Style Transfer in AI writing

Shreyas Prakash headshot

Shreyas Prakash

The first draft usually generated by AI is so devoid of your style and personality, and to get to the stage where it even comes remotely close to that requires a lot of fine-tuning, and custom prompts.

Last week, I was looking at various ways in which I could “style transfer” my writing style to AI generated drafts. Due to the time consumed in style-transfer, I ended up writing on my own. (Read more here by watching this video essay)

And yet, I still feel that there is a way in which this could be more efficient. I used Claude Projects to get this working prototype in place, and here’s how I did this:

I knew that the best evergreen content for training would be my published blog posts. Rough drafts or private notes would be too crude for the AI to make patterns out of it. So I used this as a good benchmark to emulate my style. I uploaded the markdown files into the Project knowledge section of the Claude Projects.

You can upload relevant documents, text, code, or other files to a project’s knowledge base, which Claude will use to better understand the context and background for your individual chats within that project. In addition, you can define project instructions for each project to further tailor Claude’s responses. For example, instructing Claude to use a more formal tone or answer questions from the perspective of a specific role or industry.

I attempted to make a sample blog with this instruction:

  
I want to write essays that follow the style and flair of my previous blog posts. I don't want it to look like any AI-generated gibberish, copypasta. I want it to sound, feel and smell as if it's written by me: Shreyas Prakash.

Write a blog on 'moats for open source large language models'

While thinking through the possible answer for this prompt, Claude was able to analyse my language, patterns etc, to then adjust various aspects of the essay to better match my voice:

I've written this essay incorporating several elements I noticed in your writing style:

1. Use of vivid analogies and metaphors (Robin Hood, master chef's signature dish)
2. Breaking down complex topics into clear, digestible sections
3. Incorporation of real-world examples and companies
4. A conversational yet insightful tone
5. Forward-looking analysis that goes beyond the obvious
6. Personal perspective and observations

In the first version it produced, I noticed a slightly better draft than the generic ones usually generated.

I still felt like the headers and descriptions where making it look AI-ish, so I gave an additional prompt for it to flow well:

This was a good start. But it’s just that — a start. The next part of using this material before publishing is to reflect on the sentences, and to add my own spin to it.

To summarise, these are the steps I followed —

  1. Upload your blog markdown exports to the Claude projects knowledge base.
  2. Give a description to the project on Claude as : ‘Write new blogs on my personal website based on “style transfer” from the way I have written previous blogs’
  3. Provide a prompt to guide this process further — I want to write essays that follow the style and flair of my previous blog posts. I don't want it to look like any AI-generated gibberish, copypasta. I want it to sound, feel and smell as if it's written by me: Shreyas Prakash. Write a blog on 'TOPIC'
  4. Provide more guardrails such as — A strong opening hook, natural progression of ideas without explicit headers, characteristic use of concrete examples and metaphors, a more conversational tone throughout and with a signature email sign-off.

Subscribe to get future posts via email (or grab the RSS feed). 2-3 ideas every month across design and tech

