context-builder（上下文构建器）

Context Builder

Generate interactive context-building prompts for consulting clients. These prompts are designed to be run in Claude Code -- they guide a team through structured questions using AskUserQuestion, generate output files per section, and compile everything into a reusable CLAUDE.md.

Workflow

Phase 1: Intake (AskUserQuestion)

Ask all intake questions using AskUserQuestion with closed-list options. Gather:

Question 1: Company identifier

• Options: "I have a website URL", "I have a company name", "I have both"
• Follow up to get the actual URL/name

Question 2: Who will use this prompt?

• Options: "Specific person (name + role)", "A team (no specific person)", "Unknown / TBD"
• If specific person: follow up for name and role

Question 3: Primary consulting focus (multiSelect)

• "AI automation of current operations"
• "Existential strategy (what survives AI)"
• "New business models / pivots"
• "Product development with AI"

Question 4: Industry

• "Marketing / Advertising"
• "Manufacturing / Construction"
• "SaaS / Software"
• "Professional Services / Consulting"
• (Other)

Question 5: Existing context in vault?

• "Yes, there's a call transcript"
• "Yes, there are notes/files"
• "No existing context"
• If yes: ask for filename or search term to locate it

Question 6: Session language

• "Russian (questions in Russian, output in English)"
• "English throughout"
• "Other"

Phase 2: Research (automated)

Run these research steps in parallel where possible:

1. Web research: Use WebSearch and WebFetch (via Task agent) to gather:

- What the company does, products/services - Target market, company size, geography - Tech stack, partnerships - Recent news, funding, team info - Competitive landscape

1. Vault search: Search the Obsidian vault for:

- Transcripts mentioning the company name (Grep in vault root and Daily/) - People files for contacts at the company (People/ folder) - Any existing notes or research

1. Transcript analysis (if found): Extract from call transcripts:

- Team members and their roles - Current AI tool usage - Pain points and concerns mentioned - Specific processes described - Questions raised by the team

Phase 3: Section Selection (AskUserQuestion)

Present a curated set of sections based on the consulting focus. Use AskUserQuestion with multiSelect to let the user pick which sections to include.

Section Library

Draw from references/section-library.md for the full section catalog. Default section sets by focus:

AI Automation focus:

1. Process Inventory, 2. Pain Points & Waste, 3. Current Tech Stack, 4. AI Opportunity Mapping, 5. People & Org, 6. Data Reality Check, 7. Quick Wins

Existential Strategy focus:

1. Revenue & Service Map, 2. The Existential Question, 3. Client Value Chain, 4. New Business Models, 5. Data & Knowledge Assets, 6. People & Org, 7. Quick Wins & Pilots

Full Assessment (both): All 10 sections from the library.

After section selection, ask:

Express mode grouping: Present a suggested grouping of selected sections into 4 Express mega-sections. Let user confirm or adjust.

Phase 4: Generation

Generate two files:

1. The Context-Builder Prompt

Save to: Claude-Drafts/{company-slug}-context-prompt.md

Structure (follow the template in references/prompt-template.md):

---
created_date: '[[YYYYMMDD]]'
type: draft
topic: consulting, AI transformation, {industry}
for: {contact person or team name}
---

# AI Transformation Context Builder -- {Company Name}

## About {Company}
  [Generated from research -- company description, size, market, positioning]

## Current State
  **What's working:** [from research + transcript]
  **The gap:** [from research + transcript]
  [If existential concerns found: **Existential context:**]

## Mode Selection
  [Express vs Deep Dive with section descriptions]

## How This Works
  [Standard interactive session instructions]

## Session Resumability
  [Standard resumability logic]

## Interactive Flow
  [Selected sections with tailored questions]

## Output Files
  [One file per section + final CLAUDE.md]

## Relevant Frameworks
  [Selected from references/frameworks.md based on focus]

2. Instruction File (optional)

If the prompt will be sent to someone external, generate a short instruction file: Claude-Drafts/{company-slug}-context-instructions.md

Containing:

• What this file is and how to use it
• Prerequisites (Claude Code or similar)
• The two modes explained simply
• What they'll get on output
• Privacy note (they can share as much or as little as they want)

Phase 5: Delivery (AskUserQuestion)

Question: What to do with the generated files?

• "Save to vault only"
• "Save and send via Telegram"
• "Save and let me review first"

If Telegram: ask for the recipient handle/name, then send using the telegram skill (intro message + file).

Key Principles

• Maximize closed-list questions: Every AskUserQuestion should have concrete options. Minimize free-text input.
• Research before asking: Don't ask the user things that can be found via web search or vault search.
• Tailor sections to context: If the transcript reveals specific concerns (e.g., existential fears, specific tech stack), customize the section questions to reference those specifics.
• Bake in discovered context: The generated prompt's "About" and "Current State" sections should be rich with researched details so the person running the prompt gets a warm start.
• Language awareness: If session language is Russian, all AskUserQuestion interactions during prompt execution should be in Russian, but output files in English.

Resources

references/

• section-library.md -- Full catalog of available sections with question templates
• prompt-template.md -- Structural template for the generated prompt
• frameworks.md -- Consulting frameworks to selectively include