows a slight uptick in Jun 16 (79.7%) followed by regression on Jun 18–20 — the Jun 18–20 dip correlates with increased [WIP] activity (27 open WIP issues vs 16 merged).

---

References: §27883699237

Generated by 🔍 Copilot PR Prompt Pattern Analysis · 64.2 AIC · ⌖ 7.85 AIC · ⊞ 8K · ◷

• expires on Jun 21, 2026, 1:06 PM UTC-08:00

[pelikhan]

/q optimize

[github-actions[bot]]

🔧 Pay attention, 007! Q is preparing your gadgets for this discussion comment...