oldcircle/geo-sleuth写真がどこで撮影されたかを特定するエージェントスキル — OpenStreetMapのジオメトリ、標高スカイライン、衛星画像、ストリートビューを使い、その根拠を提示します。Claude Code、Codex、Cursor、Gemini CLI、OpenCode、GitHub Copilotで動作します。
<div align="center">
# 🧭 geo-sleuth
**An agent skill that finds where a photo was taken — and shows its work.**
<p align="center">
<a href="https://github.com/oldcircle/geo-sleuth/blob/main/README.md"><img alt="English" src="https://img.shields.io/badge/English-1f6feb?style=flat-square"></a>
<a href="https://github.com/oldcircle/geo-sleuth/blob/main/README.zh-CN.md"><img alt="Simplified Chinese" src="https://img.shields.io/badge/%E7%AE%80%E4%BD%93%E4%B8%AD%E6%96%87-dbeafe?style=flat-square"></a>
<a href="https://github.com/oldcircle/geo-sleuth/blob/main/README.ja.md"><img alt="Japanese" src="https://img.shields.io/badge/%E6%97%A5%E6%9C%AC%E8%AA%9E-dbeafe?style=flat-square"></a>
</p>
<p align="center">
<a href="https://github.com/oldcircle/geo-sleuth/blob/main/LICENSE"><img alt="License: MIT" src="https://img.shields.io/badge/License-MIT-yellow.svg"></a>
<a href="https://www.python.org/downloads/"><img alt="Python 3.10+" src="https://img.shields.io/badge/python-3.10+-blue.svg"></a>
<a href="https://agentskills.io"><img alt="Agent Skill: SKILL.md" src="https://img.shields.io/badge/Agent%20Skill-SKILL.md-8A2BE2.svg"></a>
<a href="https://github.com/oldcircle/geo-sleuth/blob/main/CONTRIBUTING.md"><img alt="PRs welcome" src="https://img.shields.io/badge/PRs-welcome-brightgreen.svg"></a>
<a href="https://github.com/Oldcircle/geo-sleuth/stargazers"><img alt="GitHub stars" src="https://img.shields.io/github/stars/Oldcircle/geo-sleuth?style=social"></a>
</p>
<p align="center">
<b>Works with</b><br>
<a href="https://code.claude.com/docs/en/skills"><img alt="Claude Code" src="https://img.shields.io/badge/Claude%20Code-D97757?style=for-the-badge&logo=claude&logoColor=white"></a>
<a href="https://developers.openai.com/codex/skills"><img alt="Codex" src="https://img.shields.io/badge/Codex-000000?style=for-the-badge"></a>
<a href="https://cursor.com/docs/skills"><img alt="Cursor" src="https://img.shields.io/badge/Cursor-000000?style=for-the-badge&logo=cursor&logoColor=white"></a>
<a href="https://geminicli.com/docs/cli/skills/"><img alt="Gemini CLI" src="https://img.shields.io/badge/Gemini%20CLI-1A73E8?style=for-the-badge&logo=googlegemini&logoColor=white"></a>
<a href="https://opencode.ai/docs/skills/"><img alt="OpenCode" src="https://img.shields.io/badge/OpenCode-211E1E?style=for-the-badge&logo=opencode&logoColor=white"></a>
<a href="https://docs.github.com/en/copilot/concepts/agents/about-agent-skills"><img alt="GitHub Copilot" src="https://img.shields.io/badge/GitHub%20Copilot-000000?style=for-the-badge&logo=githubcopilot&logoColor=white"></a>
<br><sub>…and any other agent that reads <code>SKILL.md</code> and runs shell commands.</sub>
</p>
<img src="https://assets.kitploit.com/production/public/readmes/57233/40ae2ae80ba3f250652ba57961b21f85ff1d24815d2f4b6ee9a48535072dd4a6/b5cf43e2092a767f6a84d4c07b7d8c06098f8c61d23d59ac060cb3423b77cb60-display-v1.webp" width="880" alt="From one photo to a camera position: the photo, the region scan, the skyline overlays, the evidence image">
*No text. No plates. No landmarks. One bridge, one mountain. Located to within 2 m.*
</div>
---
## Quick start
```bash
npx skills add Oldcircle/geo-sleuth
```
Pick your agents when prompted. Then hand your agent a photo and say:
> find where this photo was taken
On first use, ask the agent to run `doctor.py` from the installed skill’s `scripts/` folder and address any failed checks (see [Requirements and setup](#requirements-and-setup)). That is the whole interface. The agent reads `SKILL.md`, runs the scripts, and comes back with the camera position, the direction it was facing and a satellite evidence image. Prefer to copy the folder yourself? See [Installation](#installation).
