Enterprise AI Scenario Map Skill: a prompt-and-template pipeline that turns a company name into a 15 to 30 page AI adoption report

The gap this fills is a scoping problem, not a modelling problem

The README frames the target user bluntly: a business leader who has sat through conference talks and still cannot say which of their own processes should get an AI budget first. That is a different problem from building a model or wiring an API. It is the problem of producing a defensible list of candidate use cases, ordered, with some reasoning attached, in a document short enough to survive a management meeting. The repository's stated output is a 15 to 30 page report with a fixed six-part spine: executive summary, company profile, industry scan, scenario map, implementation path, next steps. Whoever wrote it has clearly delivered consulting documents before, because the page counts per part are specified (one page for the summary, three to four for the profile, five to six for the scenario map) rather than left to the agent. The intended audience is consultancies, internal strategy teams, and anyone who currently answers the 'where can we use AI' question with a slide deck assembled by hand. The README also positions it against that deck directly, describing the output as something you can take into a board meeting rather than as empty slides.

Four stages, and only one of them is code

The pipeline diagram in the README shows four stages: deep research, analysis and diagnosis, scenario mapping, report generation. The important detail is the design note underneath it. The Python script, scripts/deep_research_wrapper.py, does not perform research. According to the README, it generates a research framework and a list of search queries. The agent then executes those queries through its own web-search tool, reads the results, and writes the analysis. The remaining three stages are driven by the Markdown references: business-analysis-framework.md for the diagnosis, scenario-priority-framework.md for ranking, typical-ai-scenarios.md as a seed library, and report-template-v2.1.md to shape the final document. SKILL.md is the entry point the agent reads. This split has a clear consequence the README states outright: there are zero external API dependencies and zero extra cost, and research quality depends entirely on the agent's capability. That is an honest framing, and it is also the project's main architectural risk. Nothing in the repository validates that a cited case study exists or that a claimed pain point appears in any source. The structure is enforced; the content is not.

Three modes, and the page counts are the real specification

The README defines three modes rather than one. Quick Scan targets 15 to 20 scenarios across 5 to 8 pages. Standard targets 30+ scenarios across 15 to 20 pages. Deep Plan targets 50+ scenarios across 20 to 30 pages, and is the only mode the README associates with ROI. The mode is selected through the natural-language prompt, not a configuration flag: you ask for a report and state the depth you want. This is worth pausing on. Because there is no schema, no required argument list, and no validation step described in the material, the difference between Standard and Deep Plan is a difference in instruction-following, not a difference in code path. If your agent is weak at holding a long structural spec, the 50+ scenario mode will degrade toward the 30+ mode and you will not get an error. The page-count table is therefore best read as a prompt specification, and the practical way to test it is to run the same company through two modes and compare the scenario counts in the output.

Getting it running is three commands and one platform constraint

The Quick Start is short. Clone the repository, load SKILL.md into your agent, then issue a prompt such as: Generate an AI scenario map report for "XX Company". The prerequisites are Python 3.8+ for the research framework script and an agent that supports a web-search tool, with Claude Code and ChatGPT with browsing given as examples. The README notes that the loading method for SKILL.md varies by platform, and does not enumerate those methods, so that step is the one most likely to need improvisation. There is a discrepancy worth flagging: the clone command in the README points at github.com/YuanziAI/enterprise-ai-scenario-map, while the repository described here is MetaInFLow/Enterprise-ai-scenario-map-skill. The README also points to a company-info-config.md under references/ for configuring the issuing organisation's details, which is what fills the cover page fields (client name, date, issuing unit). Anyone integrating this should confirm the actual remote URL before scripting a clone, and should treat company-info-config.md as the first file to edit, since an unedited cover page is the fastest way to expose that a report was generated rather than written.

