Magic Paste

Workflow

Turn Slack rants, email threads, and meeting notes into structured decisions.

What is Magic Paste?

Magic Paste uses AI to extract structured decisions from unstructured text. Instead of filling out forms, you paste raw content and Arbtr figures out the structure.

It handles:

Slack threads — Copy-paste a heated discussion
Email threads — Forward-style chains work great
Meeting notes — Messy bullet points are fine
RFC drafts — Markdown or plain text
ADRs — Existing Architecture Decision Records

How to Use

Open Magic Paste

Click + New Decision in the header, then select "From Text (Magic Paste)".

Paste Your Content

Paste any text that contains a decision. The messier, the better—Magic Paste handles noise.

Example input:

From: sarah@company.com
To: engineering@company.com
Subject: Re: Re: Database decision

Hey team,

After the outage last week, I think we need to revisit the database
choice. MongoDB's lack of transactions is killing us.

Options I see:
1. PostgreSQL - proper ACID, good JSON support
2. CockroachDB - distributed Postgres-compatible
3. Stick with MongoDB - add application-level locking

I'm leaning PostgreSQL. Thoughts?

---
Mike replied:
+1 for Postgres. The JSON support is actually really good now.

---
Lisa replied:
CockroachDB seems overkill. We're not that big yet.
Let's do Postgres and revisit if we need horizontal scaling.

Review and Edit

Arbtr extracts the structure. Review and adjust:

Title: Database migration from MongoDB
Positions: PostgreSQL, CockroachDB, Stay with MongoDB
Arguments: ACID compliance, JSON support, scaling concerns
Status: Under Discussion (detected as ongoing)

Save

Click "Create Decision" to save. You can optionally save the original source text for reference.

What Gets Extracted

Always extracted

• Decision title
• Positions (options being considered)
• Context summary

When detected

• Arguments for/against positions
• Outcome (if decision was made)
• Author attributions
• Suggested tags
• Related decisions (by title match)

How It Works

Magic Paste uses structured extraction with JSON mode. The AI is prompted to output a validated schema, not free-form text.

Extraction Pipeline

Input validation: 50–50,000 characters accepted
Model selection: Uses purpose: 'high-quality' tier (GPT-4o or equivalent)
Structured output: JSON mode with temperature: 0.3 for consistency
Schema validation: Response validated against Zod schema
Team member matching: Author names fuzzy-matched to team profiles

DraftDecision schema

interface DraftDecision {
  title: string             // Extracted decision title
  context: string           // Background/problem statement
  status: "Active" | "Concluded"
  tags: string[]            // Suggested tags
  positions: DraftPosition[]
  arguments: DraftArgument[]
  outcome?: string          // If concluded, the rationale
  winningPositionLabel?: string
  confidence: number        // 0-100, extraction confidence
  relatedDecisions?: DraftRelationshipHint[]
}

interface DraftPosition {
  label: string             // Position name (e.g., "PostgreSQL")
  description: string       // What this option means
}

interface DraftArgument {
  claim: string             // The argument text
  positionLabel: string     // Which position this supports
  substantiation: "opinion" | "reasoned" | "data-backed"
  evidence: string[]        // URLs or references
  strength: number          // 1-10 strength rating
  authorName?: string       // Who made this argument
}

interface DraftRelationshipHint {
  reference: string         // Raw text reference (e.g., "./adr-001.md")
  suggestedType?: RelationshipType
  context?: string          // Sentence where reference appears
}

Failure Modes & Edge Cases

Input Rejected

<50 characters: "Input too short. Paste at least a few messages."
>50,000 characters: "Input too long. Try a shorter excerpt."

Extraction Limitations

No clear decision: If the text has no identifiable choice, extraction may produce a generic title or fail validation.
Single option: If only one option is mentioned, it becomes the sole position. Consider adding a "Status Quo" position manually.
Ambiguous outcome: If text implies a decision was made but isn't explicit, status defaults to "Active" for review.
Author matching: Names must closely match team member profiles. "Sarah" won't match "Sarah Chen" without fuzzy matching enabled.

Reproducibility

With temperature: 0.3, extraction is mostly deterministic but not guaranteed identical. The same input may produce slightly different argument phrasing or tag suggestions. Core structure (title, positions) is highly stable.

AI Configuration

Magic Paste uses the high-quality model tier for complex extraction tasks. By default, Arbtr provides AI with usage limits. For higher limits or to use your own provider, configure BYOK.

Bring Your Own Key (BYOK)

For higher usage limits or to keep AI requests on your own account:

Go to Team Settings → AI Configuration
Add your OpenAI or Gemini API key
Select your preferred model

Model Selection

• GPT-4o: Best quality, higher cost (~$5/1M tokens)
• GPT-4o-mini: Good balance (~$0.15/1M tokens)
• Gemini Pro: Fast, competitive pricing

Typical Token Usage

• Short email: ~500 input, ~300 output tokens
• Slack thread: ~2,000 input, ~500 output tokens
• Long RFC: ~10,000 input, ~800 output tokens

AI usage is logged per-team. View token consumption in Team Settings → AI Usage.

Tips for Best Results

Good inputs

• Discussions with clear options mentioned
• Text that includes reasoning (why X over Y)
• Multiple perspectives/voices
• Technical context about constraints

Poor inputs

• Single-sentence statements
• Text without any options or alternatives
• Pure status updates (no decision content)
• Highly redacted/anonymized text

Importance

Bulk Import

Command Palette