Document Generation via AgentCore Runtime

Overview

This skill generates professional documents by calling an AWS Bedrock AgentCore Runtime agent. The remote agent uses Claude + Code Interpreter to produce high-quality files — no local office software required.

Supported Document Types

--type pptx

--type docx

--type pdf

--type xlsx

--type frontend-design

Prerequisites

1. AWS credentials configured (

   ~/.aws/credentials

   , env vars, or IAM role)
2. AgentCore Runtime deployed with document generation agent
3. Python 3.10+ with

   boto3

   installed

Required IAM Permissions

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": "bedrock-agentcore:InvokeAgentRuntime",
      "Resource": "<your-runtime-arn>"
    }
  ]
}

Environment Variables

RUNTIME_ARN

AWS_REGION

ap-northeast-1

The script also auto-loads

.env

~/codes/document-generation-mcp/.env

Quick Start

# Generate a PowerPoint
python scripts/docgen.py --type pptx \
  --prompt "Create a 10-slide presentation about cloud computing" \
  --output presentation.pptx

# Generate a Word document
python scripts/docgen.py --type docx \
  --prompt "Write a technical design document about microservices" \
  --output design.docx

# Generate from a prompt file (for long prompts)
python scripts/docgen.py --type pptx \
  --prompt-file my-prompt.txt \
  --output slides.pptx

# Large PPT (15+ slides) — auto-splits into parallel batches
python scripts/docgen.py --type pptx \
  --prompt-file 22-slide-prompt.txt \
  --output big-deck.pptx

# Force single-batch mode (disable auto-split)
python scripts/docgen.py --type pptx \
  --prompt-file prompt.txt \
  --output output.pptx \
  --no-split --timeout 1800

Python API

from scripts.docgen import generate

result = generate(
    skill_type="pptx",
    prompt="Create a 12-slide deck about AI trends in 2026",
    output_path="output/ai-trends.pptx",
)

if result["success"]:
    print(f"Saved: {result['path']} ({result['size']} bytes, {result['elapsed']:.0f}s)")
else:
    print(f"Error: {result['error']}")

How Claude Code Should Use This Skill

When a user asks to generate a document, follow this workflow:

Step 1: Determine the Mode

• Mode A (Article -> Document): User has an existing article/markdown and wants it converted
• Mode B (Topic -> Document): User provides a topic and wants a document from scratch

Step 2: Construct the Prompt

The prompt quality determines the output quality. Follow these guidelines:

PPTX Prompt Template

Create a [N]-slide professional presentation about [topic].
Use [color scheme] with [text color] on [background color].

IMPORTANT: Every slide MUST include speaker notes. Add speaker notes using
the python-pptx notes_slide feature: slide.notes_slide.notes_text_frame.text = "notes content".
Speaker notes should be 80-150 words per slide, written in conversational style,
explaining the key talking points. Do NOT just repeat the bullet points.

Slide 1: Title - [title], [subtitle]
Slide 2: [Section] - [key point 1], [key point 2], [key point 3]
Slide 3: [Section] - [data table or key facts]
...
Slide N: Thank You - [closing info]

PPTX Rules:

• Each slide description should be under 50 words (keywords and data, not paragraphs)
• Total slides should be <= 22 (more affects generation quality)
• Use "table with N rows" for data tables
• Always include speaker notes instructions
• Specify color scheme explicitly

DOCX Prompt Template

Create a professional [document type] about [topic].
Include sections: [section list].
Section 1: [Title] - [content points]
...

XLSX Prompt Template

Create an Excel spreadsheet for [purpose].
Sheet 1: [name] - columns: [col1], [col2], ...
Sheet 2: [name] - [description]

Step 3: Save the Prompt (Optional)

For debugging and reuse, save the prompt to a file:

# Save to prompt file
cat > prompt.txt << 'EOF'
[your prompt here]
EOF

Step 4: Generate the Document

python <skill-path>/scripts/docgen.py \
  --type pptx \
  --prompt-file prompt.txt \
  --output output/presentation.pptx

Expected generation times:

• HTML: 30s - 1min
• PPTX (<=10 slides): 2-5 min
• PPTX (10-22 slides): 3-15 min
• DOCX/PDF: 2-5 min
• XLSX: 2-4 min

Complex presentations (15+ slides with detailed content) can take 10-25 minutes. The default timeout is 1200s (20 min). Use

