Prompt Engineering: RTITCO + JSON Handoffs for Consistent Output

TL;DR

• Structure wins: Use RTITCO (Role, Task, Input, Tools, Constraints, Output) to turn vague asks into reliable outputs.

• JSON-first: Plan in JSON first, then produce the final asset from that plan (two-pass).

• Automation-ready: Machine-readable outputs plug directly into Notion, Slack, Gmail (draft), n8n/Zapier/Make.

• Fewer rewrites: Clarify first, be blunt, lock length/format, validate against “Success is…”.

• Scale: Treat prompts + schemas like product configs (version, reuse, govern).

Why Structured Prompting + JSON?

Free-form prompts = free-form answers. That’s fine for ideation, painful for production.
Structured prompting gives the model an explicit contract; JSON outputs make results parsable, testable, repeatable. You’ll see:

• Higher prompt adherence (tone/length/required sections)

• Consistent formatting across runs/models

• Easy handoff to automations (no brittle regex)

• Clear audit trail (prompt → plan JSON → final)

  
  

The RTITCO Blueprint (Copy-Paste)

Use this as your default skeleton for any high-stakes prompt.

Role: Act as a [senior role] for [company/industry].

Task: [exact deliverable and intent]. If any required info is missing,
ask clarifying questions before providing an Output.

Input:
- [facts, source text, links, brand voice rules, data tables]
- Resolve conflicts by asking first.

Tools (manifest):
- Files: [Brand-Voice.pdf — tone rules], [Offers.xlsx — prices/codes]
- Integrations: [Notion read-only], [Gmail draft-only], [Slack post-only]
- Forbidden: [payments, PII, bulk sends]

Constraints:
- Tone: [confident, concise, on-brand]. Be blunt; skip filler.
- Length/Format: [e.g., ≤200 words, exactly 5 bullets, Markdown table].
- Evidence: use only Input; no outside facts.

Output:
- Success is: [acceptance criteria].
- Output format: [JSON/table/plain text]. Output only that format

The JSON-First, Two-Pass Workflow

Pass 1 — Plan (JSON only)

Ask for a structured plan as valid JSON you can review or validate.

Prompt (example):

Create a JSON plan for a 1-page product sheet.
Keys: "headline", "subhead", "proof_points" (3 items), "cta".
Use only Brand-Voice.pdf and Offers.xlsx; ask clarifying questions if facts are missing.
Return JSON only

Sample plan JSON:

{
  "headline": "Faster Onboarding, Fewer Tickets",
  "subhead": "Ship product education that reduces support volume.",
  "proof_points": [
    "30% fewer new-user tickets in 60 days",
    "Tutorials mapped to top 10 support themes",
    "Real-time progress tracking in admin console"
  ],
  "cta": "Start your 14-day trial"
}

Pass 2 — Produce (from approved plan)

Feed the plan back. Enforce tone/length and “Success is…” criteria.

Prompt (example):

Using the approved JSON plan, write the final copy.
Follow Brand-Voice.pdf. Use real prices from Offers.xlsx.
Output plain text only.
[PASTE PLAN JSON]

Why it works: The model “thinks” structurally first, then “writes” to spec. You get coverage of all required elements, consistent format, and a machine-readable artifact for QA.

The Tools Manifest (Make Capabilities Explicit)

Put this “runbook” in your prompt so the model knows what it may use and what’s forbidden.

Bullet form

• Files

  • Brand-Voice.pdf — tone/phrases to enforce

  • Offers.xlsx — product names, prices, promo codes (no guessing)

• Integrations

  • Notion (read-only) — cite page titles

  • Gmail (draft-only) — create drafts in Drafts/AI; do not send

  • Slack (post-only) — post preview to #content-review

• Forbidden — payments, PII, bulk sends

JSON form

{
  "files": [
    {"name": "Brand-Voice.pdf", "use": "Tone rules & vocabulary."},
    {"name": "Offers.xlsx", "use": "Authoritative prices & codes."}
  ],
  "integrations": [
    {"name": "Notion", "mode": "read", "scope": "kb/*"},
    {"name": "Gmail", "mode": "draft-only", "folder": "Drafts/AI"},
    {"name": "Slack", "mode": "post", "channel": "#content-review"}
  ],
  "forbidden": ["payments", "PII", "bulk_sends"]
}

Real-World Recipes

1) Weekly Promo Email

• Mode: Thinking

• Flow: Plan JSON (subject, preheader, 3 bullets, CTA) → final copy → n8n renders HTML, creates Gmail draft, posts Slack preview.

• Success is: Subject ≤ 60 chars; 3 bullets use real prices; clear deadline.

2) Blog → Social → Email Cascade

• Mode: Thinking → Extended

• Flow: Research brief (citations) → outline (JSON) → long-form → auto-extract meta/title/CTAs/social snippets via schema keys.

• Tip: Keep one master plan JSON; generate multiple assets from it.

3) Paid Media Variations

• Mode: Instant for ideation, Thinking for finals

• Flow: Rapid 20 headlines in Instant → choose 5 → Thinking refines copy to exact character limits → image prompts derived from a creative brief JSON.

4) Data Extraction / Triage

• Mode: Thinking

• Flow: Input text → JSON extract (priority, issue_type, due_date) → automation routes to Notion/Asana/Slack.

• Guardrail: Validate against a JSON Schema before actions.

Guardrails That Actually Reduce Rewrites

• Clarify first: “Ask me clarifying questions before providing an Output.”

• Be blunt: Removes hedging/filler; increases review speed.

• Length locks: “Exactly 5 bullets”, “≤120 words”.

• Style anchors: Paste 2–3 on-brand lines and say “match this style.”

• Self-check: “Validate against ‘Success is…’; fix before output.”

• Schema checks: If JSON → validate; retry on failure.

ChatGPT Agents vs n8n Agents (Quick Decision)

• ChatGPT Agents (in-chat): Multi-step reasoning, tool calls, great for creative/knowledge work and ad-hoc tasks.

• n8n Agents (workflow): Schedules, webhooks, retries, logs, hits any API, sends emails, writes to CRMs—best for production.
  Pattern: Design the schema + reasoning in ChatGPT; run at scale in n8n.

Implementation Checklist

• Choose mode deliberately (Instant for drafts, Thinking/Extended for finals)

• Use RTITCO every time

• Get a JSON plan before production

• Version schemas and prompts (treat as product configs)

• Validate JSON; log prompt → plan → output

• Wire to Notion/Slack/Gmail; schedule in n8n