a builder ships promo codes, a red team attacks the change, and a judge blocks or approves it, citing the evidence.
recorded run · a small sample app in this repo · an illustration
demo 2 of 6
audio retrieval eval
search 30 clips by how they sound, not their file names. add up to 10 of your own; they stay in your browser.
try
type a word and press search.
sounds
ElevenLabs text-to-sound, 30 clips with deliberate near-duplicates.
model
LAION CLAP (larger_clap_general, 8-bit ONNX): words and audio in one space.
why this
newer audio-text models score higher on retrieval benchmarks. i picked CLAP because it's light and runs in the browser, so the search happens on your device with no server.
runtime
Transformers.js in your browser. ~205 MB once, then cached.
audio
Web Audio API for decoding and playback, canvas for waveforms.
score
cosine similarity, scaled so .12 = 0 and .45 = 100.
prep
FFmpeg: 5 s, 64 kbps mono mp3.
demo 3 of 6
self-healing test
a UI change breaks a browser test. an agent reads the failure, patches the selector, and the test reruns green.
recorded run · a small sample app in this repo · an illustration
demo 4 of 6
bug to fix
a tester files an issue, an agent reproduces it with a failing test and opens a pull request, and the CI check passes.
recorded run · a small sample app in this repo · an illustration
demo 5 of 6
adobro
tap a mouse button, click any panel on screen, and a short AI answer opens beside it, looking things up on the web when it needs to. a Windows tool i built and use.
recorded answer · the screen is a stand-in, the answer is real
arm it, then click a panel
click a panel · esc cancels
✦ask about this
what's thiscopy texttranslate
C# · .NET 8 · WPF · UI Automation · Windows.UI.Composition · Windows OCR · xAI Responses API with web search
demo 6 of 6
skill eval
is the deslop skill better than no prompt or a same-length generic one? 6 tasks × 3 runs.
recorded runs on a sample app
judge score out of 5 per task, mean of 3 runs per arm, and the skill's lift over the other two arms
task · tap to replay
baseline
placebo
skill
vs baseline
vs placebo
how i work
methodology
verification is the most important part, then encoding what i learn from an agent's failure modes, because they're very teachable.
$ idea --to ship
01researchyou and agents explore the idea with the right playbooks and skills.
02gatea gating agent proves it's feasible, or it gets scrapped.
03briefthe chief of staff (CoS) gets briefed and oversees it across projects.
04specPMs draft and lock the spec, and verify the needed skills are built before any code.
05buildengineer and QA agents work with evidence, not claims.
06attackred teams try to break it, and failures loop back.
07ship gatepre-ship-check must pass, and failures loop back.
08learnfailures and lessons feed back into the playbooks and skills.
it's exactly how i set up my multi-bot harness.
verification and shipping
the ship gate. no proof, no ship.
give agents hands and eyes for each app.
this site's own proof run.
drives the browser to test web apps.
finds what to clean up before it gets expensive.
build and review
settles the types and structure first.
runs parallel attempts and keeps the best.
adversarial review across models.
shows what a change could break.
measures before claiming something got faster.
design and UI
a design rulebook on top of a shared design system.audit · critique · polish · layout · adapt · colorize · bolder · distill · delight · animate · clarify · optimize
craft and feel.
makes it usable for everyone.
quick, strong UI drafts.
point at a skill to see what it does.
no proof, no ship.
every failure gets fed back into the playbook.
attack it before anyone else does.
when an agent fails, i usually catch it at the end, in the red-team checks or the pre-ship and verification checks, and the cause is usually the spec or a lack of coverage. those failures always get fed back into the skills.
what broke
while building this site, the first audio demo scored "cat" on a real cat clip at 32 of 100.
why
a gap in coverage.
the fix
send the bare query and calibrate the score. cat clips rank first
my guardrails are the verification skills at the end.