## Why geo-sleuth
- **One photo, one sentence.** Give your agent a photo and say *find where this photo was taken*. You get back the camera position, the direction it was facing, and a satellite evidence image.
- **It works when there is nothing to read.** No sign, no plate, no landmark: OpenStreetMap geometry, elevation data, satellite tiles and street view carry the search on their own.
- **Geometry instead of guesswork.** Pier spacing becomes a distance ruler, shadows become a bearing, a ridge line becomes a fingerprint that elevation data can be matched against.
- **Every claim points at a file.** A conclusion has to name the command that ran in the session and the file it produced. Population and fame are not evidence.
- **Scripts rank, the model judges.** Twenty single-purpose scripts search, score and sort; the model only picks among the top few.
- **One skill, every agent.** A standard Agent Skill — `SKILL.md` plus plain Python scripts — so the same folder runs in Claude Code, Codex, Cursor, Gemini CLI, OpenCode and GitHub Copilot.
- **Answers carry an error radius.** Coordinates ± radius, the camera heading, an evidence image and a graded confidence.
## Cases
None of these photos had GPS data. Each case lists what the skill read from the photo, how it narrowed the search, and how far the result landed from the confirmed camera position.
Each case is a record of one run by one model. The skill tells the agent how to work; what to query, how to filter and how many top results to check are the model's own decisions, so a different model or different decisions may not reach the same result. Under each case, *Reproduction conditions and checkpoints* lists the input, the hint, the key parameters and what each stage should still cover, so you can find the step where your own run diverged.
<table>
<tr>
<td width="33%" align="center"><a href="#1-a-rice-paddy-and-a-viaduct"><img src="https://assets.kitploit.com/production/public/readmes/57233/882ec01d5cf5c4b4c2066f36bee4401ed4cb7516f03091448894c634067c7d22/2d8673969b8afd2fcd0582938ad0dcd126bf7be291106e1db2a1c3d209405f7a-display-v1.webp" alt="An oven at the edge of a rice paddy, a viaduct and a mountain behind"></a></td>
<td width="33%" align="center"><a href="#2-a-desert-skyline-one-degree-off"><img src="https://assets.kitploit.com/production/public/readmes/57233/c24c06416480a6fbfc197cc721e9751bc8c0ae442d1210ce97600b48ceed3b0f/d14e9205971dbd9c277350789f734c26e0c2809a57bb0c2bd028daf0bc5b2713-display-v1.webp" alt="A sand dune ridge and a range of bare dark mountains"></a></td>
<td width="33%" align="center"><a href="#3-a-cherry-tree-on-a-vancouver-street"><img src="https://assets.kitploit.com/production/public/readmes/57233/0529013c2a11217725ed0daa21afb7337cd1cfaa0cc9b7719fa3c57e3e0ffc6e/3ac476bfc68be6c91be541ba2ff808bd49d65506dc2cd180ac622e5d81b813bb-display-v1.webp" alt="A flowering cherry tree over a sidewalk"></a></td>
</tr>
<tr>
<td><b>1 · Rice paddy and viaduct</b><br><sub>Railway bridges × skyline × pier count.</sub></td>
<td><b>2 · Desert skyline</b><br><sub>Power lines × skyline × a phone tilted by 1°.</sub></td>
<td><b>3 · Cherry street</b><br><sub>Street-tree open data × shadows × street view.</sub></td>
</tr>
<tr>
<td align="center"><b>±2 m</b><br><sub>Qingyuan, Guangdong</sub></td>
<td align="center"><b>within 100 m</b><br><sub>Da Qaidam, Qinghai</sub></td>
<td align="center"><b>3 m</b><br><sub>Vancouver, Canada</sub></td>
</tr>
</table>
### 1. A rice paddy and a viaduct
A phone photo with the EXIF stripped: a white oven at the edge of a harvested rice paddy, a long viaduct in the distance, a steep mountain on the right. Not a single character in the frame. One message to an agent with this skill installed, and it came back with the camera position and the direction the camera was facing.
