메뉴
Run pre-implementation AI consulting engagements—company research, discovery, maturity assessment, opportunity prioritization, and proposal/SOW/roadmap creation—and use only for advisory work, not implementation.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | ai-consultant |
| description | Run pre-implementation AI consulting engagements—company research, discovery, maturity assessment, opportunity prioritization, and proposal/SOW/roadmap creation—and use only for advisory work, not implementation. |
Guide a seasoned AI consultant through structured client engagements — from initial research through deliverable generation — with emphasis on generative AI and AI agents.
This skill provides a repeatable engagement framework for AI consulting. It assumes you are an experienced AI/ML practitioner who needs structured guidance on the consulting process: who to talk to, what to ask, how to assess readiness, how to prioritize opportunities, and how to package findings into professional deliverables.
The skill does NOT teach you AI — it teaches you how to run an AI consulting engagement systematically so nothing falls through the cracks.
Do NOT use when:
Always structure the final response with these top-level sections, in this order:
None if nothing remains.Rules:
None when a section does not apply.Before any client conversation, build a briefing. This prevents wasting discovery time on publicly available information.
Save this as company-briefing.md. This becomes your conversation anchor.
Score the client across 5 dimensions using the framework in references/ai-maturity-model.md. Each dimension is rated 1-5.
| Dimension | What you're assessing |
|---|---|
| Data | Quality, accessibility, governance, cataloging |
| Infrastructure | Cloud readiness, compute, MLOps, API architecture |
| Talent | AI/ML skills, data literacy, willingness to upskill |
| Governance | AI policy, ethics framework, risk management, approval |
| Culture | Innovation appetite, change tolerance, exec sponsorship |
Produce an ai-maturity-assessment.md with scores, evidence, and
recommendations per dimension. The aggregate score determines engagement
approach:
Use the question bank in references/question-bank.md to guide interviews. The order matters — start at the top for strategic context, then drill down.
Interview sequence:
For each interview, capture:
Produce a stakeholder-discovery.md summarizing findings across all interviews.
Before investing time in opportunity mapping, check for engagement killers. These are conditions that make AI initiatives likely to fail regardless of technical approach.
| Red Flag | Severity | Mitigation |
|---|---|---|
| No executive sponsor | Critical | Pause engagement until sponsor identified |
| No data strategy or data is siloed/inaccessible | High | Recommend data foundation work first |
| Unrealistic expectations ("AI will 10x revenue in 3 months") | High | Reset expectations with industry benchmarks |
| Political resistance from key stakeholders | High | Identify champions, propose change management |
| No budget allocated or "we'll figure out budget later" | Medium | Get budget range commitment before scoping |
| Previous failed AI initiative with unresolved blame | Medium | Address explicitly — diagnose what went wrong |
| Compliance/legal has not been consulted | Medium | Loop them in immediately |
| No plan for GenAI-specific risks (hallucination, data leakage, prompt injection) | Medium | Require GenAI risk assessment before any GenAI deployment |
| Paying for enterprise AI licenses nobody uses | Medium | Audit utilization — quick win may be adoption, not new tooling |
Document flags found in red-flags.md with specific mitigation plans. Also
include a "Positive Indicators" section documenting favorable conditions (e.g.,
strong exec sponsor, data team enthusiasm) — because knowing what IS working
is as valuable as knowing what isn't. If critical red flags exist, address them
before proceeding to Phase 5.
Map every AI opportunity discovered during stakeholder interviews. For each opportunity, score on two axes:
Business Impact (1-5):
Feasibility (1-5):
Plot opportunities on an impact-vs-feasibility matrix:
| High Feasibility | Low Feasibility | |
|---|---|---|
| High Impact | Quick Wins — do these first | Strategic Bets — plan carefully |
| Low Impact | Fill-ins — only if resources allow | Deprioritize |
For every opportunity in the matrix, include these three assessments (not just top priorities — all scored opportunities need them because they inform comparison):
Produce opportunity-matrix.md with scored and prioritized opportunities.
Generate the appropriate deliverables based on engagement scope. Use templates in references/deliverable-templates.md.
Core deliverables (always produce):
Engagement-dependent deliverables (produce when scoping specific projects):
Stakeholder-specific packaging:
Check references/industry-playbooks.md for vertical-specific guidance. High-level playbooks are available for:
For unlisted industries, apply the general framework and research industry-specific AI adoption patterns via web search.
| Mistake | Fix |
|---|---|
| Skipping pre-engagement research and wasting discovery on public info | Always do Phase 1 first — clients respect prepared consultants |
| Jumping straight to solutions without understanding maturity | A company at maturity level 1 needs data foundations, not an AI agent fleet |
| Treating all stakeholders the same | CFO wants ROI, CTO wants architecture, ops wants workflow impact — tailor messaging |
| Ignoring red flags to keep the engagement moving | Address them early — failed projects destroy consulting relationships |
| Proposing GenAI for everything | Traditional ML or even simple automation may be more appropriate for many use cases |
| Scoping too many initiatives at once | Recommend 1-2 quick wins plus 1 strategic bet. Prove value, then expand |
| Not including change management in the SOW | AI projects fail more from adoption issues than technical ones |
| Vague SOW scope that invites scope creep | Be explicit about inclusions, exclusions, and assumptions |
| Scoping without understanding the client's capacity | Get budget and team signals early — don't propose 5,000 engineer-hours to a team with no AI staff |
| Ignoring existing AI tool investments | Always inventory what's already licensed — configuring an underused Copilot deployment is cheaper than building from scratch |
| Writing deliverables in AI jargon | Match the audience — execs want business outcomes, not model architectures |
Research before you ask — never waste a client's time asking questions you could have answered with 10 minutes of web research. It signals preparation and builds trust immediately.
Maturity determines strategy — the right AI recommendation depends entirely on where the client is today. A maturity-1 company needs data foundations; a maturity-4 company needs optimization and differentiation. Match your recommendations to their reality.
Quick wins fund strategic bets — always identify at least one high-impact, high-feasibility opportunity that can show value in 30-60 days. This builds credibility and budget for larger initiatives.
Deliverables are stakeholder-specific — the same findings need different packaging for different audiences. A board deck is not a technical architecture document. Write for your reader, not for yourself.
Red flags are gifts — discovering engagement risks early is valuable, not discouraging. Addressing them proactively differentiates good consultants from bad ones.
Run Microsoft SkillOpt experiments for agent skills, prompts, or instruction files with clean train/validation/held-out splits, adapter setup, bounded rollouts, artifact collection, and honest reporting. Use when the user wants to use SkillOpt, optimize a skill with SkillOpt, benchmark prompt variants, run multi-epoch skill optimization, check held-out quality, or diagnose SkillOpt runs.
Guide annuity and life-insurance sales workflows—discovery, suitability, product selection, objections, closing, and servicing—when supporting licensed insurance sales conversations or deliverables.
Generate API reference documentation from source code—endpoints, schemas, auth, errors, and examples—when the task is to document an existing API rather than design or implement one.
Turns an approved build spec into a minimal technical implementation plan, including system components, human-in-the-loop boundaries, workflows, and delivery milestones. Use when an internal automation, concierge workflow, or lightweight product has already been chosen and now needs a concrete build plan.
Generate release notes and changelogs from git history with SemVer classification, grouped changes, and migration notes when documenting what changed between versions or refs.
Screens business ideas, offers, and product plans for employer IP, confidentiality, regulatory overreach, sensitive-data misuse, and unsafe claims, then proposes safer clean-room alternatives. Use when a user's edge may depend on employer information, regulated workflows, customer data, or legally sensitive positioning.