Prompt Distance (PD)

A whitepaper proposing Prompt Distance: the minimum number of prompts required to reproduce a software artifact from an empty directory using a frontier LLM.

In short: a way to say "this product is one prompt away from not existing" with a citation.

"How defensible is the product?" "Two prompts." (overheard at a board meeting, 2026)

Read the full paper: WHITEPAPER.md, or as a two-column PDF: paper/prompt-distance.pdf (prepared for SIGBOVIK 2027; build it yourself with tools/build-paper.sh, requires pandoc + texlive)

TL;DR

$$PD(x) = \min \{ |P| : \mathrm{eval}(M, P) \cong x \}$$

Class	Meaning	Example
PD-0	Already exists as a library or Unix utility. No project was needed.	rsync with a landing page; `cron` as a service, $9/mo
PD-1	One-shottable. The project is its own README.	AI summarizer for {meetings, emails, PDFs}; ChatGPT with your logo and a waitlist
PD-2..9	A short conversation. Usually one real design decision.	CRUD dashboard with dark mode; the fourth Trello clone this quarter; "Uber for X" where X didn't ask
PD-10..99	Actual software.	your product, allegedly; the billing system everyone is afraid of
PD-100+	Compilers, kernels, anything touching timezones.	Postgres; anything that parses dates and apologizes for it
PD-∞	The hard part is not code. See: the Dave Boundary.	the reverse proxy nobody dares restart; whatever Dave maintains

Scores decay as models improve (prompt rot), so they must be dated: PD(x) = 3 (claude-fable-5, 2026-06).

Usage

When reviewing a product, estimate its PD, date it, and disclose the transcript. Undated or transcript-free claims of "I one-shotted it" are inadmissible and shall be treated as Weekend Conjectures.

Low PD means the software is trivial. It does not mean the business is. Those are different papers.

Citation

The whole point. Now you can say it with a citation:

@misc{promptdistance2026,
  title        = {Prompt Distance: A Unified Metric for Software Triviality in the Post-Generative Era},
  author       = {thereisnotime},
  year         = {2026},
  doi          = {10.5281/zenodo.20677177},
  url          = {https://doi.org/10.5281/zenodo.20677177},
  note         = {Version 0.2. Undated citations are inadmissible (see section 5).}
}

Archived on Zenodo: 10.5281/zenodo.20677177 (concept DOI, always latest).

Getting Started

There is nothing to install. For a substantial fraction of the software this metric applies to, that was also true.

There is, however, a skill. skills/pd-review teaches your coding agent to perform PD reviews: dated score, rubric breakdown, sardonic verdict. The same file works across agents:

# Claude Code
mkdir -p ~/.claude/skills/pd-review
curl -fsSL https://raw.githubusercontent.com/thereisnotime/prompt-distance/main/skills/pd-review/SKILL.md \
  -o ~/.claude/skills/pd-review/SKILL.md

# Codex CLI (as a custom prompt)
mkdir -p ~/.codex/prompts
curl -fsSL https://raw.githubusercontent.com/thereisnotime/prompt-distance/main/skills/pd-review/SKILL.md \
  -o ~/.codex/prompts/pd-review.md

# OpenCode (as a command)
mkdir -p ~/.config/opencode/command
curl -fsSL https://raw.githubusercontent.com/thereisnotime/prompt-distance/main/skills/pd-review/SKILL.md \
  -o ~/.config/opencode/command/pd-review.md

Then ask your agent to "PD review" anything. The agent will comply. It has no Dave to consult.

Get your badge

Measured your project? Wear the score. Replace the score, model, and date with your own (undated badges are inadmissible):

![PD](https://img.shields.io/badge/PD-1_(claude--fable--5,_2026--06)-red)

Color scale: red for PD-1, orange for PD-2..9, yellow for PD-10..99, green for PD-100+, lightgrey for PD-∞ (measurement abandoned; Dave declined to comment). This repo wears PD-2: one prompt for the whitepaper, one for the LaTeX. We display this proudly, which by Corollary 1 makes it wrong.

Peer Review

This paper is under continuous open review. To submit a formal review, open a peer review issue using the structured form. Every submission receives an individual decision from the Program Committee, typically within one minute, which we consider a feature of the venue.

The acceptance rate to date is 0%. The committee is aware and considers this evidence of rigor.

Contributing

Pull requests are welcome and will be processed by the Program Committee (see Peer Review). To date the committee has maintained a perfect record, which we are statutorily prohibited from describing as "acceptance."

Acknowledgments

This paper was written entirely by hand, by a human, in a text editor, the way Kolmogorov intended. No language model was involved at any stage of its production. The author states this plainly and asks that you extend the claim the same scrutiny you would extend a PD score submitted without a transcript.

⭐ Star this repo. It will not increase its PD.

Prompt Distance

README