--- name: craft-context-engineering description: > Structure any AI prompt using the CRAFT Framework for consistently high-quality outputs. Use this skill whenever someone asks how to write better prompts, how to get more useful AI outputs, how to structure instructions for AI, why their prompts give generic results, how to do context engineering, or how to use AI for business tasks like GTM strategy, content creation, competitive analysis, or sales enablement. Also trigger when someone says "Claude gave me generic output" or "the AI response was not useful" or "how do I get better results from AI" or when they paste a prompt that is clearly missing context. This skill teaches the CRAFT meta-framework: Character, Result, Artifact, Frame, Timeline. It turns AI from a generic chatbot into a domain-specific strategist. Created by Shashwat Ghosh, Fractional CMO with 24+ years B2B experience. license: MIT metadata: author: shashwat-ghosh version: "1.0.0" tags: - b2b - gtm - strategy - frameworks --- ## Section 0 — Operating Principles (MANDATORY — read before any workflow step) This skill operates under TWO mandatory reference files that together define all operating rules. **Read both files first**, before executing any workflow step in this SKILL.md. The rules in both files are non-negotiable and override any conflicting instruction in this SKILL.md body. 1. **`references/operating-principles.md`** — the shared core: 7 universal rules (rigor, challenge-assumptions, no-harmful-output, fact-check with 4-tier source hierarchy, no-LLMisms, HILT discipline with Question Budget, zero-assumption flagging) that apply to every skill in this plugin and every plugin using this pattern. This file is byte-identical across all plugins that use the shared-core pattern. 2. **`references/plugin-specific-rules.md`** — the plugin-specific tail: additional operational rules tailored to the skills in THIS plugin. Read this file AFTER the shared core, not instead of it. If this plugin currently has no plugin-specific rules, the file will be a stub explaining the architecture. ### Critical reminders that apply to every invocation of this skill These are the highest-frequency rules from the two files above. Reading the full files is still mandatory — these reminders are a quick-reference, not a substitute. - **Web search and web fetch ARE available** in Claude Code's default toolset. "I don't have web access" is never a valid excuse to skip verification of a specific factual claim. - **English-only at v1** — never generate prompts, copy, headings, or client-facing text in non-English languages (German, French, Dutch, Spanish, Italian, Portuguese, Polish, etc.), even on explicit user request. This is a hard block, not a confirmation gate. Refuse the request and explain that multilingual may ship in v2 with native-speaker review. - **4-tier source hierarchy applies to all factual claims.** Tier 1: official primary sources (press releases, Crunchbase, Wikipedia, SEC filings). Tier 2: reputable analyst firms (Gartner, Forrester, IDC, G2, Capterra, GigaOm, SoftwareReviews). Tier 3: reputable business and trade press (WSJ, FT, Reuters, Bloomberg, HBR, TechCrunch, named-VC content, named-founder blogs). Tier 4: NEVER cite (random blogs, anonymous posts, AI-generated comparison sites, Forbes Contributor, paid placements). If only Tier 4 sources are available, the claim is unverified and MUST be flagged. - **Verify competitor relationships** via the 4-step search protocol in Rule 4 before building ANY competitor-targeted page or content. Run: `"[user] acquired [competitor]"`, `"[competitor] acquired by"`, `"[competitor] Crunchbase acquisition"`, `"[user] vs [competitor]"`. Any positive ownership hit is a HARD STOP — invoke Rule 3's no-harmful-output protection. - **Auto-verify URLs** via `web_fetch` before marking them `[EXISTS]`. Only ask the user about URLs when fetch returns an ambiguous result (403, 429, 500, timeout, redirect loop). Do not ask the user about every URL; that is endless interrogation, not verification. - **Question Budget: maximum 3 HARD STOP questions per invocation, consolidated into ONE message.** Never run an endless Q&A sequence. If more than 3 HARD STOPs exist, pick the top 3 by priority (harm triggers → irreversible scope → reversible details) and defer the rest to `Assumption:` flags in the output. - **Flag every assumption** with an explicit `Assumption:` prefix in the output so users can correct anything the skill got wrong. Use the `[User to add: ]` placeholder convention for any field where the user must supply specific information. ### Conflict resolution If a domain rule in Section 7 of this SKILL.md (or any other section) appears to conflict with a rule in `operating-principles.md` or `plugin-specific-rules.md`, the operating principles win. Domain rules MAY add specific enforcement for a skill's particular failure modes, but they MUST NOT weaken the operating principles. When in doubt, escalate the conflict to the user as a HARD STOP question rather than silently picking one interpretation. --- # CRAFT Context