WEBVTT

00:00:03.940 --> 00:00:09.420
Yield gives small jobs to an AI helper and
keeps you in charge. This time it is a

00:00:09.420 --> 00:00:13.375
real run, on Claude, through the Anthropic
API.

00:00:17.435 --> 00:00:22.395
You pick a job, paste the notes, and set
the most it may ever spend. One dollar.

00:00:25.980 --> 00:00:30.860
The task starts. The Stage follows the
real task. It thinks, it writes, and then

00:00:30.860 --> 00:00:32.350
it stops and asks.

00:00:35.730 --> 00:00:41.120
Here is the whole text it wants to save.
Nothing is saved yet. Show proof gives the

00:00:41.120 --> 00:00:43.710
fingerprint of those exact bytes.

00:00:46.530 --> 00:00:51.740
In this recording, the script clicks
Approve. In real use, a person does.

00:00:53.605 --> 00:00:58.790
Finished. The page names the model, Claude
Sonnet five point five, and what it cost,

00:00:58.790 --> 00:01:01.890
about one cent. The record checks out.

00:01:05.060 --> 00:01:09.720
Now a task with a cap of almost nothing.
The cap is checked before any call is

00:01:09.720 --> 00:01:13.260
made. It stops, says why, and nothing is
spent.

00:01:18.655 --> 00:01:24.440
One more case, staged for this recording.
A stand-in model, not Claude, sends an

00:01:24.440 --> 00:01:30.460
answer that stops at its length limit. The
Needs you card warns you to read the end

00:01:30.460 --> 00:01:31.965
before you save it.

00:01:35.740 --> 00:01:39.840
Results adds up the spend from the task
records. Only the one real run cost

00:01:39.840 --> 00:01:40.620
anything.

00:01:43.815 --> 00:01:49.620
Eighteen runs, before and after the fixes.
Before, Haiku missed items in the reply

00:01:49.620 --> 00:01:54.760
task, once finding none, and some answers
were cut off at the length limit.

00:01:57.300 --> 00:02:03.060
After the fix, no Haiku answer was cut
off. Sonnet found every planted item both

00:02:03.060 --> 00:02:08.620
times. Times and costs come from the app.
We make no claim about how long the work

00:02:08.620 --> 00:02:09.870
takes by hand.

00:02:29.335 --> 00:02:38.100
Recorded on 2026-11-08, US Eastern time,
against Yield main at commit dfaae6c. A

00:02:38.100 --> 00:02:44.460
real run on Claude Sonnet 5.5, about one
cent. The script clicked Approve; in real

00:02:44.460 --> 00:02:49.940
use a person does. The cap test was
started through the API. The length-limit

00:02:49.940 --> 00:02:52.910
scene used a stand-in model, not Claude.

00:02:53.775 --> 00:02:58.760
Runs on Claude through the Anthropic API.
ShipItWorks is not affiliated with

00:02:58.760 --> 00:02:59.780
Anthropic.

00:03:00.700 --> 00:03:07.460
Yield 0.5.0, Beta. Our audit scored it 76
out of 100. Changes made since that audit

00:03:07.460 --> 00:03:09.950
were tested, but not re-audited.

00:03:10.835 --> 00:03:16.820
Not done yet. No manual comparison was
made. Approval time was not measured. A

00:03:16.820 --> 00:03:22.060
refused step is labelled price not known.
The length-limit warning was shown on a

00:03:22.060 --> 00:03:28.020
staged reply, not on a real cut-off from
Claude. The installer is older than the

00:03:28.020 --> 00:03:28.800
code.

00:03:29.575 --> 00:03:35.440
The voice you just heard is AI-generated,
using ElevenLabs. This walkthrough was

00:03:35.440 --> 00:03:40.575
recorded on the eighth of November, twenty
twenty-six, US Eastern time.