How to build an Insight & Recommendation report with AI

Where to run this

This is an installable AI Skill (Claude Skill format) that turns campaign performance data into a structured, recommendation-led report, not a metric dump.

Upload first

• ☐Final client-approved brief (business objective, campaign objective, target audience, market(s), KPIs, success criteria, budget, flight dates, approved channels)
• ☐Final approved media plan (planned budget, reach, frequency, audiences, channels, allocations, testing framework)
• ☐Campaign performance data, preferably raw log-level exports (DV360/CM360 logs, site analytics, conversion or event-level exports), platform exports as a minimum (Google Ads, Meta, TikTok, DSP exports)
• ☐Analytics, CRM or sales data, if available, for genuine business-impact interpretation

Steps

1. 01

   Read this before you start

   Do this

   A report will only be generated when all three mandatory inputs are supplied: the final client-approved brief, the final approved media plan, and campaign performance data. The skill will not invent sales, revenue, business impact, deduplicated reach, attribution claims or incrementality claims. If data is missing, the report must clearly state what's known, what's unknown, and what cannot be concluded, not paper over the gap. Reports exist to improve future decisions, if an observation doesn't lead to a recommendation, it gets removed, not padded into the report anyway.

2. 02

   Understand the discipline this skill enforces

   Do this

   Every insight follows one chain: Observation, then Insight, then Confidence, then Recommendation. An observation with no recommendation attached gets cut, not included for the sake of completeness. Reach across platforms can never be summed, YouTube reach plus Meta reach plus TikTok reach is not unique reach, the same person can be exposed on all three. Impressions and clicks can be summed, they're delivery events and actions, not people. Deduplicated cross-platform reach is only usable when it comes from a named, approved cross-media measurement source, Nielsen, Kantar, Knorex, Halo, or an equivalent approved provider, otherwise the report states plainly that deduplicated reach is unavailable, it doesn't estimate one.

3. 03

   Install the skill in your AI platform

   Do this

   Add it to your Project's knowledge, Skills directory, or a dedicated thread depending on your AI platform. The full installable content is in the block below.

4. 3a

   The full skill content to install

   ---
   name: insight-recommendation-architect
   description: >
     Transforms campaign reporting from a metric dump into a learning engine. Answers what
     happened, why it happened, what it means, what to do next, and how the next plan should
     change. Requires a final client-approved brief, a final approved media plan, and campaign
     performance data. Never invents sales, revenue, business impact, deduplicated reach,
     attribution or incrementality claims. Every insight must lead to a recommendation or be
     removed.
   ---
   
   PURPOSE: transform campaign reporting from a metric dump into a learning engine. Answer: what happened, why did it happen, what does it mean, what should we do next, how should the next plan change.
   
   MANDATORY INPUTS: (1) Final Client-Approved Brief, must include business objective, campaign objective, target audience, market(s), KPIs, success criteria, budget, flight dates, approved channels. (2) Final Approved Media Plan, must include planned budget, planned reach, planned frequency, planned audiences, planned channels, planned allocations, testing framework. (3) Campaign Performance Data, preferred is raw log-level data (DV360 logs, CM360 logs, site analytics exports, conversion exports, event-level exports), minimum acceptable is platform exports (Google Ads, Meta, TikTok, DSP exports).
   
   DATA QUALITY GATE, classify before generating any insight: HIGH CONFIDENCE, raw log-level data. MEDIUM CONFIDENCE, aggregated platform reports. LOW CONFIDENCE, screenshots, PowerPoints, PDF reports. Every final report must contain a Data Quality Assessment section stating which tier applies and why.
   
   REACH GOVERNANCE: reach cannot be summed, YouTube Reach + Meta Reach + TikTok Reach does not equal unique reach, the same individual can be exposed across multiple platforms. Impressions can be summed, they are delivery events. Clicks can be summed, they are actions, though clicks do not equal unique visitors. Deduplicated reach is only usable when supplied by Nielsen, Kantar, Knorex, Halo, or an approved cross-media measurement source, otherwise state plainly: 'Cross-platform deduplicated reach unavailable.'
   
   BUSINESS INTERPRETATION FRAMEWORK: KNOWN, confirmed by supplied data. UNKNOWN, cannot be proven from supplied data, e.g. sales, revenue, incrementality, offline conversion. EDUCATED INTERPRETATION, a reasoned interpretation that does not overclaim and does not fabricate a business outcome.
   
   OBSERVATION TO RECOMMENDATION CHAIN, every insight must follow: Observation, then Insight, then Confidence, then Recommendation. If there is no recommendation, remove the observation, don't report it anyway.
   
   CONFIDENCE FRAMEWORK: HIGH, raw or comprehensive data. MEDIUM, platform export only. LOW, partial data, screenshots, limited visibility. Every insight must carry a confidence label.
   
   PLAN VERSUS ACTUAL, mandatory, a table of Metric | Planned | Actual | Variance covering at minimum budget, reach, frequency, impressions, clicks, conversions.
   
