Agent Academy Report Generator

Generate a comprehensive feedback analysis report for Agent Academy. The report combines data from Excel survey exports and GitHub issues, analyzes sentiment, generates charts, and produces a single styled PDF.

Folder Structure

The workspace folder should be organized as follows:

<workspace>/report/
├── data/            ← Excel files (.xlsx) with survey responses
├── badges/          ← Badge PNG images (downloaded from GitHub)
├── charts/          ← Generated chart PNG images (output)
├── markdown/        ← Generated markdown files (output)
└── pdf/             ← Generated PDF files (output)

If the user provides a workspace path, use it. Otherwise, ask for it.

Overview

The report generation has 6 phases:

1. Extract feedback from Excel files and GitHub issues
2. Analyze sentiment using keyword-based scoring
3. Generate charts with matplotlib
4. Build markdown files (management summary + detailed analysis)
5. Export a single PDF with cover page, management summary, and full analysis
6. Verify PDF formatting — visually inspect and fix any layout issues

Run the bundled Python scripts in order. Each script is self-contained and reads/writes to the folder structure above.

Phase 1: Extract Feedback

Run scripts/extract_feedback.py:

python3 <skill-path>/scripts/extract_feedback.py "<workspace-path>/report"

This script:

• Reads all .xlsx files from <workspace>/report/data/
• Maps filenames to human-readable course names
• Extracts feedback text, completion dates, and respondent names/emails
• Name normalization: If only an email address is available, it converts it to a display name — e.g. john.doe@example.com → "John Doe" (split on dots/underscores, title-cased). If neither a name nor email is found, the name defaults to "Anonymous".
• Skips trivial feedback entries (e.g. "no", "none", "n/a", "ok", single punctuation)
• For files without a feedback column (Operative, Recruit), it extracts completion dates only — their feedback comes from GitHub issues
• Outputs <workspace>/report/data/_extracted.json

GitHub Issues (Operative & Recruit)

After running the extract script, fetch feedback from GitHub issues. Do not use mcp_github_mcp_search_issues — the search API caps results at 1,000, which is insufficient for Recruit (1,390+ issues). Instead, use mcp_github_mcp_list_issues which supports full GraphQL cursor pagination:

Recruit issues:

owner: "microsoft", repo: "agent-academy", labels: ["recruit-completed"], state: "all", perPage: 100

Operative issues:

owner: "microsoft", repo: "agent-academy", labels: ["operative-completed"], state: "all", perPage: 100

Paginate through all pages using the endCursor from each response until hasNextPage is false. The API returns max 100 per page — Recruit requires ~14 pages, Operative ~4 pages. Do not stop early; collect every issue.

For each issue, parse the body for structured sections:

• ### Key Learnings and Takeaways
• ### Challenges Faced
• ### Final Thoughts

Combine these sections as the feedback text. The name field should be a plain string (the GitHub username) — if the API returns a user object like {"login": "username"}, extract just the login value.

Save the combined data (Excel + GitHub) to <workspace>/report/data/_all_feedback.json.

Each record should have: {course, date, month, sentiment, text, name, type, source, is_closed}.

• source: "excel" (set automatically by extract script), "github_completed" (badge submission issues), or "github_issue" (non-badge issues)
• is_closed: true/false — for GitHub records only; indicates whether the issue is closed (closed = badge awarded for completed issues)

Non-Badge GitHub Issues

After collecting badge submission issues, also fetch all other open issues that are NOT badge submissions. These are bug reports, questions, and general feedback that should be classified into the correct course.

Fetch issues without the completed labels:

owner: "microsoft", repo: "agent-academy", state: "all", perPage: 100

Filter out issues that already have recruit-completed or operative-completed labels. For each remaining issue, classify it to a course using this keyword map (match against issue title and labels, case-insensitive):

If no match is found, skip the issue. Set source: "github_issue" and type: "feedback" for these records. Use the issue title + body as the feedback text.

How Non-Badge Issues Are Used

Non-badge GitHub issues (source: "github_issue") are included in:

• Sentiment analysis (scored and themed)
• Course grades (affect positive/negative percentages)
• Top themes (feed into per-course theme detection)
• Representative quotes (can appear as selected quotes)
• Next steps (negative themes from these issues drive recommendations)

Non-badge GitHub issues are excluded from:

• Submission/completion charts (completions over time, cumulative, monthly, per-course)
• Monthly trend tables (submission counts)
• Total completions count in the management summary

This is because they are feedback and bug reports, not actual course completions.