The 8-industry library is a starting point, not coverage

references/typical-ai-scenarios.md ships with typical scenarios for construction, e-commerce, finance, manufacturing, healthcare, education, legal, and logistics, listed at 8+ scenarios each. The README is explicit that industries outside that list are handled by live web search rather than by the library. Two things follow. First, for the eight listed industries the agent has a seed set to work from, which should improve consistency across runs. Second, for everyone else the entire scenario list is generated at inference time, and the quality of that list is a function of search results and the agent's synthesis. A public-sector body, a mining company, or a two-person design studio gets no seed material at all. The '30+' figure in the Standard mode is a target across both sources combined, so a report for an unlisted industry is doing proportionally more work from scratch. If you work mainly outside those eight verticals, the library is close to irrelevant to you and the value of the repository reduces to the template and the priority framework.

Where it breaks: no ground truth, and a bias toward volume

The failure mode is not a crash. It is a plausible report. Because research is delegated to an agent's web search, the output inherits every weakness of that search: stale sources, marketing pages presented as case studies, and confident claims about a company's internal pain points that no public source supports. The README's own framing ('90% of business leaders') is a rhetorical figure, not a cited statistic, and it is the kind of number that tends to reappear inside generated reports unless the agent is instructed otherwise. The second issue is the scenario count as a success metric. A mode that targets 50+ scenarios rewards enumeration, and a long list of shallow entries is worse than fifteen well-argued ones for a reader trying to decide where to spend money. The priority framework in references/scenario-priority-framework.md is the mechanism meant to counteract this, but the README does not describe its criteria, so you cannot judge from the material here whether it ranks on feasibility, cost, data readiness, or something else. Read that file before you trust the ordering. Finally, this is the wrong tool if you need a deterministic pipeline: the same company run twice through a different agent, or through the same agent on a different day, will not produce the same report, and nothing in the repository offers reproducibility controls.

The alternative is a prompt chain you write yourself

The obvious comparison is not another repository. It is the thing a competent consultant already does: a research prompt, a spreadsheet of candidate use cases, a scoring column, and a slide template. The difference in approach is that this project externalises the structure into files. The scenario library, the priority framework, the case template, and the report template are all versioned Markdown, so a change to how scenarios are ranked is a diff in references/scenario-priority-framework.md rather than a paragraph someone edits in a private prompt. That is the real argument for adopting it: not that the agent is smarter, but that the assessment method is inspectable and shared across everyone on the team. A hand-rolled prompt chain gives you the same flexibility with none of the version history, and a commercial AI-roadmap tool gives you a hosted interface with none of the file-level control. The trade this repository makes is that you supply the agent, the search capability, and the quality check, and it supplies the skeleton.

Licence and the cost of keeping the templates current

The repository is MIT licensed, which permits commercial use, modification, and redistribution provided the copyright notice and permission notice are retained. Note the README's sign-off line asking for a star; that is a request, not a licence condition, and MIT imposes no attribution requirement beyond preserving the notice. The maintenance picture is unusual for a code project. There are no releases retrieved, and the substantive content lives in Markdown templates and a scenario library rather than in the Python file. That means the upgrade cost is editorial, not technical: report-template-v2.1.md is versioned by filename, so a future v2.2 would need to be diffed by hand against your own edits rather than merged through a package manager. If you fork and customise the template for your own branding, you have taken on the job of tracking upstream changes to it. The scenario library has the same property. Budget for a periodic re-read of references/typical-ai-scenarios.md, because AI use cases in the eight listed industries move faster than the file will.

Editorial conclusion

Adopt this if you already run an agent with web search, you want a repeatable structure for AI opportunity assessment, and you accept that the report's accuracy is bounded by what your agent can find and cite. Do not adopt it if you need audited figures, a deterministic pipeline, or a deliverable you can hand to a client without a human reading every claim. Before you commit, open references/typical-ai-scenarios.md and check whether your industry's entries are specific enough to be useful, and read report-template-v2.1.md to see how much of the 30+ scenario count is structure rather than researched content.

Official sources

1. Issues
2. License: MIT
3. MetaInFLow/Enterprise-ai-scenario-map-skill on GitHub
4. README