Agent Skills: Expert Consultation Document Creator

Expert Consultation Document Creator

Create comprehensive technical consultation documents by extracting code, diffs, and architectural context within LLM token limits (125KB).

Document Structure

Part 1: Problem Context (~15-25 KB)

1. Problem — Issue, errors, test failures
2. Our Solution — What was implemented and why
3. Concerns — Code smells, coupling, architectural questions
4. Alternatives — Other approaches, trade-offs

Part 2: Complete Architecture (~60-90 KB)

5. Architecture Overview — ASCII diagram, data flow, patterns
6. Components — Frontend, tests, controllers
7. Services — Implementation and interfaces
8. Models — Domain entities with relationships

Part 3: Expert Request (~5-10 KB)

9. Questions — Specific technical questions
10. Success Criteria — Requirements and priorities

Workflow

Step 1: Write Problem Context

Create a descriptive file like {topic}-consultation.md with the problem context, your solution, concerns, alternatives, and an architecture overview.

Step 2: Extract Code

Use the bundled extraction script with size tracking. The script accepts multiple files in one call — batch for efficiency.

node scripts/extract-code.js \
  --track-size --output=doc.md \
  --section="Core Files" \
  file1.ts file2.ts file3.ts \
  --section="Tests" \
  test1.ts test2.ts

File format options:

• Full file: src/Service.cs
• Line ranges: src/Service.cs:100-200 or src/Service.cs:1-30,100-150
• Git diff (per-file): src/Service.cs:diff or src/Service.cs:diff=master..HEAD

Git diff options (all changes):

• Staged changes: --staged
• Commit changes: --commit=abc123 or --commit=HEAD~1

Branch-level options (PR consultations):

• All branch changes: --branch-diff=master (auto-resolves merge base)
• Full source of new files: --new-files-from=master (syntax-highlighted, not diff format)

Prefer FULL files over chunks for better expert analysis. Use chunks only for very large files.

Step 3: Add Expert Request

Append an "Expert Guidance Request" section to the end of the document with:

• Questions — Numbered list of specific technical questions derived from the user's concerns
• Success Criteria — Bullet list of requirements and priorities
• "Please answer in English" — Always include as the final line

This section is critical — without it, the expert has no actionable request.

Step 4: Verify Size

Check the file size — it should be 100-125 KB. DO NOT read the full file back (exceeds context).

Code Extraction Examples

See EXAMPLES.md for detailed usage patterns.

With sections:

node scripts/extract-code.js \
  --track-size --output=doc.md \
  --section="What Changed" \
  src/Service.cs:diff \
  --section="Implementation" \
  src/Service.cs src/Model.cs

Using config file:

node scripts/extract-code.js \
  --config=extraction-plan.json

Staged or commit changes:

node scripts/extract-code.js \
  --staged --track-size --output=doc.md

node scripts/extract-code.js \
  --commit=abc123 --track-size --output=doc.md

PR consultation (single command for all branch changes):

node scripts/extract-code.js \
  --branch-diff=master --track-size --output=consultation.md

# With full source of new files (syntax-highlighted, not diff format)
node scripts/extract-code.js \
  --branch-diff=master --new-files-from=master \
  --track-size --output=consultation.md

Config File Format

{
  "output": "consultation.md",
  "trackSize": true,
  "sections": [
    {
      "header": "What Changed",
      "files": ["src/Service.cs:diff"]
    },
    {
      "header": "Core Implementation",
      "files": ["src/Service.cs", "src/Model.cs"]
    }
  ]
}

Code Inclusion Priority

Must include:

• Core interfaces/abstractions
• Modified/bug-fix code
• Domain models
• Key service methods
• Test examples

Skip if tight on space:

• Boilerplate
• Simple DTOs
• Repetitive test setups

Gotchas

• The 125KB limit is enforced by the extraction script — it warns at 100KB and 115KB, then exits at 125KB
• Always use --track-size to stay within the limit
• Don't read the completed document back into context — it's too large
• The script appends to existing files, so you can build incrementally
• Shell constructs like $() and variable assignments in Bash calls can trigger permission prompts. Prefer simple, single-command calls when possible.