Recruit & Operative Comparison Data

The combined _all_feedback.json will contain records with different source values. The build script uses these to generate a Submission Pipeline comparison table for Recruit and Operative:

Phase 2: Analyze Sentiment

Run scripts/analyze_sentiment.py:

python3 <skill-path>/scripts/analyze_sentiment.py "<workspace-path>/report"

This script:

• Reads <workspace>/report/data/_all_feedback.json
• Scores each feedback item using regex-based sentiment analysis (word-boundary \b matching, case-insensitive)
• Detects common themes and adds a themes array to each record (list of matched theme names, e.g. ["Hands-on & Practical Learning", "MCP & Agent Development Skills"])
• Computes per-course grades using weighted average (positive=10, neutral=6, negative=2)
• Outputs <workspace>/report/data/_analyzed.json with sentiment labels, themes, and course stats

Output Structure

The _analyzed.json file contains:

{
  "records": [ ... ],        // All records with added "sentiment" and "themes" fields
  "summary": { "total_records", "total_feedback", "positive", "neutral", "negative", "overall_grade" },
  "course_stats": {           // Per-course: total, positive, neutral, negative, grade, themes (with counts)
    "Recruit": { "total": 123, "positive": 100, ..., "themes": { "Hands-on & Practical Learning": 45, ... } }
  }
}

Sentiment Keywords

All patterns use \b word boundaries for precise matching (e.g. \bgreat\b matches "great" but not "greater").

Positive patterns include: great, excellent, amazing, awesome, fantastic, wonderful, love/loved, enjoy/enjoyed, helpful, useful, informative, well done, good job, thank/thanks, appreciate, impressive, learned, learning, eye-opening, recommend, looking forward, fun, excited, brilliant, clear, intuitive, practical, hands-on, well-structured, well-organized, good, nice, cool, neat, super, stellar, outstanding, perfect, phenomenal, beginner-friendly, easy to follow, straightforward, engaging, interesting, valuable, insightful, comprehensive, thorough, solid, smooth.

Positive emojis (also scored as positive): 👍 🎉 🙌 ❤️ 🔥 💯 ⭐ ✨ 😍 😊 🤩 💪

Negative patterns include: error(s), issue(s), problem(s), bug(s)/buggy, broke/broken/crash/crashed, fail/failed/failure, didn't work, not work, couldn't, stuck, struggling/struggle, missing, trouble, difficult/difficulties, cumbersome, challenge(s), troubleshoot, confusing/confused, frustrating/frustrated, slow, complex, unclear, improve/improvement, wish, lack/lacking, disappoint/disappointed.

Scoring rule: If negative matches > positive matches → negative. Else if positive > 0 → positive. Else if negative > 0 → negative. Else → neutral.

Theme Detection

The analyzer detects 10 predefined themes by matching regex keywords against feedback text:

A feedback item can match multiple themes. Themes are stored as r['themes'] = [...] on each record.

Phase 3: Generate Charts

Run scripts/generate_charts.py:

python3 <skill-path>/scripts/generate_charts.py "<workspace-path>/report"

Requires matplotlib and numpy. Install if needed: pip3 install matplotlib numpy.

This generates overview charts plus individual per-course charts in <workspace>/report/charts/ at 180 DPI.

Note: Submission/completion charts only count actual submissions (source: "excel" and source: "github_completed"). Non-badge GitHub issues (source: "github_issue") are excluded from these charts — they are only used for sentiment analysis.

Course Grouping & Colors

Charts group courses into 4 categories, each with a consistent color:

Display order (used in all charts): Recruit, Operative, Special Ops: YAML Specialist, Special Ops: Microsoft Copilot Studio MCP, Special Ops: Microsoft Learn Docs MCP Server, Special Ops: Power Platform CLI MCP Server, Cowork Collective: Badge Check, Cowork Collective: Compliance Packet, Cowork Collective: Out of Office.