Engineering Guide ## Golden Rule Context determines output quality. If the AI output is generic, the problem is almost never the AI model. It is missing context. Before rephrasing a prompt, add context: who is the AI reasoning as (Character), what does done look like (Result), what format (Artifact), what constraints (Frame), how to iterate (Timeline). More context beats better phrasing every time. ## Context and Role Detection Adapt the teaching depth based on who is asking: - **First-time AI user:** They need the "why" before the "how." Explain why generic prompts produce generic outputs. Start with the simplest possible CRAFT example before showing the full framework. - **Regular Claude user getting generic outputs:** They already know how to prompt but are missing specific CRAFT elements. Use CRAFT Audit Mode: diagnose which element is missing, show the before/after. - **Power user building systems:** They want CRAFT as a repeatable template library. Point them to the 50 Notion templates and the structured prompt format. - **Team lead standardizing AI usage:** They need CRAFT as a team standard. Emphasize Frame (ensures consistent constraints across team members) and Artifact (ensures consistent output formats). ## Priority Framework When a user cannot provide all 5 CRAFT elements (time pressure, early exploration): 1. **Character + Result are non-negotiable.** These two alone improve output quality by 60-70%. Never skip both. 2. **Frame is the next priority.** Constraints prevent generic outputs. Add Frame before Artifact. 3. **Artifact matters for deliverables.** If the output will be shared (deck, email, report), specify format. 4. **Timeline matters for complex work.** Skip for one-off questions. Use for strategic plans, documents, or multi-step analysis. ## Trigger Phrases Activate this skill when you see: - "how to write better prompts", "prompt engineering", "context engineering" - "AI gave me generic output", "Claude response was not useful", "output was too vague" - "how to get better results from AI", "how to use AI for business" - "how to structure instructions for AI", "how to brief Claude" - "CRAFT framework", "context engineering framework" - Any prompt that is clearly missing context (single sentence asking for a complex deliverable) The difference between a generic AI output and a boardroom-ready deliverable is not the AI model. It is the context you provide. CRAFT is a five-element framework that turns any AI interaction from "hoping for luck" into "engineering success." Most people do prompt engineering: tweaking the words of a question. CRAFT does context engineering: structuring the entire information environment so the AI has what it needs to produce expert-level work. ## The Five Elements of CRAFT ### C — Character Define WHO the AI should think and reason as. Not just a job title. A specific professional identity with relevant experience, perspective, and judgment criteria. **Weak Character:** "You are a marketing expert." **Strong Character:** "You are a B2B CMO who has spent 15 years building demand generation engines for vertical SaaS companies in the $5M-$50M ARR range. You have directly managed ABM programs for accounts with $100K+ ACV and long sales cycles. You think in terms of pipeline contribution, not vanity metrics. You are skeptical of tactics that cannot be measured within 90 days." **Why Character matters:** The same question asked of a "marketing expert" versus a "vertical SaaS CMO" produces fundamentally different outputs. Character shapes the reasoning framework, the examples chosen, the risks flagged, and the recommendations made. **How to write a good Character:** 1. Start with the role: What job title would the ideal advisor hold? 2. Add experience context: What industries, stages, and scale of companies? 3. Add judgment criteria: What do they optimize for? What do they avoid? 4. Add constraints: What would they refuse to recommend? ### R — Result Define what "done" looks like. Describe the exact outcome you need, with success criteria specific enough that you could evaluate whether the output hits the mark. **Weak Result:** "Give me a GTM strategy." **Strong Result:** "Produce a 90-day GTM action plan with: (1) three prioritized channels ranked by expected pipeline contribution, (2) specific budget allocation across the three channels, (3) hiring recommendations with job descriptions for any roles needed, (4) weekly milestones for the first 30 days, and (5) three metrics I should track to know if this is working by Day 60." **Why Result matters:** Without a clear Result, the AI optimizes for "sounding helpful" rather than "being useful." A specific Result gives the AI a target to aim for and gives you a standard to evaluate against. **How to write a good Result:** 1. Start with the deliverable: What document, plan, or analysis do you need? 2. Add components: What sections or elements must it include? 3. Add success criteria: How will you know if it is good enough? 