**photo → 27,335 → 171 → 14,372 → 22 → 3 → 1 → ±2 m**
| Step | What it did | Candidates left |
|---|---|---|
| **Read the photo** | Poles on the viaduct are catenary masts, so it is an electrified railway. Pier spacing used as a ruler (32 m span *assumed*): the left segment is about 0.5 km away, the right one over 1 km. A steep mountain about 3 km away. Rice harvested but grass still green, so no frost yet. | South China, as a bet, not a proof |
| **Region scan** | Pulled every railway bridge in the region from OpenStreetMap: **27,335 segments**. Sampled a point every 400 m and computed the 360° horizon from elevation data at each one. Kept points with flat ground nearby, a clear mountain within a few km, and a flat horizon next to it. | **171 sites** |
| **Skyline fit** | Placed candidate camera positions around each site and rendered the ridge line seen from each one: **14,372 positions**. The top 20 were within 0.1° of each other, so it added a constraint: the bridge must be near on the left and far on the right. | **22** |
| **Overlay check** | Drew the top three ridge lines back onto the photo. Score #1 (Fuzhou) had a bump hidden behind the oven, which is why it scored well. #3 (Huizhou) sloped where the photo is flat. #2 (Qingyuan) fit from the foot of the mountain to the edge of the frame. | **1** |
| **Pier count** | 17 piers in the photo become 17 bearings from the camera. Where they hit the railway line, the intersections must be evenly spaced. Combined with the skyline: first a band about 300 m long, then a single spot. | **±2 m** |
<div align="center">
<img src="https://assets.kitploit.com/production/public/readmes/57233/543d1f592ac283bab887b3b6e66187ad0efae04cc9b98d26384a8cd5950be1a1/704c7423c6ca8f9af8cabfe5f02bbdda6a6ca7bc9291d86992378586839ef5a3-display-v1.webp" width="820" alt="17 piers marked on the viaduct, spacing used as a ruler"><br>
<sub>Piers as a ruler: wide spacing on the left means near, tight spacing on the right means far.</sub><br><br>
<img src="https://assets.kitploit.com/production/public/readmes/57233/5cdc5c06cf67b32ba0ef88f6f4f12b7624d26667c4172c29acf9b2bf0179a0d7/ddd087c5f499e06536b3a87a6220487ae53b4f73ac0a444fe3afd696a805d299-display-v1.webp" width="520" alt="Top three skyline overlays: Fuzhou, Qingyuan, Huizhou"> <img src="https://assets.kitploit.com/production/public/readmes/57233/4f4665122277f510573bd405f173ef1347fcb5e01cc7ecb893ae3197019c1bff/853b5fddd3225939ca6a9422fd320a8a87f732f690b707adbe58625fcdf8c87d-display-v1.webp" width="292" alt="Evidence image: camera position, field of view, the railway and the mountain"><br>
<sub>Left: the top three ridge lines drawn onto the photo. Right: the evidence image the skill produced.</sub>
</div>
<details>
<summary>More figures from this run</summary>
<br>
<img src="https://assets.kitploit.com/production/public/readmes/57233/1b6a76662fce75ace9683df6e13f8e322de9098c85106280d3836618f5158eaa/c020739f4ba3da17062a996d29fb99f6cf9a96dec764b3e8fdb9110faa226ce3-display-v1.webp" width="720" alt="Region scan: railway bridges in grey, candidate sites in orange"><br>
<sub>Region scan: every railway bridge in the region (grey), sites that pass the horizon test (orange).</sub><br><br>
<img src="https://assets.kitploit.com/production/public/readmes/57233/eaa18ab3ad13c3893be898b4e377eb7eda7795fbb7b38cdd815bcd46c69a7cec/ae7ba7dcf1f40a95ecc479d148ef056832b529d365cd427f52c3ab312376a71b-display-v1.webp" width="720" alt="Bearings to 17 piers intersecting the railway line"><br>
