Marketing case analysis skill

Analyze Marketing Cases

Turn marketing information into evidence-aware analysis and interview-ready communication. Keep the causal chain explicit: business problem -> communication goal -> strategy -> insight -> creative idea -> execution -> result or measurement plan.

Select a mode

Infer the mode from the request. Ask only when the requested deliverable is genuinely ambiguous.

1. Case PPT mode: Analyze an existing company campaign and deliver a polished 2-3 slide .pptx with images and concise text; never exceed four slides.
2. Interview answer mode: Analyze an existing successful or failed case and deliver a natural 2-3 minute Chinese speaking script that foregrounds the success or failure mechanism, plus likely follow-up questions.
3. Campaign advisory mode: Help design a future marketing plan. Ask focused questions when essential inputs are missing, then give actionable recommendations. When the user asks for slides, deliver an actual marketing-proposal .pptx of no more than four slides.
4. Illustrated Word report mode: Analyze an existing case and deliver a polished .docx combining structured prose, a timeline, and a restrained number of sourced images.

If the user requests more than one mode, reuse one evidence base and produce each requested output.

Establish the request boundary

Before using conversation history, decide whether the current request is an explicit follow-up or a new case.

• Treat it as a follow-up only when the user clearly refers to the previous case, such as “继续刚才的案例”, “在上一版基础上修改”, “把刚才内容做成 PPT”, or an unambiguous pronoun referring to the same brand and campaign.
• Treat it as a new case whenever the brand, product, company, industry, campaign, assignment, or planning target changes, even inside the same chat.
• For a new case, create a fresh evidence ledger, source list, image set, outline, and artifact workspace. Do not reuse previous-case facts, industry assumptions, competitor names, dates, metrics, copy, images, search results, or visual assets.
• Carry forward only durable user preferences that are independent of the case, such as requested language, page limits, white-background styling, citation preferences, and output format.
• Do not mention or compare with a previous case unless the user explicitly asks for a cross-case comparison.
• When continuity is genuinely ambiguous, prefer a fresh case boundary and state the new-case assumption briefly instead of blending unrelated industries.

Establish the evidence base

1. Use the user's supplied materials as the primary source.
2. For an existing real-world case, research current or missing facts when browsing is available. Prefer official brand releases, original campaign pages, original platform posts, agency case studies, and reputable reporting.
3. Record four categories before reasoning: verified fact, attributed claim, inference, and recommendation.
4. Never invent campaign dates, budgets, sales, ROI, reach, conversion, or audience research. If results are unavailable, state that and propose measurement indicators instead.
5. Research images as evidence, not decoration. Prefer the brand's official account and original posts on the campaign's actual platform, including Xiaohongshu, Weibo, Douyin, Bilibili, WeChat, e-commerce pages, or event listings when accessible. For each image, record the source URL, account or publisher, date when available, and the claim it supports. Do not treat a repost or unlabeled screenshot as proof.
6. Preserve source links near claims, in DOCX captions or source notes, and in PPT notes or compact slide footnotes.

Keep the evidence ledger, uncertainty classification, alternative assessment, and scoring rubric as internal working material. Do not expose them as full-page methodology or evidence-boundary tables in an audience-facing report unless the user requests an audit trail. Express material uncertainty through concise attribution, footnotes, or source notes.

Apply the core framework

Read references/framework.md before analyzing or planning a case. Define the objective and strategic boundary before developing insights or creative ideas. Refine the strategy if later evidence contradicts it.

For all existing-case outputs—PPT, interview answer, and illustrated Word report—use this order:

1. Reconstruct the event: time, place, company, event, objective, distinctive visual or spatial features, and the online-offline connection.
2. Reconstruct the exact communication rhythm: warm-up, burst, and long-tail dates, content, channels, and handoffs.
3. Reconstruct the core actions: identify which celebrity, KOL/KOC, account, content unit, or offline activation did what, when, where, and for what role.
4. Explain why users wanted to spread it: identify the participation hook, social currency, emotional or identity value, and UGC loop.
5. Evaluate verified outcomes and extract reusable methods, limitations, and improvements.
6. Connect the observations back to strategy, audience insight, creative mechanism, execution, and business or brand value.

Do not merely list STP, 4P, SWOT, or platform names. Explain why each choice serves the objective and audience.

Read references/comparison-and-failure.md for every existing-case analysis. Compare the chosen tactic with at least one plausible alternative, distinguish channel roles, and extend the conclusion to competitors and the industry. For a failed case, switch from breakout analysis to the mirrored failure workflow and include a grounded counterfactual.

