The original design asked one LLM call to handle platform selection, situation diagnosis, content ideation, scheduling, ad strategy, and KPI setting simultaneously. Six independent model reviews agreed this was the core flaw. The fix is to separate each cognitive task into its own step, with deterministic code handling the logic that doesn't need AI reasoning.

When attention dilutes across 7 simultaneous cognitive tasks in a single LLM call, the output in Sections 4-7 goes weak. Each step below gets a focused call with one job. That's the entire architectural change.

Block 1: What you're building

Q1. What does your business do? Options: sells a physical product / sells a digital product or app / offers a service to consumers / offers a service to other businesses (B2B) / creates content or runs a newsletter / runs a local brick-and-mortar business

Q2. [NEW] What is your average sale price? Options: under $50 (impulse buy) / $50-500 (considered purchase) / $500-2,000 (high consideration) / over $2,000 (high-ticket or ongoing retainer). Why it matters: high-ticket businesses need trust-first content strategies. Low-ticket businesses can drive impulse purchases via Instagram and TikTok ads.

Q3. Who is your ideal customer? Age range: under 35 / 35-50 / 50-65 / 65+. Type: everyday consumer buying for themselves / professional buying for work or their business

Q4. Geographic scope: local or regional (city/state) / national or global

Block 2: Your starting point

Q5. What is your current situation? Options: zero social accounts for this business / have accounts but almost no followers / accounts with followers but growth stalled / accounts that were growing then stopped working

Q6. What have you already tried? (check all that apply) [CONDITIONAL: only shows if Q5 is not "zero accounts"] Options: posting consistently / paid ads / hashtags / collaborations / went viral once then nothing / posting randomly / comment-to-DM automation / influencer partnerships

Block 3: Honest capacity check

Q7. How many hours per week can you realistically spend on social media? Options: under 2 hours / 2-5 hours / 5-10 hours / I have a team or VA who can help

Q8. [NEW] How comfortable are you being on camera? Options: very comfortable, I enjoy it / I could try with some practice / I strongly prefer to stay off camera

Q9. [NEW] Will you (or a specific person) be the recognizable face of this business? Options: yes, personal brand / no, the business is the brand

Q10. What content formats feel natural? (pick all that apply) Options: writing captions and articles / teaching or explaining things on video / behind-the-scenes and lifestyle content / none of these feel comfortable yet. [NEW addition: professional graphics and designed content]

Block 4: Goals and voice

Q11. What does success look like in 90 days? Options: people finding me and joining my email list / direct sales from social / a community of engaged followers / brand awareness

Q12. [NEW] How would you describe your brand's personality? Options: funny and casual / warm and relatable / serious and authoritative / bold and direct. Why it matters: brand voice shapes every post example. Without this, all examples come out generic.

Optional (improves output significantly): "Describe your specific niche in one sentence" (free text, 100 chars max) and "Monthly budget for paid promotion" ($0 / under $100 / $100-500 / $500+).

This is pure code. No LLM. The rules are deterministic: if/then based on quiz answers. LLMs aren't used here because they hallucinate platform recommendations when given conflicting signals. The function accepts a quiz-answers dict and returns a ranked list of platforms capped by the user's time budget.

Key routing decisions baked in: B2B/professional audiences route to LinkedIn first. Local businesses route to Facebook Groups. Under-35 consumers go to TikTok unless they're camera-averse, in which case Instagram carousels. The 50+ audience routes to Facebook as primary, with YouTube as a trust-builder for high-ticket offers. Content creators go to Substack first. Time budget caps the platform count (1 platform for under-2-hours/week; 3 for teams).

def select_platforms(q):
    candidates = []
    no_video = q['on_camera'] == 'no_camera'

    if q['audience_type'] == 'professional' or q['business_type'] == 'b2b':
        candidates.append({'platform': 'LinkedIn', 'priority': 1,
            'reason': 'your audience is professionals who make decisions on LinkedIn'})
        if not no_video and ('teaching' in q['content_formats']):
            candidates.append({'platform': 'YouTube', 'priority': 2,
                'reason': 'long-form educational content builds authority with professional buyers'})
        elif 'writing' in q['content_formats']:
            candidates.append({'platform': 'Substack', 'priority': 2,
                'reason': 'thought leadership content reaches professional decision-makers'})

    elif q['geography'] == 'local':
        candidates.append({'platform': 'Facebook', 'priority': 1,
            'reason': 'local community lives in Facebook Groups'})

    elif q['audience_age'] == 'under35' and q['audience_type'] == 'consumer':
        if not no_video:
            candidates.append({'platform': 'TikTok', 'priority': 1,
                'reason': 'under-35 consumer audience is most active on TikTok'})
        candidates.append({'platform': 'Instagram', 'priority': 2,
            'reason': 'secondary discovery platform for under-35 consumers'})

    elif q['audience_age'] in ['50-65', '65+'] and q['audience_type'] == 'consumer':
        candidates.append({'platform': 'Facebook', 'priority': 1,
            'reason': 'your audience is heavily concentrated on Facebook'})

    limits = {'under2': 1, '2-5': 2, '5-10': 3, 'team': 3}
    max_p = limits[q['hours_per_week']]
    candidates.sort(key=lambda x: x['priority'])
    return candidates[:max_p]

Based on quiz answers, the system assigns one of six content archetypes. The archetype becomes the hero identity injected into every LLM prompt, shaping the tone, focus, and positioning of the entire report.