<sub>Pier count: bearings to the 17 piers intersect the line; only one camera position makes the spacing even.</sub><br><br>
<sub>The run took about 72 minutes end to end, roughly half of it waiting on computation.</sub>
</details>
<details>
<summary>Reproduction conditions and checkpoints</summary>
<br>
- **What this record is**: the table is the first time this method solved the photo (Claude Opus 5.0), not a blind re-run; its steps were later written into the skill.
- **Input**: a 1280×1707 phone photo with EXIF stripped. The images in this README are display versions and are not suitable as input.
- **Candidates**: all 27,335 OSM railway-bridge segments in the region, excluding only those tagged `electrified=no` (as `rail_mtn_scan.py` did then and `terrain.py scan` does by default now), with no `highspeed` filter. Catenary masts only show that the line is electrified; the correct line, the Guangzhou–Qingyuan intercity, has 55 bridge segments in OSM and none is tagged `highspeed=yes`.
- **Checkpoints**: the 171 sites that pass the horizon test should include intercity bridge segments near the answer; the skyline fit covers 14,372 camera positions, the top 20 lie within 0.1°, and "bridge near on the left, far on the right" cuts them to 22; the top three are overlaid; the 17 pier bearings fix the point.
- **Case scripts**: `examples/rail-skyline-session/` holds the scripts tuned to this photo. OSM splits one line into many ways with the same name; `joint.py` tries every same-name way near the candidate center in turn, so the result does not depend on the order the query returns them.
</details>
### 2. A desert skyline, one degree off
A sand dune in the foreground, a range of bare dark mountains behind it, and a few pylons at the foot of the range. There is no text, road or building in the frame, and reverse image search returned only generic desert photos from northwest China.
<div align="center"><img src="https://assets.kitploit.com/production/public/readmes/57233/a74a438373e9d057ce59e7bd0863ac9d2f9d97e34d63d2db5fdcc6c54e43fdaa/5d629bfef73d4b2170bcd61823db70648025e34f839d50e6efe07c3735b561fe-display-v1.webp" width="820" alt="The photo: a sand dune ridge, bare dark mountains, pylons at the foot of the range"></div>
**photo → 1,537 sites → stuck around 10 px → solve a 1° tilt → 1 site → one dune**
| Step | What it did | Left |
|---|---|---|
| **Read the photo** | The pylons are the only man-made objects, so the camera is within a few kilometres of a power line. Sand, gravel and bare rock point to northwest China. | 4 provinces |
| **Power lines × terrain** | `osm.py geom` pulled every power line in Xinjiang, Gansu, Ningxia and Qinghai. `terrain.py scan` worked along them with elevation data and kept the places where mountains fill the view. | **1,537 sites** |
| **Skyline fit** | `terrain.py ridge` traced the ridge in the photo, and `terrain.py fit` compared it with the ridge seen from each site. In Qinghai the top 25 sites all scored between 9.6 and 11.7 px; the true site was 9th. | **stuck** |
| **The tilt** | A phone held 1° off level moves the ridge at the edge of a 2048 px frame by 1024 × tan 1° ≈ 18 px, about 10 px on average, which is the size of the gap between the sites. At the true site the error across the frame lies on a line whose slope is tan θ, with θ ≈ 1°. `fit` was changed to solve the roll together with the horizon offset. | |
