Part 2 · Synthesis & Checklist

7. Synthesis — how the six elements interact in one agent

7.1 Flow of one request (general form)

Example request: “Analyze this data, turn the results into a report, and email it to the owner.”

[Instructions] Always-loaded rules: calculate accurately / send only after confirmation / follow skills for format
   ↓
[Knowledge]    Retrieve connected knowledge (data files and documents)
   ↓
[Skill]        Load analysis skill → "calculate accurately, organize as tables"
[Tool]         Aggregate the full dataset with code execution (not RAG, Part 1 9.4)
   ↓
[Skill]        Load writing skill → apply format and design rules (single source of truth)
[Tool]         Generate report → upload to storage → draft email + attach link
   ↓
[Instructions] "Confirm recipients + preview before sending" → user approval gate
[Tool]         Send email after approval
   ↓
[Memory]       Conversation is recorded in session memory (used for follow-up questions); persistent output storage is the file store

7.2 Full composition at a glance

7.3 Recommended build order

1. Start with instructions — role, guardrails, and HITL in plain language (short).
2. Connect knowledge — upload Excel and Word to SharePoint and connect Knowledge.
3. Create one core skill — start with the most important one as a single file.
4. Keep tools minimal — code interpreter + only the MCP tools the scenario truly needs.
5. Expand skills — bundle resources (templates) where quality is needed.
6. Split into connected agents when needed.

FAQ response: “Shouldn’t we have everything ready from the start?” → No (Part 1, 1.4). Start with instructions + one skill + minimum tools, run it, and add tools, skills, or agents only where you get blocked. Tools and exception rules are usually best added after encountering the problem.

Appendix. Build checklist & common pitfalls

Build checklist

• Are the instructions short and focused on “what”? (No implementation details.)
• Are details such as colors and procedures in skills, not instructions? (Single source of truth.)
• Is there a confirmation gate for irreversible actions such as sending?
• Does each skill have name and description in YAML?
• Does the description contain clear trigger keywords?
• Is SKILL.md at the zip root? (Zip the folder contents.)
• Does the skill name use only lowercase letters, numbers, and hyphens? (No underscores or uppercase.)
• Are core design rules and requirements written directly in the SKILL.md body? (Resource files are not automatically injected.)
• Are tools minimal? (No overlapping tools, and no tools for what code can do.)
• Are numbers separated into Excel, and tone/format into Word?

Common pitfalls