Produce the selected output

Read references/mode-outputs.md and follow the selected mode exactly.

For any PPT output, enforce a hard maximum of four slides and at least five distinct, relevant images. Use the user's visual references only for layout logic unless they explicitly ask to copy a palette or brand system. Default to a restrained, low-saturation professional style.

• For Case PPT mode, use the available presentation-creation skill or tooling, perform image research, create an actual .pptx containing at least five distinct evidence-bearing images, render every slide, and visually inspect it before delivery. Do not return only an outline unless the user explicitly requests one.
• For Interview answer mode, optimize for spoken Chinese, not report prose. Lead with a sharp thesis about why the case succeeded, broke out, underperformed, or backfired, then use case facts and a relevant comparison to prove it.
• For Campaign advisory mode, consult iteratively. Ask only high-value questions that materially change the plan; otherwise state reasonable assumptions and proceed. If a PPT is requested, read references/marketing-plan-ppt.md and apply its strategy-first, top-down structure and layout rules.
• For Illustrated Word report mode, use the available document-creation skill or tooling, create an actual .docx of no more than six rendered pages, and embed at least three distinct sourced images. Reserve page 1 entirely for a case introduction containing the case title, positioning subtitle, large real-world hero scene, core judgment, essential event metadata, image caption or source, and research date. Begin detailed analysis on page 2. At least two embedded images must be real scene-based visuals rather than text-heavy posters or screenshots. Render every page and visually inspect it before delivery.

For DOCX rendering, try the host agent's packaged document renderer first. On Windows, if no renderer is available, run powershell.exe -NoProfile -ExecutionPolicy Bypass -File scripts/render_docx_windows.ps1 -InputDocx <docx> -OutputDir <clean-output-dir>. The script searches PATH, common installation locations, optional explicit executable paths, and known local fallbacks before converting through headless LibreOffice and rasterizing with pdftoppm. Treat this fallback as a full render gate: inspect every produced page image, verify that the first-page hero image is visible, and revise until the report is at most six pages. Do not skip visual QA merely because the host agent lacks a built-in renderer.

Before delivery, run scripts/check_artifact_constraints.py with the bundled workspace Python. For PPTX use --min-images 5; for DOCX use --min-images 3 --render-dir <render-output> --max-pages 6. A passing count does not replace visual review: confirm that every counted image is distinct and relevant, and that the Word hero image is visibly rendered on page 1.

Use references/channels-and-metrics.md when selecting channels, phases, or KPIs. Treat its mappings as heuristics, not fixed rules.

Read references/breakout-logic-and-writing.md before writing the conclusion, interview answer, or executive summary. Select only the transformation lenses supported by evidence; never force fashionable phrases onto a case.

Polish the language

Apply a final editorial pass inside this skill; do not require a separate generic writing skill. Preserve facts while improving judgment, causality, rhythm, and audience fit. Follow the carrier-specific rules in references/breakout-logic-and-writing.md.

Quality check

Score the work with references/evaluation-rubric.md. Revise before delivery when any critical item fails.

Confirm that:

• the business problem and communication objective are distinct;
• strategy precedes and constrains creative execution;
• the audience insight is more specific than demographics;
• the creative idea expresses product or brand value;
• channel choices match audience behavior and funnel stage;
• facts, inferences, and recommendations are visibly separated;
• outcomes are supported, or clearly labeled as proposed KPIs;
• the final output matches the requested format and duration.
• the analysis states the breakout mechanism and proves it with case-specific evidence;
• named people, dates, locations, content, and platform actions are verified or visibly qualified;
• the language is concise and insight-led without becoming slogan-heavy.
• the analysis explains why the brand chose this tactic rather than a plausible alternative;
• competitor and industry conclusions distinguish portable methods from brand-specific assets;
• failure analysis separates intended strategy, observed outcome, likely mechanism, and counterfactual improvement without hindsight bias.
• PPT outputs contain at least five distinct, relevant images; logos, icons, decorative shapes, and repeated crops do not count;
• every PPT output contains no more than four slides, uses one consistent white background across the entire deck by default, and keeps each slide centered on one decision or conclusion; never insert an isolated dark-background slide;
• Word outputs contain at least three distinct, relevant images, including at least two real scene-based images; page 1 is a dedicated case introduction with a large scene-based hero image, and the complete rendered document is no more than six pages;
• a new brand, industry, campaign, or planning target does not inherit any case-specific text, facts, sources, images, competitors, or assumptions from an earlier request.