| **Fit again** | With the roll solved, the true site went from 11.1 to 6.2 px. The next best site was at 9.4 px. | **1 site** |
| **Satellite** | Within 1 km along the line of sight there are three crescent dunes and a tent camp. A 750/330/110 kV bundle runs 1–2 km to the east, which is why the pylons appear only on the left of the photo. The camera stood on the crest of the dune west of the camp, facing 205°. | **one dune** |
<div align="center">
<img src="https://assets.kitploit.com/production/public/readmes/57233/22ce1bb037e1110875246a441826b85a4cdb7e780ea28e2fc634469d81442368/f8ae37a34d6c057c73abb80f47e2fed80a641e4d0ca3f79b0139b74276356775-display-v1.webp" width="820" alt="Power lines in Qinghai, Gansu and Ningxia, candidate sites in orange, the camera as a red star"><br>
<sub>Power lines from OSM (gold), sites where mountains fill the view (orange), and the camera (red star). Xinjiang's sites are not drawn.</sub><br><br>
<img src="https://assets.kitploit.com/production/public/readmes/57233/1c486e619ab76b3a6cdc8186455ff71cacefffe75ac85e253db37a521f403988/aab9002fe325d5877ba7465fc2e5fb05b95b79617186da09ec2cb771a22db065-display-v1.webp" width="820" alt="Ridge in the photo against the computed ridge at the true site, before and after solving the roll, with the error plotted across the frame"><br>
<sub>The true site, before and after solving the roll. Top: the error grows from one side of the frame to the other, and its slope gives the angle. Bottom: with the roll solved, the trend is gone.</sub><br><br>
<img src="https://assets.kitploit.com/production/public/readmes/57233/f7c403801bcb88b05d04891840396339de070fecf7489fc338f18d8c87184d9b/10809485b8741490c7295655d6367a9ee29d6e4824efe92b505109acda83f315-display-v1.webp" width="820" alt="Skyline error of 408 Qinghai sites assuming a level camera, and of the 12 best sites with the roll solved"><br>
<sub>Skyline error per site. Assuming a level camera, the true site sits inside the pack; with the roll solved, it is the only one below 9 px.</sub><br><br>
<img src="https://assets.kitploit.com/production/public/readmes/57233/6eb3bb67fa010d7fa0fd63e6acc0e7aaa4c5de7936f81a8f764e898e6d60d73b/15846a4d5798355446f50b4832cfc027cfd44556fd232bdf68ed62fc395a1254-display-v1.webp" width="600" alt="Evidence image: the camera on the dune crest, field of view, the dunes, the camp and the power-line bundle"><br>
<sub>Evidence image. The OSM power lines (gold) run over the pylons visible in the satellite image.</sub>
</div>
**Blind re-run.** Later we gave the photo to a new agent running the current skill, with one hint: it is in China. It scanned the 5,334 power lines in Qinghai, fit 792 sites, and picked one at 0.224° against 0.357° for the next. It ended on the same dune crest, **87 m from the confirmed camera position**, after 32 minutes. The roll fit added for this photo is now part of `terrain.py fit`; [Contributing](#contributing) explains why this case is the reference for changes to the core scripts.
<details>
<summary>Reproduction conditions and checkpoints</summary>
<br>
- **Two runs**: the table is the original solve that led to the roll fit (Claude Opus 5.0); the blind re-run came later, with a new agent on Claude Opus 5.5.
- **Blind run input and hint**: the original 2048×1152 photo; the only hint was "it is in China". The images in this README are display versions.
