Removendo ambiguidades das solicitações

O que a ambiguidade causa na IA

O clássico: 'Torne isto melhor'

import anthropic

client = anthropic.Anthropic(api_key='sk-ant-your-key-here')

original_text = (
    'Our software helps companies manage their data. '
    'It has many features. Customers like it a lot.'
)

# Ambiguous improvement request
vague = f'Make this better:\n\n{original_text}'

# Unambiguous improvement request
specific = (
    f'Rewrite this product description to be exactly 50% shorter, '
    f'more confident in tone, and replace vague phrases like "many features" '
    f'and "a lot" with specific claims. Do not add new features I have not mentioned.\n\n'
    f'{original_text}'
)

for label, prompt in [('VAGUE', vague), ('SPECIFIC', specific)]:
    response = client.messages.create(
        model='claude-opus-4-5',
        max_tokens=200,
        messages=[{'role': 'user', 'content': prompt}]
    )
    print(f'--- {label} ---')
    print(response.content[0].text)
    print()

Múltiplas interpretações: exemplos práticos

import openai

client = openai.OpenAI(api_key='sk-your-key-here')

# Disambiguating 'fix this code'
buggy_code = 'def divide(a, b): return a / b'

ambiguous = f'Fix this:\n{buggy_code}'

unambiguous = (
    f'Fix only the division-by-zero bug in this function. '
    f'Add a guard that raises a ValueError with message "b cannot be zero" when b=0. '
    f'Do not change anything else — keep the function signature and return type identical.\n\n'
    f'{buggy_code}'
)

response = client.chat.completions.create(
    model='gpt-4o',
    messages=[{'role': 'user', 'content': unambiguous}]
)
print(response.choices[0].message.content)

Técnica de desambiguação: declarar a interpretação

import anthropic

client = anthropic.Anthropic(api_key='sk-ant-your-key-here')

response = client.messages.create(
    model='claude-opus-4-5',
    max_tokens=256,
    messages=[{
        'role': 'user',
        'content': (
            'Edit the following paragraph. '
            'When I say "edit", I mean: fix grammar and punctuation only. '
            'Do NOT change vocabulary, sentence structure, or content. '
            'When done, list each change you made in a numbered list below the edited text.\n\n'
            'The team have went to the meeting early, but the manager '
            'werent there so they waited for alot of time before leaving.'
        )
    }]
)
print(response.content[0].text)

Técnica de desambiguação: definir o escopo

import openai

client = openai.OpenAI(api_key='sk-your-key-here')

code_block = '''
def calculate_tax(income, rate):
    return income * rate

def calculate_net(income, tax):
    return income - tax
'''

response = client.chat.completions.create(
    model='gpt-4o',
    messages=[{
        'role': 'user',
        'content': (
            'Add Python type hints to the following code.\n'
            'IN SCOPE: adding type hints to parameters and return values only.\n'
            'OUT OF SCOPE: changing function names, logic, docstrings, or formatting.\n'
            'Do not add any comments or docstrings.\n\n'
            + code_block
        )
    }]
)
print(response.choices[0].message.content)

Técnica de desambiguação: especificar o formato da saída

import anthropic

client = anthropic.Anthropic(api_key='sk-ant-your-key-here')

candidates_text = (
    'Alice: 5 years Python, worked at Stripe, has a CS degree.\n'
    'Bob: 3 years JavaScript, worked at a startup, self-taught.\n'
    'Carol: 8 years Java, worked at Google, has an MS in CS.'
)

response = client.messages.create(
    model='claude-opus-4-5',
    max_tokens=300,
    messages=[{
        'role': 'user',
        'content': (
            'Extract the candidate data below into a JSON array. '
            'Each object must have exactly these keys: name, years_experience, primary_language, '
            'previous_employer, education_level. '
            'education_level values: DEGREE, MASTERS, SELF_TAUGHT.\n\n'
            + candidates_text
        )
    }]
)
print(response.content[0].text)

Técnica de desambiguação: pedir esclarecimentos ao modelo

import openai

client = openai.OpenAI(api_key='sk-your-key-here')

response = client.chat.completions.create(
    model='gpt-4o',
    messages=[
        {
            'role': 'system',
            'content': (
                'You are a professional copywriter. '
                'Before starting any writing task, ask exactly 3 clarifying questions '
                'that would most improve the output quality. '
                'Only proceed to write after the user answers those questions.'
            )
        },
        {
            'role': 'user',
            'content': 'Write a landing page headline for my business.'
        }
    ]
)
print(response.choices[0].message.content)

Termos relativos que precisam de ancoragem

import anthropic

client = anthropic.Anthropic(api_key='sk-ant-your-key-here')

# Relative terms mapped to concrete equivalents
mapping = [
    ('Give me a brief summary',   'Summarize in exactly 2 sentences'),
    ('Write something formal',    'Write using no contractions, no first person, and Flesch-Kincaid grade 12+'),
    ('Make it simple',            'Use only words a 10-year-old would know; max 15 words per sentence'),
    ('Be comprehensive',          'Cover at least 5 distinct subtopics with one example each'),
]

for vague, concrete in mapping:
    print(f'Vague:    "{vague}"')
    print(f'Concrete: "{concrete}"')
    print()

Ambiguidade do sujeito: quem executa a ação?

import openai

client = openai.OpenAI(api_key='sk-your-key-here')

# Clear subject reference prevents wrong-section edits
contract_text = (
    '## Section 1: Payment Terms\nPayment is due within 30 days.\n\n'
    '## Section 2: Termination\nEither party may terminate with 30 days notice.'
)

response = client.chat.completions.create(
    model='gpt-4o',
    messages=[{
        'role': 'user',
        'content': (
            'In the contract below, rewrite ONLY Section 2 (Termination). '
            'Change the notice period from 30 days to 90 days. '
            'Do not change Section 1 or any other text.\n\n'
            + contract_text
        )
    }]
)
print(response.choices[0].message.content)

Ambiguidade temporal: quando?

import anthropic
from datetime import date

client = anthropic.Anthropic(api_key='sk-ant-your-key-here')

today = date.today().isoformat()

response = client.messages.create(
    model='claude-opus-4-5',
    max_tokens=256,
    system=f'Today is {today}. Use this as the reference for any time-related words.',
    messages=[{
        'role': 'user',
        'content': (
            'Describe the best practices for Python async programming '
            'as of January 2025. If your knowledge does not cover this timeframe, '
            'say so explicitly and share what you know up to your cutoff.'
        )
    }]
)
print(response.content[0].text)

Criar um radar de ambiguidade

def scan_prompt_for_ambiguity(prompt):
    '''Simple heuristic scanner for common ambiguity patterns.'''
    warnings = []
    vague_verbs = ['fix', 'improve', 'make it', 'update', 'change it', 'redo']
    relative_adj = ['better', 'shorter', 'longer', 'formal', 'simple', 'recent', 'comprehensive']
    missing_format = ['json', 'table', 'list', 'bullet', 'paragraph', 'word', 'sentence']

    lower = prompt.lower()
    for v in vague_verbs:
        if v in lower:
            warnings.append(f'Vague verb detected: "{v}" — specify what change exactly')
    for a in relative_adj:
        if a in lower:
            warnings.append(f'Relative adjective: "{a}" — anchor with a measurable definition')
    if not any(f in lower for f in missing_format):
        warnings.append('No output format specified — add format, length, or structure')
    return warnings

test = 'Make the report better and more formal.'
print('Prompt:', test)
for w in scan_prompt_for_ambiguity(test):
    print(' WARNING:', w)

Verificação de conhecimento

Eliminar a ambiguidade — recapitulação