what does lulu draft do?

Turns one domain into a measured question set, asking the customer nothing else. It reads the customer's own site, asks a model for candidate questions, and keeps only the ones that survive.

how do you run it?

In the parser’s own words, turn one domain into a measured question set, asking nothing else. Every flag this build accepts, on one line:

lulu draft --site SITE --out PATH --floor FLOOR --budget USD [--rivals RIVALS] [--ledger LEDGER] [--provider PROVIDER] [--model MODEL] [--max-searches MAX_SEARCHES] [--pages PAGES] [--candidates CANDIDATES] [--harvest-only] [--dry-run]

lulu draft --site https://example.com --floor 0.5 --budget 5.00 --rivals RIVALS --out PATH  ± the whole path
lulu draft --site https://example.com --floor 0.5 --budget 5.00 --out PATH --harvest-only  ± read the site, spend nothing
lulu draft --site https://example.com --floor 0.5 --budget 5.00 --out PATH --dry-run  ± stop before the draws

where do the questions come from?

The person running it types one url. The pages that get read are ranked by what the site itself links to and lists, and every question arrives carrying the page it came from and a quote from that page, checked as a literal substring of that page before anything is asked for real.

what does --floor decide?

What a kept question has to beat, and the number of draws follows from it rather than from a constant. At a floor of 0.5 a clean sweep needs four draws before its lower bound clears the floor, so four is what it buys; at 0.75 it buys twelve, and the bill is printed before the money moves.

It costs money in two places and both are priced before either happens. One call proposes candidates, and one round asks the survivors for real. The second is the expensive one and it is also the pilot: the draws that decide which questions survive are the same draws the variance split is read off.

why does it count rivals and never the customer?

The gate measures how often an answer names a rival. Screening a question set on the customer's own mentions keeps the questions they scored well on and drops the rest, which raises the published number by deleting the evidence against it, so with no rival declared the paid round is refused rather than run on the one name available.

The rival list is still typed in by hand, through --rivals. The round prints the names the list missed, and the subject file records which round and which arm each name came from, so the list can be corrected between rounds.

where can it stop?

--harvest-only stops after reading the site and spends nothing. --dry-run adds the proposal call and the free gates and stops before the draws. Either way, every candidate that died is written to <out>.draft.json with the gate that stopped it, because a set of nine questions that started as forty is a different object from one that started as nine.

which flags does it take?

--site SITE ± the customer's site, and the only thing they type, required

--out PATH ± where the subject file is written, required

--floor FLOOR ± rate a question has to beat on rivals to be kept; also sets how many draws it takes, required

--budget USD ± hard ceiling in dollars; a draft with no ceiling is refused, required

--rivals RIVALS ± comma separated names the questions are screened on; without them the paid round is refused

--ledger LEDGER ± directory the screening round is written to, default ./ledger

--provider PROVIDER ± which engine answers, default anthropic

--model MODEL ± model to ask; the provider's check model by default

--max-searches MAX_SEARCHES ± searches one call may run, default 3

--pages PAGES ± how many of the site's pages are read, default 12

--candidates CANDIDATES ± questions asked for in one call, default 40

--harvest-only ± read the site and stop; no model, no spend

--dry-run ± corpus, one proposal call and the free gates; stops before the draws that measure

where does it lead?

lulu screened ± lulu rivals ± lulu collect

the engines it can ask: perplexity sonar, claude.

the parts of the method it reaches: intervals, variance, screen, names.

every command: setup, init, doctor, collect, draft, usage, ablate, lift, screened, publish, rivals, report, verify, plan, size. back to the docs