LitHarness writes LitRPG serial chapters from a short brief. A pitch call writes a bible you glance
at before any chapter money is spent; each ~1,500-word chapter then costs two calls (plan, then
draft), plus one located rewrite when the tells counter says so. Code owns the numbers: the hero opens
his status as [Label: value] lines, and code reads them back into the sheet. Everything is plain files;
nothing to install beyond Python 3.11+ (standard library only) and an agent CLI signed in with its
subscription: Codex, Claude Code, or both.
The previous implementation is kept, unchanged, at the tag legacy/incumbent-2026-09-27 and the branch
legacy/main. The design is frozen in experiments/2026-09-27-rebuild-plan/PLAN.md;
LEARNINGS.md says why it is this small.
| Verb | Does | Calls |
|---|---|---|
new SLUG --brief FILE [--words 1500] (500-1700) |
creates the serial, draws the pitch, stops for your glance | 1 (up to 3) |
next SLUG [-n K] (K up to 10) |
plan, draft, checks, sheet, tells rewrite, report per chapter | 2-3 each |
agent [codex|claude] |
shows, or switches, the one agent that writes for the whole project | 0 |
status [SLUG] |
chapters, words, failing checks with quotes, tokens, hand edits, locks, the agent | 0 |
redraw SLUG --from N |
moves chapter N and later to attempts/<utc>/ (0: the pitch too) |
0 |
check FILE [--stage pitch|plan|chapter] [--serial SLUG] |
every deterministic check on any text | 0 |
canary [AGENT] |
live isolation canary; pins that agent's CLI version (default: the project's agent) | 1 small |
Resuming means re-running the same command: finished calls are adopted from their receipts and never bought twice. Exit codes: 0 done, 1 needs you (a failed stage with its quotes, a held lock, the canary), 2 fault.
python -m litharness agent claude # or codex; holds until the next switch
python -m litharness canary # once per agent, and after every upgrade of its CLI
python -m litharness new slot --brief briefs/slot.txt
python -m litharness next slot -n 3 # after you have read ch00/bible.md
python -m litharness status slotOne agent writes for the whole project, Codex or Claude Code, and agent switches it between any two
commands; a full model id and an effort can ride along (agent claude:claude-opus-5-5:high). The choice lives in
$LITHARNESS_HOME/agent, every receipt names the agent that made its call, and a switched agent is a
changed input, so it opens draws 4-6.
A stage that fails a hard check (money, leak, person, rise, fields, pitch-shape,
plan-shape, length) is redrawn unchanged, at most 3 draws per input set. A money word that our own
request carried stops at once with its file and line. Editing the brief, bible, state or plan opens
draws 4-6; chapter text is never hand-edited, redraw instead.
Serials live in $LITHARNESS_HOME/serials/<slug>/ (default ~/LitHarness-data), outside every git
tree: serial.json, brief.md, manifest.json, ch00/ (the pitch) and chNN/ with state.md,
plan.md, final.md (the raw draft), chapter.md, sheet.txt and report.md. Every call keeps its
request, system prompt, events, stderr, answer and receipt under chNN/calls/. Nothing is deleted.
Back up the generated books to OneDrive with:
Compress-Archive -Path "$HOME\LitHarness-data\serials" -DestinationPath "$env:OneDrive\LitHarness-backups\serials-$(Get-Date -Format yyyy-MM-dd).zip"Chapters posted to Royal Road carry its AI-Generated tag.
python bench.py baseline SLUG places every chapter of a serial among 182 LitRPG titles on RoyalRoad at the
same chapter number, by code alone: no model call, seconds after the first run. The report lands in
serials/SLUG/baseline/ and describes; it never decides. The corpus and its protocol are in
experiments/2026-09-28-genre-baseline.
python draws.py run --agent codex draws fresh pitches for 11 past briefs and reports how often a draw passes
its hard checks, which check fails and what a draw costs: one call a row, graded by code, describing and never
deciding. python draws.py approve is yours to run first. Protocol: experiments/2026-10-02-pitch-draws.
python -m unittest -q: serial, stdlib only, a few seconds, and no test spawns a model CLI.