The Multi-Pass AI Content Engine: From Brief to Publish-Ready Asset
Objective: Execute the lesson's 3-pass AI writing framework (outline generation, section drafting with voice calibration, and editorial polish with banned-words pruning) to produce a publish-ready 1,200-word educational guide that passes brand voice and fact-checking standards.
You are the senior content marketer at Freshworks producing a comprehensive guide on 'Customer Service SLA Management' for the Freshdesk blog. Rather than single-shotting an essay, you will orchestrate a 3-pass workflow across Claude and ChatGPT.
Step through the complete production pipeline: choose the right model profile, construct a 4-part outline prompt, generate section drafts anchored on real case examples, and run a dedicated editorial cleaning pass to eliminate AI-tells.
Before you start
What you'll need
Free path (everything below is enough to finish)
Free tier provides access to Claude 3.5 Sonnet for long-form drafting
Free tier provides fast constraint enforcement and formatting checks
Free, collaborative spreadsheet
Free word processor with revision history
Paid upgrades (optional, faster/deeper)
The free multi-model workflow (Claude + ChatGPT) achieves full production quality; paid enterprise tools like Jasper automate shared style guides for large writing teams.
Scales company-wide style guide enforcement across multi-author teams
The process
4 steps
Step 01 of 04
The lesson details that matching model strengths (Claude for nuanced long-form tone and prose flow, ChatGPT for structured schemas and bulk variants, Gemini for Workspace integration) prevents voice degradation and reduces rewrite cycles.
For a 1,200-word strategic B2B support guide requiring natural editorial voice and nuanced tone, which primary drafting model and secondary editing tool should you configure?
Procedure
- Review the content deliverable: 1,200-word strategic guide with nuanced B2B advice
- Select Claude (Sonnet/Opus) as the primary drafting model for superior natural prose and long-form voice
- Select ChatGPT as the structured editing and constraint-checking model
- Document the handoff protocol between drafting and editing environments
Freshworks Model Routing Matrix
Deliverable: 1,200-word B2B Guide ('Customer Support SLA Management')
Drafting Engine: Claude 3.5 Sonnet (excels at nuanced, natural long-form voice without five-paragraph stiffness)
Editing Engine: ChatGPT-4o (excels at rigid rule adherence for banned-words pruning and formatting audits)
Verification Layer: Human Editor (verifies SLA calculation benchmarks and citations)Healthy
Routing long-form voice tasks to Claude and rule-based editorial checks to ChatGPT based on distinct model strengths.
Unhealthy
Defaulting to a single tool for all workflows without considering output prose quality or structural bias.
What this means
Matching model specializations cuts human editing time by ensuring first drafts start with strong sentence variety and natural cadence.
So what do I do about it?
| Symptom | Action | Effort |
|---|---|---|
| Drafts consistently feel rigid and sound like high school five-paragraph essays | Switch the drafting engine from default GPT models to Claude and enforce the 3-pass workflow | 5 min |
Step 02 of 04
Pass 1 generates the architectural outline. Using the 4-part structure (Role, Task, Context, Constraints), you prompt the model to deliver a detailed H2/H3 hierarchy with bulleted talking points before writing prose.
What structured prompt ensures the model produces an actionable 4-section outline with specific subheadings rather than high-level generic advice?
Procedure
- Draft the Role: 'You are a veteran B2B SaaS customer success strategist writing for Freshworks'
- Draft the Task: 'Create a detailed outline with 4 H2 sections and 2-3 H3 subsections per topic for a guide on Customer Service SLAs'
- Draft the Context: 'Audience is support team leads managing 10-50 agents. Focus on first-response time vs. resolution time tradeoffs'
- Draft the Constraints: 'Return only headings and 2 bulleted subpoints per heading. Do not write the full draft yet'
- Inspect the outline and adjust section order before proceeding to drafting
Generated Freshworks Outline (Pass 1): H1: The Modern Customer Service SLA Playbook ## 1. Defining SLAs That Protect Revenue Without Burning Out Agents - First-response time (FRT) vs Mean Time to Resolution (MTTR) - Tiered SLAs based on customer ARR and ticket severity ## 2. Setting Realistic Baseline Metrics (With Industry Benchmarks) - Analyzing historical ticket volume spikes in Freshdesk - Building SLA buffer thresholds for omnichannel queues ## 3. Automation and Escalation Workflows - Automated routing rules before breach warnings trigger - Multi-tier escalation trees for VIP accounts ## 4. SLA Breach Post-Mortems: Turning Misses into Process Fixes - Root-cause tagging in ticketing analytics - Team-wide SLA review meetings that focus on systems over blame
Healthy
Locking in a granular, approved outline with specific operational topics before generating a single paragraph of prose.
Unhealthy
Skipping the outline stage and asking the model to write the complete 1,200-word draft in one prompt.
What this means
The outline is your editorial blueprint; adjusting hierarchy and angles at the outline stage takes 2 minutes versus 30 minutes of rewriting prose.
So what do I do about it?
| Symptom | Action | Effort |
|---|---|---|
| AI-generated content drifts off-topic and misses key technical nuances | Require outline sign-off in Pass 1 before permitting section drafting | 5 min |
Step 03 of 04
Pass 2 drafts each section individually. By pasting the approved outline section and 2-3 paragraphs of past high-performing company copy as in-context style examples, the model maintains authentic brand voice and depth.
