Skip to content

PROCESS.md

Latest commit

 

History

History
314 lines (235 loc) · 11.2 KB

PROCESS.md

File metadata and controls

314 lines (235 loc) · 11.2 KB

How to build your own .md file

An agent-executable plan for extracting a writing voice from a Twitter/X archive. Paste this into Claude Code (or any coding agent) along with your archive path and author name.

Prerequisites

  • A Twitter/X data archive (.zip from Settings > Your Account > Download an archive)
  • Unzip it. The file you need is data/tweets.js
  • Node.js / npx tsx available

Tell the agent:

Build a writing profile for [AUTHOR NAME] using the process in PROCESS.md.
My Twitter archive is at [PATH TO data/tweets.js].
Output the final profile to [AUTHOR]_profile.md.
Ask me to validate findings before finalizing.

Phase 1: Pre-process the archive

Input: data/tweets.js Output: tweet-index.json

Write and run a script that:

  1. Reads data/tweets.js and strips the window.YTD.tweets.part0 = prefix to get valid JSON
  2. Filters out:
    • Retweets (full_text starts with "RT @")
    • Pure replies (full_text starts with "@")
    • URL-only tweets (no text after stripping https?://\S+)
    • Empty text
  3. Extracts per tweet: id, text (with URLs stripped), date (created_at), likes (favorite_count as int), rts (retweet_count as int)
  4. Sorts by likes descending
  5. Writes to tweet-index.json
  6. Logs total tweets scanned and total indexed

Checkpoint: Print "[N] tweets scanned, [M] indexed" before proceeding.

Phase 2: Quantitative analysis

Input: tweet-index.json (use top 500 by likes) Output: Sections written to the profile markdown file

Run each analysis as a separate task. Read the tweet index, compute the numbers, and write each section to the output file. Do not combine analyses into one pass.

2a. Shape

Compute:

  • Median word count per tweet
  • % under 10 words, under 20 words, under 50 words
  • % that are a single sentence (no period-separated clauses, no line breaks creating multiple statements)
  • Average likes by word count bracket (1-5, 6-10, 11-15, 16-20, 21-30, 31+)

Write as ## Numbers section.

2b. Perspective

Compute:

  • % of tweets containing "you" or "your"
  • % containing "I" or "my"
  • % containing "we" or "our"
  • Same ratios for top 100 only

Write as ## Perspective section.

2c. Punctuation fingerprint

Compute average per tweet:

  • Periods, commas, line breaks (\n), question marks, exclamation points, colons, semicolons, em dashes
  • % ending with period vs no punctuation vs other
  • Average likes for period-ending vs bare-word-ending tweets

Write as ## Punctuation section.

2d. Structure templates

Classify each of the top 500 tweets:

  • Single sentence
  • Two-part parallel ("X does A. Y does B.")
  • Conditional ("If X, then Y.")
  • Numbered list ("1. ... 2. ...")
  • Explicit contrast ("X vs. Y")
  • Other (describe)

Give percentages. Write as ## Structure templates section.

2e. Opening patterns

Classify the opening of each tweet:

  • Observation/declaration (starts with noun or statement of fact)
  • Numbered list (starts with "1." or "1)")
  • Conditional (starts with "If" or "When")
  • Imperative verb (starts with a command)
  • Quote (starts with attributed quote)
  • Question

Give percentages. Write as ## Opening patterns section.

2f. Closing patterns

Analyze the final word of each tweet:

  • 20 most frequent final words with counts
  • Dominant part of speech of final word
  • % ending with period vs no punctuation
  • Average engagement for each ending type
  • Structural closing patterns (punchline inversion, imperative ending, question ending, simple declarative)

Write as ## Closing patterns section.

2g. Verb mood

Classify each tweet by dominant verb mood:

  • Declarative (statement of fact)
  • Imperative (command)
  • Conditional (if/when)
  • Interrogative (question)

Give percentages and average engagement per mood. Write as ## Verb mood section.

Phase 3: Rhetorical pattern extraction

Input: Top 300 tweets from tweet-index.json Output: Sections appended to the profile

3a. Rhetorical moves

Read the top 300 tweets. Identify every recurring rhetorical move. For each:

  • Name it
  • Describe the mechanics (what makes it work, not just what it is)
  • Give 3-5 example tweets with exact text

Look for but don't limit to: contrast pairs, reframes, quantification of abstract ideas, humor through understatement, uncomfortable truths, compressed frameworks, definitional reframes.

Write as ## Rhetorical patterns section.

3b. Word-level mechanics

Read the top 500 tweets. Run separate searches for each device:

  1. Alliterative contrasts: Find every tweet where two compared concepts start with the same letter. List each with exact text and letter pair.
  2. Matched meter: Find couplets where both lines have equal or near-equal syllable counts.
  3. Chiasmus: Find A-B / B-A word-order reversals.
  4. Circular loops: Find tweets where the ending echoes or returns to the beginning.
  5. Internal rhyme: Find sound echoes within or across lines.
  6. Negation flips: Find tweets using the same words with "don't" / "no" added or removed to invert meaning.
  7. Paradox: Find self-contradictory statements that resolve into insight.
  8. Ending word analysis: Do punchlines tend toward monosyllabic or polysyllabic words? Anglo-Saxon or Latinate?
  9. Case patterns: Does lowercase vs capitalized serve a deliberate purpose? Compare engagement.

Write as ## Word-level mechanics section with subsections.

