sixwords

Package a concept into exactly six words with an LLM coach. The six words are the human-facing text; everything that went into them — the full argument they compress, the interview, the drafts, the reason every word earned its slot — is the subtext, stored in a sentree .subtext.json document.

Install

Requires Python 3.11+ and an Anthropic API key.

uv tool install git+https://github.com/TheoryVentures/sixwords
# or: pipx install git+https://github.com/TheoryVentures/sixwords
export ANTHROPIC_API_KEY=sk-ant-...

Or run it without installing:

uvx --from git+https://github.com/TheoryVentures/sixwords sixwords write

Usage

sixwords write                              # interview → drafts → accept → saved story
sixwords read <story>.subtext.json          # six words, then the reveal
sixwords list                               # stories written so far
sixwords export <story>.subtext.json        # surface markdown only
sixwords login                              # sign in to sixwordidea.com (email code)
sixwords publish <story>.subtext.json       # publish a finalized story

write runs a short interview (the coach asks one question at a time to pin the idea down), then proposes three candidate packagings with per-word rationale. Pick one, type your own six words, or ask for changes. When you lock the words in, the coach writes the title and the full prose version of the argument, and everything is saved.

read shows only the six words first, then reveals the subtext layer by layer: the idea in full, the word-by-word rationale, the draft evolution, the interview, and the human/AI provenance. Pass --all to skip the pauses.

Stories are saved to ~/.sixwords/stories by default; override with --stories-dir or the SIXWORDS_STORIES_DIR environment variable.

Publishing

Finalized stories can be published to sixwordidea.com: an index of six-word idea cards where each card links to the story's raw subtext JSON at sixwordidea.com/ideas/<slug>.json — a plain URL agents can fetch with no headers or keys.

Publishing requires signing in once with sixwords login (a one-time code is emailed to you). Ideas land in a Supabase table, and a GitHub Action rebuilds the static site (sixwords build-site) and deploys it to GitHub Pages. Backend setup lives in supabase/README.md.

The artifact

Each story is one nested-JSON sentree document with a single sentence (s-0001):

• sentence.text — the six words
• sentence.meta.backstory — the full idea they compress (claim, mechanism, stakes)
• sentence.meta.word_choices — per-word rationale and rejected alternatives
• sentence.meta.drafts — every draft, stamped with author and rationale
• metadata.interview — the coach's questions and your answers
• metadata.candidates_considered — every candidate proposed, chosen or not
• provenance + metadata.edit_log — who wrote the final words (human vs agent), when, and why, maintained by sentree's mutation API

The default model is claude-sonnet-5; override with SIXWORDS_MODEL.

Development

git clone https://github.com/TheoryVentures/sixwords
cd sixwords
uv sync
uv run pytest        # tests use a fake Anthropic client; no key needed
uv run ruff check .

The sentree library is vendored at vendor/sentree (MIT); see vendor/README.md for how to update it.