--timeout 1800

Step 5: Verify Output

Check that the file was generated and report the result:

ls -la output/presentation.pptx

Command Line Options

--type

--prompt

--prompt-file

--output

-o

--region

--runtime-arn

--timeout

--no-split

--max-slides-per-batch

Tips for Better Results

1. Be specific about colors: Always provide hex codes, not just color names
2. Keep slide content concise: Keywords and data points, not full sentences
3. Speaker notes matter: Include the speaker notes instruction for PPTX — it dramatically improves presentation quality
4. Chinese content: Specify "All slide titles and content MUST be in Chinese" if needed
5. Data tables: Describe as "table: col1, col2, col3" with row data
6. Complex prompts: Save to a file with

   --prompt-file

   instead of inline

   --prompt

7. Respect the 20-call budget: The server agent has a hard limit of 20 tool calls (code executions). Keep each slide description under 50 words so the agent can generate everything in 1-2 code runs
8. Auto-split threshold: For 15+ slides, the script auto-splits into ≤11-slide batches. This is more reliable than generating 22 slides in one shot

Troubleshooting

Timeout Errors

The script now auto-estimates timeout based on complexity. If you still hit timeouts:

# Explicit timeout override
python scripts/docgen.py --type pptx --prompt-file prompt.txt --output out.pptx --timeout 2400

# Better: let auto-split handle it (15+ slides splits into parallel batches)
python scripts/docgen.py --type pptx --prompt-file prompt.txt --output out.pptx

The script retries up to 3 times on transient errors (timeout, connection reset) with exponential backoff.

RUNTIME_ARN Not Configured

Set the environment variable or pass via CLI:

export RUNTIME_ARN="arn:aws:bedrock-agentcore:ap-northeast-1:123456789:runtime/your-agent-id"
python scripts/docgen.py --type pptx --prompt "..." --output out.pptx

Or create a

.env

~/codes/document-generation-mcp/.env

AWS_REGION=ap-northeast-1
RUNTIME_ARN=arn:aws:bedrock-agentcore:ap-northeast-1:123456789:runtime/your-agent-id

Empty or Failed Response

The AgentCore agent has a 20 tool call budget (enforced by MaxToolCallsHook on the server). If the prompt is too complex, the agent runs out of calls before producing the file. Symptoms:

• Response with no

  file_base64

• Error: "Code Interpreter did not produce an output file"

Solutions:

• Reduce slide count (keep ≤ 11 per batch, the auto-split handles this)
• Simplify per-slide content (use keywords, not paragraphs)
• The agent must: install deps (1 call) + generate file (1 call) + output base64 (1 call) = 3 minimum calls, leaving 17 for error recovery

Understanding Server Constraints

The remote agent operates under these constraints:

• Model: Claude Opus 4.6 (cross-region inference profile)
• Tool call limit: 20 code executions maximum (warning at 19, hard-stop at 21)
• Conversation window: SlidingWindowConversationManager (window=20, per_turn=True) — old messages get trimmed
• Base64 capture: A hook captures file output before conversation trimming
• Retry on server: The agent retries transient Bedrock errors up to 3 times internally

Architecture

Current Mode (Synchronous)

User / Claude Code
       |
       | python scripts/docgen.py --type pptx --prompt "..."
       |
       v
  boto3 invoke_agent_runtime()  ←── retry with backoff on transient errors
       |
       v
  AWS Bedrock AgentCore Runtime
       |
       | Strands Agent + Claude Opus 4.6 + Code Interpreter
       | MaxToolCallsHook (20 calls) + Base64CaptureHook
       |
       v
  Generated file (base64 encoded in response)
       |
       v
  Decoded and saved to --output path

Auto-Split Mode (for 15+ slide PPTX)

User / Claude Code
       |
       | python scripts/docgen.py --type pptx --prompt-file big.txt -o deck.pptx
       |
       v
  _count_slides_in_prompt() → 22 slides detected
       |
       v
  _split_pptx_prompt() → [Batch 1: slides 1-11, Batch 2: slides 12-22]
       |
       v
  ThreadPoolExecutor (parallel)
       ├── generate("pptx", batch1) → Part1.pptx
       └── generate("pptx", batch2) → Part2.pptx
       |
       v
  _merge_pptx_files([Part1.pptx, Part2.pptx]) → deck.pptx

Production Mode (Async — for team/service deployments)

For production use with Amazon Quick Suite, see the full async architecture in

sample-amazon-quick-suite-knowledge-hub/docs/use-cases/document-generation-mcp-agentcore-runtime/

This uses: submit_job Lambda → fire-and-forget → Step Function polling → direct-persist to S3 → CloudFront download URL. No client-side timeout concerns.