Short names used on chart axes (where full names are too long):

Overview Charts

Per-Course Charts

For each course, a course_{slug}.png file is generated (e.g. course_recruit.png, course_operative.png, course_special_ops_yaml_specialist.png). Each shows:

• Weekly submissions as an area chart (primary axis)
• Cumulative total as a dashed line (secondary axis)

The slug is the course name lowercased with spaces → underscores, colons removed, and ampersands → "and" (e.g. "Cowork Collective: Badge Check" → course_cowork_collective_badge_check.png).

Color scheme (navy/teal theme matching cover): Blue (#3b82f6), Green (#22c55e), Red (#ef4444), Orange (#f59e0b), Purple (#a78bfa), Teal (#14b8a6). Charts use a consistent light background (#fafbfc), subtle grid (#e2e8f0), and navy (#111827) text.

Phase 4: Build Markdown

Run scripts/build_markdown.py:

python3 <skill-path>/scripts/build_markdown.py "<workspace-path>/report"

This generates three markdown files in <workspace>/report/markdown/:

management-summary.md

• Title: "Management Summary"
• Key metrics table (total completions, feedback count, courses, overall grade, sentiment percentages)
  • Total completions excludes source: "github_issue" records (only counts actual submissions)
• Highlights section (Recruit/Operative counts, highest/lowest rated courses)
• Chart references (one per section with narrative text)
• 4 main sections — Recruit, Operative, Special Ops, Cowork Collective:
  • Single-course sections (Recruit, Operative): spotlight content directly under the section heading
  • Multi-mission sections (Special Ops, Cowork Collective): section-level summary first (combined grade, response count, cross-mission themes, section-level next steps), then individual mission spotlights as sub-sections with short names (e.g. "YAML Specialist" instead of "Special Ops: YAML Specialist")
  • Each course/mission spotlight includes:
    • Grade, response count, positive/negative percentages
    • Grade explanation — contextual note based on score range:
      • ≥ 9.0: "Exceptional satisfaction — one of the top-rated courses."
      • ≥ 8.5: "Strong satisfaction with consistently positive feedback."
      • ≥ 8.0: "Good satisfaction, though some participants encountered challenges."
      • < 8.0: "Solid course, but a higher neutral/negative ratio pulls the grade down."
    • Top 3 themes with mention counts
    • Standout callout — appears when positive % ≥ 85%: highlights most-praised theme
    • Watch area callout — appears when negative % ≥ 15%: highlights top negative theme
    • Representative quotes — 10 total, split proportionally to the grade using formula: num_positive = round(10 × grade / 10). Grade 8.0/10 → 8 positive, 2 negative. Quotes are selected by closest to ideal length (positive: ~250 chars, negative: ~200 chars).
• Recommendations section (4 static recommendations: address setup challenges, expand Cowork Collective, scale Special Ops, continue monitoring)

Quote Cleaning

All quotes in the report go through a cleaning pipeline before display:

1. Newlines → spaces
2. Bold headers (**text**) removed
3. Markdown headers (# ...) removed
4. Checkbox lines (- [ ] ...) removed
5. Leading list dashes removed, inline dashes → periods
6. HTML entities decoded (' → ', & → &)
7. Collapsed whitespace
8. Truncated at sentence boundary — positive quotes target ~250 chars max, negative quotes ~200 chars max (falls back to word boundary + ... if no sentence break found above 80–100 chars)

feedback-analysis.md

• Title: "Detailed Feedback Analysis"
• Overall summary table (sentiment counts + percentages)
• Course overview table (responses, sentiment breakdown, grade per course)
• Recruit & Operative: Submission Pipeline — funnel-style comparison table showing GitHub badge submissions, badges awarded/pending with percentages, Excel form submissions, and other GitHub issues. Includes a note explaining that GitHub and Excel are independent channels.
• 4 main sections — Recruit, Operative, Special Ops, Cowork Collective:
  • Single-course sections (Recruit, Operative): analysis content directly under the ## Section heading (h2), with subsections as h3
  • Multi-mission sections (Special Ops, Cowork Collective):
    • Section-level summary (under h2): combined response count + grade, cross-mission themes (top 5 with positive/negative indicators), section-level recommendations (up to 5 next steps based on aggregated data across all missions)
    • Per-mission blocks (under h3): each mission gets its own analysis using the short name (e.g. "YAML Specialist", "Badge Check"), with sub-headings at h4 level
  • Each course/mission block includes:
    • Response counts and sentiment percentages
    • Submissions Over Time chart (per-course chart reference)
    • Monthly trend table showing submissions, sentiment breakdown, and grade per month
    • Top 5 themes with mention counts and positive/negative indicator
    • 10 representative quotes — balanced on grade using same formula as management summary (round(10 × grade / 10) positive, remainder negative); each cleaned by the quote pipeline (see above); positive truncated to ~200 chars, negative to ~180 chars
    • Next Steps — up to 5 prioritized, actionable recommendations generated algorithmically:

