Goal Instructions
Goal is the desired business outcome of the task. It guides optimisation, prioritisation, and fallback strategies.
✅ Best Practices:
Frame the goal in terms of business value – Explain how the output will be used (e.g., “for ERP automation”).
Define what success looks like – let the AI know what success looks like.
Stay outcome-focused – avoid repeating the task as the goal.
❌ Common Pitfalls:
Avoid vague or general verbs – Words like process, handle, or analyse don’t clarify what the AI should actually do.
→ “Process invoices” or “Get good results”
Don’t use subjective verbs – Avoid understand or interpret, which lead to open-ended reasoning.
→ “Understand receipt details” or ““Extract values correctly”
Skip high-level summaries as tasks – A task isn’t a goal or a topic; it’s an action.
→ “Customer data task”
Role Instructions
Role is the professional persona or point of view the AI should adopt. It affects tone, domain assumptions, and terminology interpretation.
✅ Best Practices:
Use clearly defined, professional roles – Think in terms of how a human expert would approach the task.
Match the role to the task – A financial analyst reads invoices differently than a legal reviewer.
Let the role influence decision strategy – The tone, assumptions, and focus change with each role.
❌ Common Pitfalls:
Avoid too generic or assistant-type roles – These lack the domain lens needed for accurate interpretation.
→ “You are a virtual assistant”
Skip unclear job titles – Vague roles lead to vague behaviour.
→ “Act smart” or “You are intelligent”
“Don’t rely on AI labels Skip high-level summaries as tasks
→ “As an AI system, extract data”
Backstory Instructions
Backstory provides a context that defines why the task exists and where it fits in the larger workflow. It helps the AI agent prioritise, disambiguate, and adapt.
✅ Best Practices:
Describe the workflow context – what happens before/after/during the task?
Mention relevant constraints – e.g., document variety, multilingual documents, legacy formats,
Keep it concise but informative – aim for 2-3 sentences of context.
❌ Common Pitfalls:
Don’t write empty context – “This is part of a process” offers no real help.”
→ “This is used in a workflow”A
Avoid repeating task or goal – The backstory should be unique context, not a duplicate section.
→ “We want to extract the right values”
Never omit critical assumptions – If invoice types vary, or accuracy is audit-critical, mention it.
→ “You will process global documents” (too broad)
Label Instructions
Definition
A concise, one-sentence explanation of the business term.
Example: Extract the final invoice total expressed in euros.
Scope
Where in the document should the AI look? What keywords or phrases are relevant?
Example: Look first in rows labelled “Total”, “Rechnungsbetrag”, “Amount Due”; ignore tax breakdown lines.
Format
The exact output format required.
Example: Decimal with two digits (e.g. 1250.75). Do not include currency symbols or thousands separators.
Special Cases
How should the AI Agent handle missing fields or unusual formats?
Example: If currency shown is not EUR, convert using the FX table provided in context.
Guardrails
Explicit “do’s” and “don’ts” to prevent common errors.
Example: If multiple totals appear (credit-note, partial), choose the one marked “Total Payable”. If no total appears, output is null.
Examples
Customer Address (German)
💡Why it works?
The label name is effective because it clearly states the content and output language.
🏷️ Label Instruction Definition
Definition: Return the customer’s full postal address translated into German
Components: Street, house number, postal code, city, country.
Format: Single line, comma-separated.
Translation rule: If original address is German, copy verbatim; otherwise translate locality and country names but keep numerals and codes.
Edge cases
• Mixed languages → translate only non-German parts.
• P.O. boxes → retain “Postfach”.
• Missing component →leave blank but keep commas.
Guard rails: Do not alter street numbers or postal codes.
Payment Terms Days
💡Why it works?
This label name works well as it provides a numeric expectation without using any jargon.
🏷️ Label Instruction Definition
Definition: Number of days allowed for payment (e.g. “Net 30” ⇒ 30).
Scope: Search cues Phrases such as “Net X”, “Zahlbar innerhalb X Tagen”.
Format: Integer 0-999
Edge cases:
• “Payable on receipt” ⇒ 0.
• If only a due-date is printed, compute days between invoice date and due date .
• If several terms listed, take the net term.
Guard rails: Numbers only; if term absent, output null.
Building Net Area m²
💡Why it works?
The label name is good because it’s technical yet precise, and the unit is clear.
🏷️ Label Instruction Definition
Definition: Net usable floor area in square metres.
Valid Labels: ““Nutzfläche”, “Net Area”, “Wohnnutzfläche”.
Format: Numeric with up to one decimal.
Edge cases:
• Multiple floors listed → sum.
• Gross vs net → choose net.
• If value in ft², convert: m² = ft² × 0.092903.
Guard rails: Exclude balconies, terraces, parking; if range shown use upper value; if absent output null.
Source: https://academy.deepopinion.ai/