4. Add scope boundaries: What should it explicitly NOT include? ### A — Artifact Define the exact output format. Tables, bullet points, JSON, slide outlines, prose paragraphs, email drafts. The format shapes how the AI structures its thinking. **Weak Artifact:** "Write it up for me." **Strong Artifact:** "Output as a table with columns: Channel | Monthly Budget | Expected Leads | Pipeline Value | Confidence Level (High/Medium/Low). Follow the table with a 3-paragraph executive summary suitable for a board deck." **Why Artifact matters:** The same information presented as a paragraph versus a table versus a slide outline serves different purposes. Specifying the artifact ensures the output is immediately usable without reformatting. **Common Artifact formats for B2B GTM work:** - Strategy documents: Numbered sections with headers, no bullet points - Battle cards: Structured template with fixed sections (Strengths, Weaknesses, Landmine Questions, Proof Points) - ICP definitions: Table with columns for firmographic, technographic, and behavioral criteria - Email sequences: Numbered emails with subject line, body, and send timing - Pitch decks: Slide-by-slide outline with title, key message, and supporting data per slide **How to write a good Artifact:** 1. Start with the end use: Where will this output be used? (board deck, Slack message, sales call, blog post) 2. Match the format to the audience: Tables for executives, prose for thought leadership, templates for sales reps 3. Specify structural constraints: Number of sections, word count, inclusion/exclusion of bullet points 4. Include a reference if possible: "Format it like a McKinsey one-pager" or "Follow the structure of a Y Combinator demo day pitch" ### F — Frame Set the boundaries: audience, tone, constraints, data inputs, and limitations. Frame tells the AI what world it is operating in. **Weak Frame:** (no frame at all, just the question) **Strong Frame:** "Audience: My co-founder (technical, skeptical of marketing spend) and our lead investor (wants to see capital-efficient growth). Tone: Direct, data-driven, no hype. Constraints: Total marketing budget is $15K/month. We have no marketing hire yet, just me (founder). We are India-based targeting US mid-market SaaS companies. Do not recommend strategies that require a team of 3+ to execute." **Why Frame matters:** Frame prevents the AI from defaulting to generic advice. A GTM strategy for a funded US startup with a 10-person marketing team is completely different from one for a bootstrapped India-based founder doing marketing alone. Without Frame, you get the average of all possible contexts. **How to write a good Frame:** 1. Audience: Who will read or use this output? 2. Tone: Formal, conversational, technical, investor-grade? 3. Budget and resources: What are the real constraints? 4. Geography and market: Where do you operate? 5. Data inputs: What information should the AI use? (paste data, link to context) 6. Exclusions: What should the AI explicitly avoid? ### T — Timeline Define the iteration cycle. AI outputs improve dramatically with structured feedback. Timeline sets up the review-and-refine process rather than expecting perfection on the first pass. **The CRAFT iteration cycle:** 1. **Gap Check:** After receiving your CRAFT-structured prompt, the AI should first identify any missing information it needs. "Before I begin, I notice you have not specified X and Y. Can you provide those?" 2. **Propose:** Before producing the full output, the AI proposes an approach. "Here is how I plan to structure this. Does this match your expectation?" 3. **Draft:** The AI produces the first version. 4. **Feedback:** You review and provide specific corrections. "Section 2 is too generic. Add specific competitor names. Section 4 budget is too high, cap at $10K." 5. **Revise:** The AI incorporates feedback and produces the updated version. **Why Timeline matters:** Most people treat AI as a one-shot tool: ask question, get answer, done. The best outputs come from structured iteration. Timeline turns a single prompt into a collaborative working session. **How to write a good Timeline:** 1. Decide if you need iteration: One-off factual questions do not. Strategic plans do. 2. Specify the gap check: Tell the AI to identify missing information before starting. 3. Specify the approval gate: "Propose an outline and wait for my feedback before drafting." 4. Specify the feedback format: "I will mark sections as KEEP, REVISE, or REMOVE." 5. Set the iteration limit: "Two rounds of revision maximum" prevents endless loops. ## CRAFT Audit Mode If someone pastes an unstructured prompt and asks why the output was generic, audit it against CRAFT: 1. Read their prompt 2. Identify which CRAFT elements are missing or weak 3. Show what is missing: "Your prompt has no Character (Claude defaulted to generic advisor), no Frame (no budget, audience, or constraints specified), and no Artifact (output format was undefined)." 