Chapter 10 — Prompt Engineering

Category: Operations Reading time: 6 minutes

---

Prompt Chain Architecture

The system uses a sequential 4-prompt chain where each step’s output feeds the next. This is intentional — rather than sending the raw 150KB transcript to a single prompt, the chain progressively distills the content:

Raw merged content (~150KB)
    → Extract Signal (~20KB) — 85% reduction
    → Community Post (~15KB) — structured and polished
    → Compressed Post (~12KB) — 30-40% further reduction
    → Weekly Invite (~1.5KB) — single focused output

Each step has a clear, narrow job. This produces better results than a single “do everything” prompt.

Prompt 1: Extract Signal

Goal: Reduce noise, keep only high-value information.

Key instruction: “Extract only the information that would still be useful to a community member reading it one week later.”

This temporal framing is powerful — it automatically filters out time-sensitive reactions (“that’s cool!”, “lol”), greetings, and small talk while preserving tools, links, insights, and open questions.

Output structure:

• Shared Resources (title, URL, why it matters)
• Key Q&A (question, answer, synthesis)
• Key Insights (ideas, tips, mental models)
• Tools and Concepts (with context)
• Follow-Ups Worth Revisiting (open threads)

Rules that matter most:

• “Do not force content into a section if nothing valuable is present” — prevents padding
• “Rewrite messy chat language into clear professional language” — handles transcript artifacts

Prompt 2: Community Post

Goal: Transform extracted signal into a copy-paste-ready community post.

Critical constraint: Plain text output for Skool compatibility.

IMPORTANT: This post will be published on Skool, which does NOT render markdown.
Output PLAIN TEXT only. Do NOT use markdown syntax like # headers, ** bold **,
- bullet lists, or [links](url). Instead use emoji for section headers.

Emoji section headers:

• 📎 SHARED RESOURCES
• ❓ KEY Q&A
• 💡 KEY INSIGHTS
• 🛠️ TOOLS AND CONCEPTS MENTIONED
• 🔄 FOLLOW-UPS WORTH EXPLORING
• 📝 SUMMARY

The “omit any section that has little or no useful content” rule is important — not every call covers every category.

Prompt 3: Compress Post

Goal: Reduce length by 30-40% while preserving all high-value content.

This is the simplest prompt. The key instruction is what NOT to remove: “Do not remove important resources or key takeaways.” The LLM’s job is to cut redundancy, tighten language, and merge overlapping points.

The plain text constraint is reinforced: “Keep the same emoji-based section headers and plain text formatting as the input.”

Prompt 4: Weekly Invite

Goal: Generate a unique weekly call invite that references last week’s content.

This is the most constrained prompt. It defines a rigid 4-part structure:

1. Opening hook (LLM-generated) — playful, references specific content
2. Standard section (verbatim) — reproduced exactly, including emoji
3. Flavor bridge (LLM-generated) — references unresolved topics
4. Closing section (verbatim except date) — Zoom link and date

Tone rules:

• Playful, welcoming, energetic
• No negative call-outs
• No inside jokes that exclude non-attendees
• First names only when referencing open questions

The standard section and closing are provided verbatim in the prompt. The LLM must reproduce them exactly — this is enforced by explicit instruction: “must be reproduced EXACTLY (no modifications to wording or emoji).”

Model Configuration

All prompts use identical settings:

Setting	Value	Rationale
Model	Claude Sonnet 4.6	Best balance of quality and cost for content generation
Temperature	0.3	Consistent output while allowing fresh weekly invites
Max tokens	8192	Sufficient for even the longest community posts

Iterating on Prompts

To modify a prompt:

1. Edit the system message in the relevant LLM node (in n8n UI or in the workflow JSON)
2. If editing JSON, re-import and re-link credentials
3. Reprocess a test meeting to verify the changes
4. Compare old vs. new output

The sequential chain makes prompt iteration safe — changing Prompt 3 doesn’t affect Prompts 1 or 2. Each step can be tuned independently.

Common Prompt Issues

Markdown leaking into output: Add explicit “Do NOT use markdown” instructions. LLMs default to markdown formatting — you must actively suppress it.

Weekly invite modifying the standard section: Strengthen the verbatim instruction. Adding “EXACTLY” in caps and repeating it helps.

Extracted signal too verbose: Tighten the “one week later” framing or add explicit length guidance.

Missing sections in community post: This is usually correct behavior — the “omit empty sections” rule is working. Only investigate if sections with genuine content are being dropped.