Read more

  1. Breadboarding, shaping, slicing, and steelthreading solutions with AI agentsproduct-management
  2. How I started building softwares with AI agents being non technicalagentic-engineering
  3. Legible and illegible tasks in organisationsproduct
  4. L2 Fat marker sketchesdesign
  5. Writing as moats for humanswriting
  6. Beauty of second degree probesdecision-making
  7. Read raw transcriptsknowledge
  8. Boundary objects as the new prototypesprototyping
  9. One way door decisionsproduct
  10. Finished softwares should existproduct
  11. Essay Quality Rankerobsidian
  12. Export LLM conversations as snippetsbrowser-extension
  13. Flipping questions on its headinterviewing
  14. Vibe writing maximswriting
  15. How I blog with Obsidian, Cloudflare, AstroJS, Githubwriting
  16. How I build greenfield apps with AI-assisted codingai-coding
  17. We have been scammed by the Gaussian distribution clubmathematics
  18. Classify incentive problems into stag hunts, and prisoners dilemmasgame-theory
  19. I was wrong about optimal stoppingmathematics
  20. Thinking like a ship
  21. Hyperpersonalised N=1 learningeducation
  22. New mediums for humans to complement superintelligenceai-coding
  23. Maxims for AI assisted codingai-coding
  24. Personal Website Starter Kitai-coding
  25. Virtual bookshelvesaesthetics
  26. It's computational and AI everythingai-coding
  27. Public gardens, secret routesdigital-garden
  28. Git way of learning to codeai-coding
  29. Kaomoji generatorsoftware
  30. Copy, Paste and Citecuriosities
  31. Style Transfer in AI writingai-coding
  32. Understanding codebases without using codeai-coding
  33. Vibe coding with Cursorai-coding
  34. Virtuoso Guide for Personal Memory Systemsmemory
  35. Writing in Future Pastwriting
  36. Publish Originally, Syndicate Elsewhereblogging
  37. Poetic License of Designdesign
  38. Idea in the shower, testing before breakfastsoftware
  39. Technology and regulation have a dance of ice and firetechnology
  40. How I ship "stuff"software
  41. Weekly TODO List on CLIcli
  42. Writing is thinkingwriting
  43. Song of Shapes, Words and Pathscreativity
  44. How do we absorb ideas better?knowledge
  45. Read writers who operatewriting
  46. Brew your ideas lazilyideas
  47. Vibescreativity
  48. Trees, Branches, Twigs and Leaves — Mental Models for Writingwriting
  49. Compound Interest of Private Notesknowledge
  50. Conceptual Compression for LLMsai-coding
  51. Meta-analysis for contradictory research findingsdigital-health
  52. Beauty of Zettelswriting
  53. Proof of workproduct
  54. Gauging previous work of new joinees to the teamleadership
  55. Task management for product managersproduct
  56. Stitching React and Rails togetherai-coding
  57. Exploring "smart connections" for note takingknowledge
  58. Deploying Home Cooked Apps with Railssoftware
  59. Self Marketing
  60. Repetitive Copypromptingwriting
  61. Questions to ask every decadejournalling
  62. Balancing work, time and focusproductivity
  63. Hyperlinks are like cashew nutswriting
  64. Brand treatments, Design Systems, Vibesdesign
  65. How to spot human writing on the internet?writing
  66. Can a thought be an algorithm?product
  67. Opportunity Harvestingcareers
  68. How does AI affect UI?design
  69. Everything is a prioritisation problemproduct-management
  70. Nowlifestyle
  71. How I do product roastsproduct
  72. The Modern Startup Stacksoftware
  73. In-person vision transmissionproduct
  74. How might we help children invent for social good?social-design
  75. The meeting before the meetingmeetings
  76. Design that's so bad it's actually gooddesign
  77. Breaking the fourth wall of an interviewinterviewing
  78. Obsessing over personal websitessoftware
  79. Convert v0.dev React to Rails ViewComponentsrails
  80. English is the hot new programming languagesoftware
  81. Better way to think about conflictsconflict-management
  82. The role of taste in building productsdesign
  83. World's most ancient public health problemsoftware
  84. Dear enterprises, we're tired of your subscriptionssoftware
  85. Products need not be user centereddesign
  86. Pluginisation of Modern Softwaredesign
  87. Let's make every work 'strategic'consulting
  88. Making Nielsen's heuristics more digestibledesign
  89. Startups are a fertile ground for risk takingentrepreneurship
  90. Insights are not just a salad of factsdesign
  91. Minimum Lovable Productproduct
  92. Methods are lifejackets not straight jacketsmethodology
  93. How to arrive at on-brand colours?design
  94. Minto principle for writing memoswriting
  95. Importance of Whytask-management
  96. Quality Ideas Trump Executionsoftware
  97. How to hire a personal doctor
  98. Why I prefer indie softwareslifestyle
  99. Use code only if no code failscode
  100. Personal Observation Techniquesdesign
  101. Design is a confusing worddesign
  102. A Primer to Service Design Blueprintsdesign
  103. Rapid Journey Prototypingdesign
  104. Directory Structure Visualizercli
  105. AI git commitscli
  106. Do's and Don'ts of User Researchdesign
  107. Design Manifestodesign
  108. Complex project management for productproducts
  109. How might we enable patients and caregivers to overcome preventable health conditions?digital-health
  110. Pedagogy of the Uncharted — What for, and Where to?education
  111. Future of Ageing with Mehdi Yacoubiinterviewing
  112. Future of Equity with Ludovick Petersinterviewing
  113. Future of Mental Health with Kavya Raointerviewing
  114. Future of Tacit knowledge with Celeste Volpiinterviewing
  115. Future of Rural Innovation with Thabiso Blak Mashabainterviewing
  116. Future of unschooling with Che Vanniinterviewing
  117. Future of work with Laetitia Vitaudinterviewing
  118. How might we prevent acquired infections in hospitals?digital-health
  119. The soul searching yearsentrepreneurship
  120. Design education amidst social tribulationsdesign
  121. How might we assist deafblind runners to navigate?social-design