The Six Archetypes

The archetype appears as the hero identity at the top of the personalized report and is injected into every LLM prompt (Steps 2, 3, 4, 6), ensuring that diagnosis, playbooks, scheduling, and ad strategy are all written through the lens of that archetype's strengths and constraints.

One focused task: diagnose where this person is and what their biggest challenge is. This prompt gets full attention on diagnosis only. No recommendations, no tactics. The output is 3-4 paragraphs of honest prose addressed directly to the reader.

The prompt instructs the model to cover: an honest framing of their starting position, the most likely root cause if they have accounts that aren't working, a reality check on their time budget vs. their goal, and one sentence of specific encouragement. Hard rules: contractions throughout, no em-dashes, no bullet lists.

One call per recommended platform, all running simultaneously. Each gets the full quiz context but focuses on ONE platform only. This is where quality improves most dramatically over the single-prompt approach. Eight platform prompts are defined, each with current 2025-2026 platform knowledge baked into the system prompt so the model doesn't have to recall it from training data.

Each playbook prompt covers 4-6 specific output areas: account setup checklist with examples, format mix tailored to their on-camera comfort, a specific first action (comment-to-DM keyword, Groups strategy, board names, hook formula, etc.), and three complete copy-paste-ready example posts labeled by purpose (Reach, Engagement, Conversion) with platform-specific formatting applied. YouTube is the exception: instead of posts, it generates the word-for-word hook script for the first 30 seconds of their first video. The key instruction across all prompts: finished, publishable output specific to their business -- not topics, not angles, not generic advice.

Instagram-specific: Comment-to-DM Automation

The Instagram prompt includes current knowledge on comment-to-DM keyword automation as the highest-converting tactic right now. Tool: InstantDM ($9.99/month, Meta-certified). Open rates run 80-90%. The model is instructed to give the user one specific keyword and one specific resource to offer in the automated DM, not a generic suggestion.

Substack-specific: Cross-Recommendation Strategy

The Substack prompt includes the current cross-rec mechanic (recommendations drive 20% of all new subscriptions across the network), the specific outreach sequence (comment first for several weeks, add them to your own recommendations list before asking, then pitch), and the Reletter.com tool for finding similarly-sized newsletters to approach.

This call receives the playbook outputs from Step 3 and synthesizes them into a cohesive schedule. Its job is coordination and scheduling, not inventing new tactics. Everything it produces should draw from the playbooks already written.

Output structure: Week 1 is setup and content stockpile only (the key rule: publish nothing in Week 1). Week 2 is the warm-up engagement routine before the account has any content. Week 3 is go-live, with specific first-post guidance pulled from the playbooks. Week 4 is the first data review with explicit if/then adjustment logic.

The daily routine section adapts to their time budget: a 15-minute habit for under-2-hours/week, the full 30-5-2 routine (engage with 30 posts, comment on 5, DM 2 relevant accounts) for 2-5 hours/week, and batched content creation days for 5+ hours/week.

No LLM. Pull from a static table keyed by platform and primary goal. Each entry includes three metrics with specific numeric targets and how-to-check instructions. The table currently covers Instagram (3 goal variants), Facebook, YouTube, LinkedIn, Substack, Pinterest, and TikTok. It expands as new platform/goal combos are validated.

Example: Instagram, email-list goal returns engagement rate (target 3-6%, check via Insights Overview), DM conversations started (target 5+ per week by week 4), and email signups from bio link (target 2-5% of profile visitors, verified via email platform subscriber source). The metrics are always checkable with free native platform tools, never third-party paid analytics.

This step only runs if monthly_budget is greater than $0. If the user selected $0, the report says "Organic first -- ads after you have proof of what works" and skips the step entirely. No wasted tokens on users who aren't going to run ads.

When it does run, the prompt is heavily constrained to produce one focused recommendation: ONE platform, ONE campaign objective, ONE specific daily budget (not a range), ONE creative approach to test first, a specific evaluation window, and explicit success/failure definitions so the user knows when to stop vs. scale.

Two hard routing rules are baked in. High-ticket (price over $500): awareness or traffic campaigns first, not conversion campaigns. Cold audiences don't convert directly on high-ticket offers; the ad warms them up, the email or call closes them. Low-ticket (price under $50): conversion campaigns directly to a product page or lead magnet are appropriate for cold audiences.

Deterministic code combines all step outputs into the final report, injects the upsell CTA, and formats for delivery. The paid ads section is conditionally included. The upsell CTA is always last. Output is delivered as a formatted PDF or long-form HTML email via Resend. The report also includes a "Download 30-Day Calendar (CSV)" button compatible with Buffer, Publer, and Metricool for immediate scheduling.

def assemble_report(step_outputs, quiz_answers):
    report_sections = [
        format_header(quiz_answers),
        format_platform_stack(step_outputs['platforms']),
        step_outputs['diagnosis'],
        format_playbooks(step_outputs['playbooks']),
        step_outputs['thirty_day_plan'],
        format_milestones(step_outputs['milestones'], quiz_answers),
    ]
    if quiz_answers['monthly_budget'] != 'zero':
        report_sections.append(step_outputs['paid_ads'])
    else:
        report_sections.append(ORGANIC_ONLY_MESSAGE)
    report_sections.append(UPSELL_CTA)
    return '\n\n'.join(report_sections)