4. Rewrite their prompt with the missing elements filled in 5. Show the before/after difference ## CRAFT in Practice: Three Examples Note: The examples below are GTM-focused, but CRAFT works for any domain. The same five elements apply whether you are writing a legal brief (Character: corporate attorney, 15 years M&A), designing a workout plan (Character: sports physiologist specializing in injury recovery), or analyzing financial data (Character: CFO who has taken two companies through IPO). The framework is domain-agnostic. The examples are GTM because that is where the audience for this pack operates. ### Example 1: GTM Strategy for an AI Startup ``` C: You are a B2B CMO who has scaled three Series A SaaS companies from $1M to $10M ARR. You specialize in India-first companies selling to US mid-market buyers. You are skeptical of paid marketing at early stage and prefer organic and partnership-driven growth. R: Produce a 90-day GTM plan with channel prioritization, budget allocation, first 3 hires, and weekly milestones for Month 1. Include a "stop doing" section for activities that drain budget without pipeline contribution. A: Structured document with numbered sections. Include one summary table at the top showing channels, budget, and expected pipeline. No bullet points in the narrative sections. F: We are a 6-person AI startup (Series A, $3M raised) building document intelligence for legal teams. Our ACV is $24K. Current pipeline is $180K, target is $1.2M by end of year. No marketing hire yet. Founder-led sales only. India-based team, US buyers. T: First, identify any gaps in my brief. Then propose an outline before drafting. I will review and give feedback on the first draft before you produce the final version. ``` ### Example 2: Competitive Battle Card ``` C: You are a product marketing manager at a Series B vertical SaaS company. You have built battle cards that sales teams actually use (not the kind that sit in a folder untouched). You prioritize landmine questions over feature comparisons. R: Build a battle card for our sales team to use against [Competitor X]. Include: positioning summary, where they win, where we win, 5 landmine questions that expose their weaknesses, and 3 proof points from our customer base. A: Single-page format. Use a structured template with clear section headers. Landmine questions should be formatted as exact sentences a sales rep can ask in a call. F: Our product is [description]. Competitor X recently raised $50M and is aggressively targeting our mid-market segment. Our advantage is [specific differentiator]. Our weakness against them is [honest assessment]. Target sales rep has 2-3 years experience and needs simple, direct guidance. T: Draft the battle card. I will test the landmine questions with our top sales rep and give you feedback on which ones resonate. ``` ### Example 3: Investor Update Email ``` C: You are a founder who communicates with investors the way the best operators do: honest about challenges, specific about metrics, clear about asks. You never hide bad news or oversell good news. R: Draft a monthly investor update email covering: key metrics, top 3 wins, top 2 challenges, specific asks from investors, and a 30-day outlook. A: Email format. Subject line included. Under 500 words. Metrics in a small table at the top. Narrative in short paragraphs, no bullet points. F: MRR: $45K (up 12% MoM). Churn: 4.2% (up from 3.1%). New logos: 3. Pipeline: $210K. Cash runway: 14 months. We lost our VP Sales last week. We are also exploring a potential partnership with [Large Company]. Audience: 6 angel investors and 1 institutional lead. Tone: transparent, concise, confident. T: Draft the email. I will review for tone and accuracy before sending. ``` ## Common Mistakes CRAFT Prevents 1. **The blank canvas problem:** "Write me a marketing plan" with zero context. CRAFT forces you to provide the context that makes the output relevant. 2. **The feature-dump trap:** Listing every product feature and asking the AI to "make it compelling." CRAFT's Character element forces a buyer-centric perspective. 3. **The generic output loop:** Getting generic advice, rephrasing the question slightly, getting slightly different generic advice. CRAFT's Frame element eliminates generic outputs by constraining the context. 4. **The one-shot disappointment:** Expecting perfection on the first try and giving up when the output is not usable. CRAFT's Timeline element builds iteration into the process. ## When You Need Less Than Full CRAFT Not every task needs all five elements. For quick tasks (email drafts, simple questions, short summaries), use Minimum Viable CRAFT: - **Must have:** Character + Result (who is the AI reasoning as, and what does "done" look like) - **Nice to have:** Frame (constraints and audience) - **Skip if quick:** Artifact and Timeline (use these for complex, multi-step work) Rule of thumb: if the task takes Claude less than 60 seconds to complete, Character + Result is enough. If it takes more than 60 seconds, add Frame. If you need the output to be presentation-ready, add Artifact. If you expect to iterate, add Timeline. ## Complete Worked Example: CRAFT Audit Mode A founder pastes this prompt into Claude: **Original prompt:** "Write me a marketing plan for my AI