3c. Contrast frames

Read the top 500 tweets. Find every tweet containing a contrast or comparison. Classify the syntactic frame:

  • Reframe (single sentence repositioning)
  • Parallel declaration ("X does A. Y does B.")
  • Paradox (self-contradiction resolving to insight)
  • Conditional reveal ("If X, [surprising Y]")
  • Juxtaposed pair (two words, no verb)
  • Progression (numbered escalation)
  • Explicit vs ("X vs Y")
  • Negation flip ("Not X, Y" or "No one X, everyone Y")
  • Expectation subversion ("What you think: ... What's true: ...")
  • Labeled contrast ("[Label A]: X. [Label B]: Y.")
  • Chiastic inversion (ABBA word swap)
  • Cyclical (A->B->C->A loop)
  • Any other frames discovered

Rank by frequency. Give percentages and 3+ examples of each. Write as ## Contrast frames section.

3d. Colon / punctuation as device

Find every tweet using a colon. Classify:

  • Label for visual/thread (structural, not rhetorical)
  • Setup:punchline pivot (rhetorical)

For rhetorical colons: what's the word count before vs after? Compare engagement to non-colon tweets. Write as ## Colon as pivot section (or name for whatever punctuation device is most prominent).

Phase 4: Negative space and constraints

Input: Top 500 tweets Output: Sections appended to the profile

4a. What the voice never says

Read 500 tweets and check for the absence of each category:

  • Personal emotions ("I feel...", "I'm excited...")
  • Apologies or self-correction
  • Current events or news commentary
  • Personal biographical details (daily activities, travel, meals)
  • Engagement asks (like, share, follow, RT)
  • Gratitude performances ("so grateful...", "humbled...")
  • Vulnerability theater ("I'll be honest...", "real talk...")
  • Motivational cliches ("believe in yourself", "never give up")
  • Own success metrics (revenue, follower counts)
  • Pop culture references (movies, TV, sports, music)
  • Political opinions
  • Religious references
  • Complaining or negativity about specific people or groups
  • Self-promotion with direct calls to purchase

For each: confirm truly absent, or count rare exceptions and quote them. Write as ## What the voice never says section.

4b. Banned words

Based on everything analyzed so far, generate a list of words and phrases that would break this voice. Categories:

  • Corporate jargon
  • Filler phrases and hedging language
  • Buzzwords
  • Transition words that add no meaning
  • Cliche phrases

Write as ## Never use these words section.

4c. Anti-patterns

List structural habits the voice avoids. Things like: long explanations, preamble, summary paragraphs, excessive adjectives, em dashes, self-congratulation, hedging qualifiers. Write as ## What the voice never does section.

Phase 5: Voice description and themes

Input: All analysis completed in phases 2-4 Output: Sections prepended/appended to the profile

5a. Voice description

Write a one-paragraph description of the voice based on all quantitative and qualitative findings. No hedging. Declarative sentences only. Write as ## Voice section at the top of the file.

5b. Signature vocabulary

Find the 15-20 most frequently used meaningful words (exclude stop words). Note what the word choices reveal about the worldview. Write as ## Signature vocabulary section.

5c. Tone

Describe the tone in 3-5 bullet points based on the data. Not what the voice says, but how it feels. Write as ## Tone section.

5d. Themes

Identify 5-8 recurring themes across the corpus. For each, give a name and 2-3 example phrasings. Write as ## Themes section.

Phase 6: Teaching examples

Input: tweet-index.json + all analysis Output: Final sections of the profile

6a. Rewrite pairs

Select 10 ideas from the corpus. For each, write:

  1. A generic version (how a default writer would say it - long, hedging, jargon)
  2. The actual tweet (compressed, direct)

Write as ## Rewrite pairs section.

6b. Reference examples

Select 20-30 of the highest-performing tweets that best represent the voice. Choose for variety across rhetorical moves, structural patterns, and themes. Write as ## Reference tweets section.

6c. Long-form rules

If the corpus contains any long-form writing (threads, essays), describe the format: paragraph length, opening style, closing style, use of visuals. Write as ## Article format section. Skip if no long-form exists.

Phase 7: Validate with the author

This step is mandatory. Do not skip it.

Present all findings to the author grouped by category:

  1. Quantitative findings (shape, perspective, punctuation, structure, openings, closings, verb mood)
  2. Rhetorical moves identified
  3. Word-level mechanics found
  4. Contrast frames cataloged
  5. Negative space (what the voice never says)
  6. Banned words and anti-patterns

For each group, ask: "Which of these are intentional?"

Remove anything the author identifies as accidental or coincidental. The file should only contain deliberate patterns.

Phase 8: Assemble and order

Combine all validated sections into a single markdown file. Final section order:

## Voice
## Numbers
## Perspective
## Punctuation
## Structure templates
## What performs
## Sentence structure
## Rhetorical patterns
## Word-level mechanics
## Closing patterns
## Contrast frames
## Verb mood
## Colon as pivot
## What the voice never says
## Opening patterns
## Signature vocabulary
## Tone
## Themes
## Never use these words
## What the voice never does
## Article format
## Rewrite pairs
## Reference tweets

Write the final file to [AUTHOR]_profile.md.

Notes

  • 1,000 posts is a minimum corpus. 10,000+ gives statistical confidence.
  • Engagement data is critical. Without it you're measuring output, not signal.
  • Run each analysis in phases 2-4 separately. Combining them produces shallow results.
  • The file is never finished. New patterns surface over time.
  • This process works for any medium. Adapt phase 1 for newsletters, blog posts, transcripts, or books.