Next Steps Algorithm

Each course gets up to 5 next steps, generated in priority order:

1. Priority 1–2: Top negative themes — The top 2 themes with the most negative mentions, but only if each has ≥ 3 negative mentions. Uses a theme→advice mapping:

   Theme	Advice
   Setup & Configuration Challenges	Simplify setup — provide a pre-configured environment or improved quick-start guide...
   Errors & Technical Issues	Fix reported bugs — audit the most common failure points and add troubleshooting guidance.
   Improvement Suggestions	Act on improvement suggestions — review and prioritize the most-requested changes...
   Clear Instructions & Documentation	Improve documentation clarity — review step-by-step guides for completeness...
   VS Code Integration	Improve VS Code experience — ensure extensions and tooling work seamlessly...
   MCP & Agent Development Skills	Strengthen MCP/agent content — review labs where participants reported MCP-related issues...
   Valuable Learning Experience	Enhance learning depth — some participants felt the content could go deeper...
   Course Quality & Engagement	Refine course structure — address pacing and engagement issues...
   Hands-on & Practical Learning	Improve hands-on exercises — ensure lab environments are reliable...
   Real-World Applicability	Add more real-world examples — participants want to see how skills apply to production...

2. Priority 3: Grade-based advice — If grade < 8.0: "Boost course quality — current grade is below the 8.0 target." If grade < 8.5: "Target grade improvement — small refinements could push this above 8.5."

3. Priority 4: Leverage strengths — Highlights the top positive theme: "Double down on [theme] — most praised aspect ([count] positive mentions)."

4. Priority 5: Engagement — If < 50 responses: "Increase feedback volume — consider adding in-course feedback prompts." Otherwise: "Continue monitoring — maintain the feedback loop and review sentiment trends monthly."

feedback.md

• Curated positive-only feedback (excludes error-mentioning entries)
• Max 5 quotes per course (~45 total), selected for ideal length (~250 chars)
• Each item as a blockquote: > 👍 "text" — Name (Course)

Phase 5: Export PDF

Run scripts/export_pdf.py:

python3 <skill-path>/scripts/export_pdf.py "<workspace-path>/report" "<report-title>"

Example: python3 scripts/export_pdf.py "/path/to/workspace/report" "Agent Academy - Report April 2026"

This script:

1. Downloads badge images from https://raw.githubusercontent.com/microsoft/agent-academy/main/docs/public/images/ to <workspace>/report/badges/ (if not already present)
2. Builds a single HTML file with:
   • Cover page: Dark navy gradient background with teal accent, grid overlay, Agent Academy logo, report title, all badge images
     • Title splitting: The report title is split on " - " — the part before becomes the main title (large teal text), the part after becomes the subtitle (smaller slate text). E.g. "Agent Academy - Report April 2026" → main: "Agent Academy", subtitle: "Report April 2026".
     • Hardcoded elements: "Mission Report" tag (top-right classification label), "Microsoft Copilot Studio" pre-title above the main title, "Earned Badges" label on the badge shelf, "microsoft/agent-academy • Comprehensive Feedback & Completion Analysis" footer text.
   • Management summary: Converted from markdown with charts embedded as base64 images (H1 title stripped — the HTML provides its own)
   • Detailed analysis: Concise per-course analysis with monthly trends, themes, and representative quotes (H1 title stripped)
3. Exports to PDF using headless Microsoft Edge:

   "/Applications/Microsoft Edge.app/Contents/MacOS/Microsoft Edge" \
     --headless --disable-gpu --no-sandbox \
     --print-to-pdf="<workspace>/report/pdf/<title>.pdf" \
     --no-pdf-header-footer \
     "file://<workspace>/report/_report.html"

4. Cleans up the temporary HTML file
5. On macOS, automatically removes the com.apple.quarantine extended attribute so the PDF opens without a Gatekeeper warning

The output PDF lands in <workspace>/report/pdf/.