How do you feed the outline and context examples into the model section-by-section to generate 300-word deep-dive segments?
Procedure
- Take Section 1 from the approved outline
- Paste 2 past high-performing Freshdesk blog paragraphs to set tone, sentence length, and pacing
- Prompt Claude to write Section 1 (300 words) using direct, practical language with zero introductory fluff
- Repeat for Sections 2, 3, and 4, ensuring each section concludes with a clear transition line
- Compile the 4 drafted sections into a unified Google Docs draft
Freshdesk Guide — Section 1 Draft Sample (Pass 2): When support leaders set service level agreements, they often fall into a predictable trap: treating first-response time as the only metric that matters. An automated auto-responder can hit a 60-second first-response target every time, but it resolves zero customer frustrations. A resilient SLA structure splits commitments into two tiers: Initial Response (acknowledging and triaging the issue) and Next-Action Resolution (providing a meaningful fix or status update). High-performing support teams configure separate targets based on ticket priority: - Critical / P1 (System Down): 15-minute response, 2-hour resolution target - High / P2 (Major Feature Broken): 1-hour response, 8-hour resolution target - Standard / P3 (General Inquiries): 4-hour response, 24-hour resolution target
Healthy
Drafting section-by-section with pasted tone examples, keeping each section focused on concrete tactical frameworks.
Unhealthy
Generating all sections at once without voice examples, resulting in repetitive introductory summaries in every section.
What this means
In-context voice examples constrain the model's token distribution, yielding authentic company tone without requiring fine-tuned models.
So what do I do about it?
| Symptom | Action | Effort |
|---|---|---|
| AI drafts sound disconnected from the company's brand voice | Paste 2-3 paragraphs of published, high-performing brand copy into the prompt context | 5 min |
Step 04 of 04
Pass 3 executes a strict editorial scrub. The draft is audited against a shared banned-words list, paragraphs are tightened to a 2-3 sentence maximum, and statistical claims are flagged for mandatory human fact-checking.
What automated cleaning prompt and verification checklist guarantees the draft contains zero AI tells and only verified data before publishing?
Procedure
- Feed the full draft into ChatGPT with the editorial cleaning prompt
- Instruct the model: 'Review this draft. Remove every instance of: delve, unlock, leverage, seamless, game-changer, robust, in today's fast-paced world, it's worth noting. Split any paragraph longer than 3 sentences'
- Highlight all numerical statistics and verify each against primary research in Google Sheets
- Finalize the draft in Google Docs for staging in the CMS
Editorial Audit Log, Freshworks Content Ops Draft: 'Customer Service SLA Management Guide' (1,240 words) AI-Tells Scrubbed: - 'delve into SLA metrics' -> replaced with 'audit your SLA metrics' - 'seamlessly integrates' -> replaced with 'connects directly' - 'unlock the true potential of your support team' -> replaced with 'reduce agent burnout' - 'in today's rapidly evolving SaaS landscape' -> removed entirely Fact-Check Status: - HDI 2024 Support Benchmarks citation: VERIFIED (Source: HDI Global Report, 2024) - Zendesk Benchmark Average FRT (12.4h): VERIFIED (Source: Zendesk CX Trends, 2024) Final Status: Clean, publish-ready in CMS
Healthy
Running a systematic negative-words audit and verifying 100% of cited numbers before publication.
Unhealthy
Publishing AI output directly without scrubbing banned phrases or verifying hallucinated statistics.
What this means
The third pass is the safety net that transforms raw AI output into authoritative, trust-building enterprise content.
So what do I do about it?
| Symptom | Action | Effort |
|---|---|---|
| Published articles contain obvious AI clichés that damage executive credibility | Enforce Pass 3 automated cleaning as a mandatory pre-publish gate in your CMS workflow | 5 min |
Final deliverable
A complete 4-part AI production workbook including model routing table, approved outline, compiled section drafts, and an editorial scrub log with zero banned AI terms.
See a reference example
Klaviyo Content Ops: 3-Pass AI Production Output Asset: 'E-commerce Abandoned Cart Recovery Strategy' (1,180 words) Pass 1 (Outline Prompt via Claude 3.5 Sonnet): - H2: Anatomy of a High-Converting 3-Part Cart Recovery Sequence - H3: Timing the 1-hour transactional reminder - H3: Introducing dynamic discount incentives at 24 hours - H2: Calculating True Recovery ROI vs Margin Erosion Pass 2 (Section Draft with Context Injections): Drafted 350-word Section 1 citing Klaviyo 2024 benchmark data (3.4% average SMS recovery rate). Pass 3 (Editorial Scrub via ChatGPT): - Removed: 'In today's rapidly evolving e-commerce landscape' -> Replaced with: 'In 2026, cart abandonment rates average 70.19% across Shopify stores.' - Removed: 'unlock hidden revenue', 'seamlessly integrate', 'delve into' - Formatted paragraphs to 2 sentences max. - Status: 100% verified, zero banned terms.
Success criteria
You're done when you can:
- Executes all 3 passes sequentially without attempting a single-shot draft
- Applies negative constraints that eliminate 100% of banned AI terms ('delve', 'unlock', 'seamless', 'game-changer')
- Verifies all statistical claims with dated third-party sources in a fact-checking log