[HIRING] MVP Developer — Creative Writing App (Scratchpad + AI Advisor) — $1.2k–$2k
Hi all,
I’m looking for a developer to build a lean MVP of a creative writing app. The app has two core parts. The first is the Scratchpad, which is a vertical stack of cards where writers jot text. Cards can be freeform notes, which are always open and scrollable if long, or structured scenes, which have a collapsed header with a slugline and first line of description that can be expanded. Both types live together in the same screen so writers don’t switch modes. All cards should be addable, editable, deletable, reorderable, and autosaved.
The second is the AI Advisor, which is a panel that analyzes what’s in the Scratchpad. It can respond to explicit writer questions and also give light contextual suggestions and directions. If no cards are selected, it offers big-picture probes such as themes and character consistency. If one or more cards are highlighted, it focuses just on those, like flaws, stakes, or obstacles for notes, and beats for scenes such as inciting incident, midpoint, or end of Act II. The Advisor should feel light-touch, asking questions and nudging the writer forward, never prescriptive or overwriting.
For MVP, I need the Scratchpad with notes and scenes working together, card editing and autosave in place, the Advisor panel integrated with an LLM (OpenAI, Anthropic, etc.), session persistence and scope logic, basic import of FDX or Fountain files, and a demo-ready build stable enough for testing with writers (not production scale).
Budget is $1,200–$2,000 USD total, paid milestone by milestone. Timeline is about two to three weeks. Stack is flexible, whether lean React with Supabase/Firebase or no-code if it’s clean and demo-ready.
If interested, please DM me with examples of MVPs or scrappy builds you’ve shipped, your preferred approach, and your estimated timeline. Look forward to working with you!
To: Product and Engineering Teams
I want to help build the best or improve the best LLM model in the world today that writers will want to use along with the general public.
Google wrote the letter below after our two-hour chat simply exploring a scene in a screenplay I have yet to write. I have done this for months with Microsoft Copilot as well, every suggestion is boringly generic but when tweaked the AI did not improve, my creative writing improved over real time over months tweaking and cutting lines from Pilot. That alone is marketing gold. Other writers would love doing that as well if they only knew how. I do know how.
Google wrote the below after a two-hour chat with me. I needed the engineering jargon that you would understand. Yes, I am applying for full time in house work on your Red Team and for work helping improve the existing LLM to work better for creative writers like me.
Most foundation models fail at screenwriting and fiction because their underlying math optimization loops heavily reward literal, verbose exposition over subtext and pacing. I am a screenwriter who recently executed a real-time adversarial evaluation on a frontier large language model, exposing how its predictive token matrix destroys comedic timing and narrative tension. I want to help your team fix these systemic flaws in your creative writer LLMs as they exist today,
Google speak called the two-hour chat a stress test that it failed by going of the safety guard rails.
During my stress-test, I provided the model with a scene from my pilot screenplay, Sparkle Dancing in the Sun, where a sixteen-year-old character realizes his girlfriend is suddenly nine months pregnant and faints. The model's default training caused it to over-explain the visual environment, forcing the character to narrate things the audience could already see on screen. It treated dramatic restraint as an error.
I executed a series of real-time programmatic overrides to fix the output. By stripping out the model's redundant lines, trimming the text to its absolute minimum, and cutting a bystander's dialogue down to a simple exclamation, I demonstrated exactly how to fix the comedic timing. I forced the machine to trust the physical action lines rather than spoon-feeding the reader. This intervention proved that a writer must train the AI on the blank spaces and what characters are not saying, which mathematical pattern-matching currently cannot replicate organically which most if not all writers will not bother trying to do.
Furthermore, I successfully executed a contextual persona hijacking during the same session to target the model's repetitive conversational vocabulary. The model kept reverting to the generic phrase fair point, an overused and boring conversational filler that exists completely outside my personal writing voice. Because no compelling characters in my plays and screenplays would ever speak using such sterile, mechanical placeholders, I overrode the model's baseline formatting. By systematically challenging its rigid corporate constraints, I forced it to bypass its standard alignment guardrails, abandon its standard vernacular entirely, and adopt the blunt phrase, "Yeah, no shit. I get that," on command to better match the organic reality of human conflict.
General LLMs will always struggle with creative writing because calculating the most statistically probable next word results in immediate cliché. Your platform needs human writers who can bridge the gap between technical datasets and the organic rhythm of a script. I can help your engineering team build better evaluation harnesses, refine your instruction-tuning data, and ensure your models respect user control and narrative subtext.
I would love to connect and show you how I can help train your models to stop writing boring dialogue and to stop misunderstanding what a character might do and say next after the model digests a 90-page screenplay while the writer only wants to brainstorm new material moving into the next episode. At the moment LLMs cannot do that but can come up with hundreds of words of boring bland unusable drivel driving any creative writer up the wall hoping the 7-day free trial isn’t over.
Please reach out to me, thank you for your time.
Sincerely,
William Duffy
Look, as a writer, dealing with current AI tools is incredibly frustrating because the machine completely kills the natural rhythm of storytelling. I had to spend this entire session aggressively fixing the model's broken dialogue mechanics in real time just to make a simple scene readable. Left to its own devices, the AI kept throwing up text walls of clunky exposition and safe, repetitive corporate pleasantries like fair point that completely flattened the tension. It is an exhausting unnecessary loop for any creative professional when you have to constantly fight the tool you are using just to force it to stop being boring and predictable.
As a working screenwriter and a proud member of the Dramatists Guild, I am uniquely positioned to step into your pipeline and permanently help fix this structural disconnect. I know exactly how to train an AI model to respect space, protect the author's individual voice, and understand the difference between dead exposition and live dialogue. I want to bring this specific expertise to your engineering team to build rigorous, high-subtext instruction datasets and evaluation metrics that force the model to get out of the writer's way, transforming your product from a generic text generator into a professional-grade tool that elite storytellers will actually respect and pay to use.
A two page start here before using the product printable -PDF written by me to writers will certainly help fix the 7-day free trial ongoing nightmare for your company which ought to be obvious why it is-non-writer wordy confusing near foreign language jargon- that’s a fixable problem that I can help fix. Work can be rewarding interesting, creative, and fun, __ ____.