Page Break Rules

The HTML includes CSS rules to prevent content from splitting across pages:

• Charts + their descriptions are wrapped in .chart-block divs with page-break-inside: avoid
• Course spotlight sections (h3 + content) are wrapped in .course-spotlight divs
• Headings (h2, h3, h4) use page-break-after: avoid
• Tables, blockquotes, and images all have page-break-inside: avoid

Badge Images

Downloaded from microsoft/agent-academy repo at docs/public/images/:

• logo.png — Cover page logo
• mcs-agent-academy-recruit-badge.png, mcs-agent-academy-operative-badge.png, mcs-agent-academy-commander-badge.png — Level badges
• YAML_Specialist_Badge.png, Academy_LearnMCP_Badge.png, CommandLine_Badge.png — Special Ops badges
• BadgeBandit-badge.png, AuditAce-badge.png, Vacay-badge.png, MCP_Joker_Badge.png — Cowork Collective badges

Edge Fallback

The script checks these browser paths in order and uses the first one found:

1. /Applications/Microsoft Edge.app/Contents/MacOS/Microsoft Edge (macOS)
2. /usr/bin/microsoft-edge (Linux)
3. /Applications/Google Chrome.app/Contents/MacOS/Google Chrome (macOS)
4. /usr/bin/google-chrome (Linux)

Phase 6: Verify PDF Formatting

After generating the PDF, always open and visually inspect it to check for formatting issues. Use the view_image tool or open the PDF to review each page.

Verification Checklist

Check the following and fix any issues found:

Common Formatting Issues & Fixes

Chart too large / stretched: The embed_md_images() function in export_pdf.py controls image sizing. It matches filenames to determine width:

• sentiment_pie → max-width:45% (small donut chart)
• course_* → max-width:85% (per-course charts)
• Everything else → max-width:100% (overview charts)

If a chart looks wrong, adjust the SMALL_CHARTS or MEDIUM_CHARTS lists and the corresponding max-width percentages.

Heading appears alone at page bottom: This means page-break-after: avoid is not being respected (can happen with large content blocks). Wrap the heading + its following content in a <div class="keep-together"> block in md_to_html().

Content runs together without spacing: The markdown-to-HTML converter treats \n\n as paragraph breaks (<p>). If content runs together, add an extra blank line in the markdown output (build_markdown.py). Use proper ### headings instead of **bold text:** for visually distinct sections.

h2 doesn't start on new page: Only works in @media print. The :first-of-type selector prevents the first h2 in each .content div from page-breaking. If ALL h2s page-break when they shouldn't, check that the CSS selector is scoped correctly.

Re-export After Fixes

If any formatting issues are found and fixed in the Python scripts, re-run the affected steps:

# If build_markdown.py was changed:
python3 <skill-path>/scripts/build_markdown.py "<workspace>/report"

# Always re-export PDF after any fix:
python3 <skill-path>/scripts/export_pdf.py "<workspace>/report" "<report-title>"

Then verify again until all checks pass.

Excel filenames map to display names as follows:

Note: filenames may contain non-breaking spaces (\xa0) — normalize these to regular spaces before matching.

Running the Full Pipeline

To generate a complete report in one go:

# 1. Extract from Excel
python3 <skill-path>/scripts/extract_feedback.py "<workspace>/report"

# 2. Fetch GitHub issues via MCP tools:
#    a. Badge submissions: recruit-completed + operative-completed labels (paginate ALL pages)
#    b. Non-badge issues: all other issues, classified to courses by keyword
#    c. Add source/is_closed fields, combine with Excel → _all_feedback.json

# 3. Analyze sentiment
python3 <skill-path>/scripts/analyze_sentiment.py "<workspace>/report"

# 4. Generate charts
python3 <skill-path>/scripts/generate_charts.py "<workspace>/report"

# 5. Build markdown
python3 <skill-path>/scripts/build_markdown.py "<workspace>/report"

# 6. Export PDF
python3 <skill-path>/scripts/export_pdf.py "<workspace>/report" "Agent Academy - Report April 2026"

# 7. Verify PDF formatting
#    Open the PDF and check: cover page, chart sizes, page breaks, spacing
#    Fix any issues in export_pdf.py or build_markdown.py, then re-export

Step 2 requires the GitHub MCP tools and should be done interactively by the agent between steps 1 and 3.