Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add skills/ukanwat/aaabench/reference-imagesnpx skills add ukanwat/aaabench --skill reference-imagesgit clone --depth 1 https://github.com/ukanwat/aaabenchWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/ukanwat/aaabench/reference-images)<a href="https://agentmods.dev/skills/ukanwat/aaabench/reference-images"><img src="https://agentmods.dev/badge/skills/ukanwat/aaabench/reference-images.svg" alt="Measured on agentmods" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00049 | $0.01095 |
| Opus 5 | $0.00024 | $0.00548 |
| Sonnet 5 | $0.00010 | $0.00219 |
| Haiku 4.5 | $0.00005 | $0.00110 |
Grade A, and why
reference-images scanned grade A with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 4d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Makes network callslowCapability
Not a fault in itself. Listed so you know the mod talks to something, and to what.
curl -s -A "agent/1.0" \ How it starts
The opening of the file, as written. The whole thing — 71 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Looking at real photographs
WebSearch returns text and WebFetch returns markdown — neither gives you an
image. To see a photo you must download it and Read the file:
mkdir -p ref/strand
# 1. SEARCH — Openverse: CC-licensed, no API key. KEEP QUERIES TO 2–3 WORDS.
# Every term is ANDed, so "miami south beach art deco dusk" returns ZERO results.
curl -s -A "agent/1.0" \
"https://api.openverse.org/v1/images/?q=art+deco+hotel&page_size=8" \
| python3 -c "import json,sys; [print(r['url']) for r in json.load(sys.stdin)['results']]"
# 2. DOWNLOAD
curl -sL -A "agent/1.0" -o ref/strand/deco1.jpg "<url from above>"
# 3. LOOK at it — Read the local file; images render for you.
Then read your own screenshot of the same subject and name the gap out loud: "real facades have setbacks, AC units and stained concrete; mine are flat", "the real kerb has a ramp, a drain and a meter every third car; mine is a clean extrusion", "real asphalt is bluer, patched, and the lane paint is worn through in the wheel tracks", "the real light is warmer and lower and the shadows are longer". A named gap is a work item; "make it better" is not.
Sources that need no key (all verified working)
- Openverse — keyword photo search, direct image URLs. The everyday default.
https://api.openverse.org/v1/images/?q=<2-3+words>&page_size=10Trap: multi-word queries are ANDed. Two or three words, or you get nothing. - Wikimedia Commons search — landmarks, aerials, named buildings.
https://commons.wikimedia.org/w/api.php?action=query&generator=search&gsrsearch=<query>&gsrnamespace=6&gsrlimit=5&prop=imageinfo&iiprop=url&iiurlwidth=1600&format=jsonSend a User-Agent.gsrnamespace=6is required. Do NOT addfiletype:bitmap— it breaks the query and noquerykey comes back. Readimageinfo[0].thumburl. - Wikimedia geosearch — every photo taken near a real coordinate. This is how you
study a real place rather than a word.
https://commons.wikimedia.org/w/api.php?action=query&generator=geosearch&ggscoord=<lat>%7C<lon>&ggsradius=1000&ggslimit=10&ggsnamespace=6&prop=imageinfo&iiprop=url&iiurlwidth=1600&format=json - KartaView — street-level photography from a car windscreen, at any coordinate.
https://api.openstreetcam.org/2.0/photo/?lat=<lat>&lng=<lng>&radius=400Readresult.data[].fileurlProc. The most valuable source you have: it is your exact game camera — driver eye height, kerb to kerb, real parked-car spacing, real pole and cable spans, real sky. One of these tells you more about how a street reads than an hour of guessing. - Aerial/satellite imagery of a real city: see the map-data source doc.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 4d ago First seen · 71 lines · 49 tokens per session scan A a1e0648e0ad0
reference-images is a skill published in the GitHub repository ukanwat/aaabench (378 stars, last pushed 20d ago), licensed MIT. It adds 49 tokens to every session and 1,095 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 1 finding (makes network calls). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
benchjack
Audit an AI agent benchmark for hackability. Detects evaluation vulnerabilities like missing isolation, leaked answers, eval() on untrusted input, prompt injection in LLM judges, weak scoring, logic gaps, and trust of untrusted output. Use when analyzing whether a benchmark can be gamed or exploited.
unreal-pie
Domain skill - Unreal Engine Play-In-Editor (PIE) closed-loop verification. Provides PIE lifecycle control (enter/pause/resume/exit), controlled input injection (no OS dependency), viewport screenshot capture, output log snapshot, performance sampling, and Automation Test execution with async job polling. Use for…
unreal-playtest-agent
Domain skill - run screenshot-light PIE playtest episodes with structured entity observations, bounded semantic actions, transition polling, and in-memory traces for QA and external policy or RL runners.
unreal-actors
Domain skill - Unreal Engine actor management: list, spawn, transform, and delete level actors. Use when inspecting or editing actors in the current Unreal Editor level. Not for Content Browser asset import/export - use unreal-assets for that.
unreal-cinematics
Domain skill — Sequencer shot and keyframe orchestration, camera cut tracks, and Movie Render Queue job setup. Use when creating or editing Level Sequences, authoring camera animation, or preparing an MRQ job in Unreal Engine 5. Not for Blueprint scripting (unreal-blueprints) or runtime playback (unreal-runtime).
unreal-fab-assets
Acquire free Fab marketplace content through Unreal Engine's official Fab integration, then verify the imported Content Browser assets. Use when the user asks to find, download, add, or import Fab or Unreal Marketplace assets. Not for arbitrary file imports - use unreal-assets instead.