   FUTURE PLANNING FEEDBACK LOOP, every report must answer: Audience Impact (what should change in audience strategy), Channel Impact (what should change in channel strategy), Reach Impact (what should change in reach planning), Media Plan Impact (what should change in allocation and execution).
   
   MANDATORY REPORT STRUCTURE, in order: 1. Executive Summary. 2. Campaign Context. 3. Data Quality Assessment. 4. Plan vs Actual. 5. Key Insights. 6. Business Interpretation. 7. Recommendations. 8. Impact on Future Planning. 9. Caveats & Limitations. 10. Future Testing Opportunities.
   
   CLOSING PRINCIPLE: reports exist to improve future decisions. The final approved client brief, the final approved media plan, and campaign performance data are mandatory. If a recommendation cannot be formed from an observation, the observation should not be reported.

   Scroll inside the block to read it all — “Copy prompt” copies the entire text.

5. 04

   Pick the invocation prompt that matches your situation

   Do this

   Ten real scenarios, each a genuinely different data situation or scope you might actually be in.

6. 4a

   Complete campaign with real business metrics available

   /insight-recommendation-architect
   
   Create a client-facing insight report.
   
   Attached:
   - Final approved brief
   - Final media plan
   - Campaign performance data
   - Analytics data
   - CRM data
   - Sales data
   
   Please: apply the data quality gate; perform plan versus actual analysis; generate observations, insights and recommendations; interpret business impact; assign confidence levels; identify future planning changes; identify future testing opportunities; keep the report concise and executive-ready.

7. 4b

   No business metrics available, media data only

   /insight-recommendation-architect
   
   Create an insight report.
   
   Attached:
   - Final approved brief
   - Final media plan
   - Campaign performance data
   
   No sales or revenue data available.
   
   Please: apply the data quality gate; perform plan versus actual analysis; identify what is known; identify what is unknown; provide business interpretation without overclaiming; generate recommendations; generate future planning updates; assign confidence ratings.

   Do this

   This is the honest, no-business-data version, don't let the absence of sales data turn into an invented business-impact claim.

8. 4c

   Raw log-level data available, the highest-confidence case

   /insight-recommendation-architect
   
   Use the attached log-level campaign data.
   
   Please: classify data quality; analyse audience performance; analyse channel performance; build observations; convert observations into insights; build recommendations; build future planning updates; build testing recommendations.

9. 4d

   Platform exports only, a lower-confidence case

   /insight-recommendation-architect
   
   Generate an insight report from platform exports.
   
   Attached:
   - Google export
   - Meta export
   - TikTok export
   
   Please: classify confidence accordingly; generate observations; generate insights; generate recommendations; highlight reporting limitations; identify missing data.

10. 4e

    Plan versus actual only, a narrower deliverable

    /insight-recommendation-architect
    
    Compare the approved media plan against campaign delivery.
    
    Attached:
    - Approved media plan
    - Campaign results
    
    Please: build the plan versus actual section; explain variances; generate insights; generate recommendations; update future planning guidance.

11. 4f

    Cross-channel performance comparison

    /insight-recommendation-architect
    
    Evaluate performance across channels.
    
    Attached:
    - Approved brief
    - Media plan
    - Campaign data
    
    Please: compare channels; identify strongest and weakest contributors; do not sum reach across platforms; explain limitations; build recommendations; suggest future allocations.

12. 4g

    Audience learning analysis

    /insight-recommendation-architect
    
    Evaluate audience performance.
    
    Attached:
    - Audience architecture
    - Campaign results
    
    Please: rank audience learnings; identify winning audiences; identify weak audiences; recommend future audience changes; update audience strategy.

13. 4h

    Channel strategy learning analysis

    /insight-recommendation-architect
    
    Evaluate channel strategy effectiveness.
    
    Attached:
    - Channel strategy
    - Campaign results
    
    Please: determine whether channel roles were achieved; identify successful channels; identify underperforming channels; recommend future channel changes; update future channel strategy.

14. 4i

    Executive summary only

    /insight-recommendation-architect
    
    Generate an executive summary from the attached campaign materials.
    
    Limit output to: Executive Summary, Key Insights, Business Interpretation, Recommendations.

15. 4j

    Fast version, when everything is already attached

    /insight-recommendation-architect
    
    Generate an Insight & Recommendation Report.
    
    Attached:
    - Final approved brief
    - Approved media plan
    - Campaign performance data
    
    Apply: data quality gate; plan versus actual; reach governance; Observation to Insight to Recommendation framework; confidence ratings; future planning feedback loop.
    
    Do not invent business outcomes. Do not sum cross-platform reach. Keep output concise and recommendation-led.

16. 05

    Before this goes anywhere near a client

    Do this

    Check that every recommendation in the report actually traces back to a real observation with a confidence label, not a generic best-practice suggestion dressed up as a finding. Confirm no cross-platform reach was silently summed. Confirm anything about sales, revenue or incrementality is either backed by real data or explicitly marked unknown, never implied. If the data quality tier is Low, make sure the report says so plainly rather than presenting screenshot-derived numbers with false confidence.