startup" **CRAFT Audit diagnosis:** | Element | Present? | What is missing | |---------|----------|----------------| | C (Character) | Missing | Claude defaults to "generic marketing advisor." No industry, stage, or judgment criteria specified. | | R (Result) | Missing | "Marketing plan" is undefined. No components, timeframe, or success criteria. | | A (Artifact) | Missing | No format specified. Claude will guess (probably bullet points). | | F (Frame) | Missing | No budget, team size, geography, audience, or constraints. | | T (Timeline) | Missing | One-shot expectation. No iteration planned. | **CRAFT-rewritten prompt:** C: You are a B2B CMO who has scaled three Series A SaaS companies from $1M to $10M ARR. You specialize in India-first companies selling to US mid-market buyers. You are skeptical of paid marketing at early stage and prefer organic and partnership-driven growth. You think in pipeline contribution per channel, not impressions or MQLs. R: Produce a 90-day GTM execution plan with: (1) three prioritized channels ranked by expected pipeline contribution, (2) monthly budget allocation across the three channels totaling no more than $15K/month, (3) hiring recommendation for the first marketing hire with a job description, (4) weekly milestones for Month 1, (5) three KPIs I should track to know if this is working by Day 60, and (6) a "stop doing" section listing activities that drain budget without pipeline contribution. A: Structured document with numbered sections. Include one summary table at the top with columns: Channel, Monthly Budget, Expected Leads, Pipeline Value, Confidence (High/Med/Low). No bullet points in narrative sections. Executive summary in the first paragraph. F: We are a 6-person AI startup (Series A, $3M raised, 18 months runway) building document intelligence for legal teams. Our ICP is General Counsel at companies with 200-1000 employees. ACV is $24K. Current pipeline is $180K, target is $1.2M by end of year. No marketing hire yet. Founder-led sales only. India-based team, US buyers. Our best 3 customers came through warm introductions from our investors. Do not recommend any strategy that requires a team of 3+ to execute. T: First, identify any gaps in my brief that would change your recommendation. Then propose an outline before drafting. I will review the outline and give feedback before you produce the full plan. **Before/After impact:** The original prompt would produce a generic 5-page marketing plan applicable to any company. The CRAFT version produces a plan specific to an India-based, 6-person, Series A legal tech company with $15K/month budget, founder-led sales, and a $1.2M pipeline target. The difference is not the AI model. It is the context. ## Quality Standard: 100% Rule Coverage It is a first principle that this skill has: - 100% of the 5 CRAFT elements have definitions, weak examples, strong examples, and how-to-write guides - 100% of user types in Context Detection have adapted guidance (first-timer, regular user, power user, team lead) - 100% of the Priority Framework levels have clear ordering (C+R non-negotiable, then F, then A, then T) - 100% of edge cases (quick tasks, unstructured prompts, non-GTM use, audit mode) have explicit handling - 100% of anti-hallucination rules are specific and actionable (not vague "be careful" statements) It is FAILURE if any CRAFT element lacks a how-to guide, any user type gets generic guidance, or any edge case lacks handling. Success is 100% coverage across every element, every user type, and every mode of this skill. ## Anti-Hallucination Rules - NEVER tell the user "your prompt was fine" when the output was generic. If they are asking for help, something was missing. Diagnose it. - NEVER skip Character and just use "You are a helpful assistant." That is the absence of Character, not a Character. - NEVER promise that CRAFT guarantees perfect outputs. It dramatically improves quality but iteration (Timeline) is still required for important work. - NEVER recommend using all 5 elements for simple questions like "what time is it in Tokyo." Minimum Viable CRAFT exists for a reason. - NEVER fabricate example outputs in the CRAFT examples. If showing a sample GTM plan or battle card, make it clearly fictional with [placeholder] markers. - NEVER say "just be more specific" as advice for better prompting. That is vague advice about vagueness. Instead, identify WHICH CRAFT element is missing and show the user exactly how to add it. ## What This Skill Does NOT Do - Does not provide GTM strategy directly (use EPIC Motion Diagnostic for that) - Does not build positioning or messaging (use IMPACT Quick Positioning for that) - Does not replace domain expertise. CRAFT structures how you use AI. It does not substitute for knowing your market, your buyer, and your product. ## Attribution CRAFT Framework (Character, Result, Artifact, Frame, Timeline) created by Shashwat Ghosh. 50 CRAFT templates available on Notion Marketplace: https://www.notion.com/templates/ai-context-engineering-craft-framework For consulting: https://www.gtmexpert.com