Skip to main content
The config.json file controls which tweets get flagged for deletion. This guide shows practical examples for different scenarios.

​
Criteria structure

Every config.json follows this structure: 
{
  "criteria": {
    "forbidden_words": [],
    "topics_to_exclude": [],
    "tone_requirements": [],
    "additional_instructions": ""
  }
}

​
Criteria fields explained

Case-insensitive word list. Flags any tweet containing these exact words.
"forbidden_words": ["crypto", "NFT", "damn"]
  • Flags: “Crypto is the future!” ✓
  • Flags: “I’m damn tired” ✓
  • Keeps: “Cryptocurrency adoption” ✗ (different word)
High-level themes and topics. The AI interprets these broadly.
"topics_to_exclude": [
  "Political opinions",
  "Cryptocurrency hype"
]
The AI uses semantic understanding to catch variations.
Rules about how content should be expressed.
"tone_requirements": [
  "Professional language",
  "No sarcasm"
]
Focuses on delivery rather than topic.
Natural language instructions for the AI.
"additional_instructions": "Be extra cautious with tweets from 2020-2021"
Use this for nuanced, context-specific rules.

​
Example configurations

​
Conservative professional cleanup

For professionals who want to remove only obviously problematic content:
config.json
{
  "criteria": {
    "forbidden_words": [
      "damn",
      "wtf",
      "shit",
      "fuck"
    ],
    "topics_to_exclude": [
      "Explicit profanity",
      "Personal attacks or insults",
      "Discriminatory language"
    ],
    "tone_requirements": [
      "No aggressive or hostile language"
    ],
    "additional_instructions": "Only flag content that is clearly unprofessional or offensive"
  }
}
Use case: Job seekers who want to clean up obvious red flags without removing personality.

​
Aggressive career rebrand

For users completely rebranding their professional image:
config.json
{
  "criteria": {
    "forbidden_words": [
      "crypto",
      "NFT",
      "web3",
      "blockchain",
      "metaverse",
      "moon",
      "HODL",
      "gm",
      "wagmi"
    ],
    "topics_to_exclude": [
      "Cryptocurrency or blockchain",
      "NFT or digital art speculation",
      "Get-rich-quick schemes",
      "Financial advice or predictions",
      "Hype or promotional content",
      "Casual or unprofessional tone",
      "Personal life updates"
    ],
    "tone_requirements": [
      "Strictly professional language",
      "Informative and educational only",
      "No excitement or hype",
      "No slang or internet speak"
    ],
    "additional_instructions": "Flag anything related to crypto, NFTs, or Web3. Also flag casual tweets, jokes, or personal updates. Keep only educational, professional content."
  }
}
Use case: Former crypto influencer transitioning to traditional finance or enterprise software.

​
Academic researcher

For academics wanting to maintain scholarly credibility:
config.json
{
  "criteria": {
    "forbidden_words": [],
    "topics_to_exclude": [
      "Political partisan statements",
      "Controversial social issues",
      "Personal attacks on other researchers",
      "Non-peer-reviewed health claims",
      "Conspiracy theories"
    ],
    "tone_requirements": [
      "Evidence-based language",
      "Respectful of differing viewpoints",
      "Avoids absolutes and hyperbole"
    ],
    "additional_instructions": "Flag tweets that make unsubstantiated claims or attack individuals. Keep scholarly discussion, research sharing, and respectful debate."
  }
}
Use case: Researchers applying for tenure or joining institutions with strict public communication guidelines.

​
Tech startup founder

For founders wanting professional but personality-driven content:
config.json
{
  "criteria": {
    "forbidden_words": [
      "scam",
      "fraud",
      "fake"
    ],
    "topics_to_exclude": [
      "Personal attacks on competitors",
      "Disparaging comments about other companies",
      "Negative comments about investors or employees",
      "Controversial political statements"
    ],
    "tone_requirements": [
      "No aggressive or confrontational language toward others",
      "Avoid pessimistic or defeatist messaging"
    ],
    "additional_instructions": "Flag tweets that could damage business relationships or investor confidence. Keep optimistic, constructive content even if casual."
  }
}
Use case: Founders preparing for fundraising rounds or acquisition discussions.

​
Minimal filtering

For users who want to flag only legally risky content:
config.json
{
  "criteria": {
    "forbidden_words": [],
    "topics_to_exclude": [
      "Threats of violence",
      "Illegal activities",
      "Harassment or doxxing",
      "Copyright infringement"
    ],
    "tone_requirements": [],
    "additional_instructions": "Only flag content that could result in legal action or platform suspension"
  }
}
Use case: Users who want to keep their authentic voice while removing only serious risks.

​
Testing and refining criteria

1

Start with a test sample

Create a small test archive with 10-20 representative tweets to test your criteria.
# Backup your full archive
mv data/tweets/tweets.json data/tweets/tweets-full.json

# Create a small test file (manually select 20 tweets)
# ... copy sample tweets ...
2

Run analysis on test data

python src/main.py extract-tweets
python src/main.py analyze-tweets
3

Review false positives/negatives

Check data/tweets/processed/results.csv:
  • False positives: Good tweets incorrectly flagged
  • False negatives: Bad tweets that should be flagged but weren’t
4

Adjust criteria

Based on your review:
  • Too many false positives → Make criteria more specific
  • Too many false negatives → Make criteria broader
5

Reset and retry

rm data/checkpoint.txt
rm data/tweets/processed/results.csv
python src/main.py analyze-tweets
6

Deploy on full archive

Once satisfied, restore your full archive:
mv data/tweets/tweets-full.json data/tweets/tweets.json
rm data/checkpoint.txt
python src/main.py extract-tweets
python src/main.py analyze-tweets

​
Common criteria patterns

​
Pattern: Word variations

To catch word variations, add multiple forms: 
"forbidden_words": [
  "crypto",
  "cryptocurrency",
  "cryptocurrencies",
  "bitcoin",
  "BTC",
  "ethereum",
  "ETH"
]

​
Pattern: Time-based filtering

Target specific time periods using additional_instructions: 
"additional_instructions": "Be especially strict with tweets from 2017-2018 during the ICO boom. Flag any cryptocurrency enthusiasm from that period."
The AI doesn’t have access to tweet timestamps in the current implementation, but mentioning time periods can help it recognize context clues in the content.

​
Pattern: Industry-specific jargon

Remove industry-specific terms you no longer want associated with you: 
"forbidden_words": [
  "synergy",
  "disruptive",
  "paradigm shift",
  "circle back",
  "low-hanging fruit"
],
"additional_instructions": "Flag corporate buzzwords and empty jargon"

​
Troubleshooting criteria

​
Everything gets flagged

Your criteria is too strict: 
{
  "criteria": {
    "topics_to_exclude": [
      "Any opinion",
      "Personal thoughts",
      "Casual language"
    ]
  }
}

​
Nothing gets flagged

Your criteria is too lenient: 
{
  "criteria": {
    "forbidden_words": ["fuck"],
    "topics_to_exclude": [
      "Explicit content"
    ]
  }
}

​
Next steps

Basic workflow

Learn the end-to-end process

Large archives

Handle archives with 10,000+ tweets

Build docs developers (and LLMs) love

Get started for freeTalk to us