- **Blind run checkpoints**: 5,334 power lines in Qinghai; `terrain.py scan` with the default `--cluster-km 3.5` turned 14,843 hits into 792 clusters; all 792 went into `terrain.py fit` (in 6 batches), with no satellite-ranking cut before the skyline fit; the best site scored 0.224° against 0.357° for the next; satellite imagery along the line of sight then found the dunes and the camp. 87 tool calls, 32 minutes.
- **Clusters to fit**: `fit` lays its camera grid around each cluster's representative point only. With a larger `--cluster-km`, hits far from the representative may never be fit; `scan` now records each cluster's reach (`extent_m`) and `fit` prints a WARNING when a cluster is wider than its `--radius`.
- **Example**: `examples/desert-roll-fix/` reproduces the fit before and after the roll change on the fixed 12 sites (true site from #8 at 12.0 px to #1 at 6.2 px).
</details>
### 3. A cherry tree on a Vancouver street
A flowering cherry over a sidewalk. The plates and street furniture place it in Vancouver. Finding the street is the hard part, because the city has close to ten thousand blocks.
**photo → Vancouver → 1,095 big cherries → 97 blocks → 72 slope checks → W 60th Ave → 3 m**
| Step | What it did | Left |
|---|---|---|
| **Read the photo** | White-and-blue BC plates, a dark-green streetlight in a grass boulevard, Vancouver Special houses. | **Vancouver** |
| **Street-tree data** | Vancouver publishes every street tree with species, trunk diameter and address. The candidates came from this data rather than from well-known cherry streets: flowering cherries with a trunk of at least 40 cm, then blocks with at least three of them. | **1,095 trees, 97 blocks** |
| **Shadows and slope** | The cars' shadows fall to the left and toward the camera, so the sun is ahead and to the right. The yards on the camera side sit above the sidewalk behind rock walls, so that side of the street is uphill. The agent measured the cross-slope of 72 blocks from elevation data and checked which side the big trees stand on. | **a few blocks** |
| **Tree by tree** | On the 100 block of W 60th Ave, the north side is lined with large cherries. On the south side, the trees 30–60 m ahead are 7.6 cm saplings, and the next large one is about 75 m away. The photo shows the same thing: nothing tall on the near right and pink canopy in the distance. | **1 block** |
| **Street view** | The 2009 and 2024 captures show the rock retaining wall with stone steps, the green streetlight and the white house. The house with a rooftop deck is in the 2024 capture and in the photo. The camera was on the north sidewalk, facing east. | **3 m** |
<div align="center">
<img src="https://assets.kitploit.com/production/public/readmes/57233/0fb6691aead52a5b2704c7016372380c8022f47e3f0c06081d50d530d4cad2e7/8033c1d4c52a54840a1cb775f8166a3f610a628c6a46dbad6066b4d3321d6a4e-display-v1.webp" width="464" alt="Vancouver's street trees, the large flowering cherries, the candidate blocks and the answer"> <img src="https://assets.kitploit.com/production/public/readmes/57233/810e71d15111a798569a543a45036602358e6b890ed8f9a7e654b2082f2b6836/2aed94f4f64b7e89750fe8dc67a80fa55e8383499c13224169a6b1a49e0788ab-display-v1.webp" width="388" alt="The photo with the clues it gives: streetlight, plates, car shadows, raised yards, rooftop deck"><br>
<sub>Left: all 184,517 street trees with coordinates in the city's data, the 1,095 large flowering cherries and the 97 candidate blocks. Right: what the photo gives away.</sub><br><br>
<img src="https://assets.kitploit.com/production/public/readmes/57233/49c8f63ee9fd07fdc7a8089d7961b7201d039cc094f33b8b65595bbf456af3de/04e9302013c9cd9cc96ef80252fa08a9e0a816a2106ce9ebb713e60b632acafd-display-v1.webp" width="820" alt="The photo next to street view captures from 2024 and 2009 on the same block"><br>
<sub>The photo was taken from the sidewalk and the street view car drove down the middle of the road, so the angles differ.</sub><br><br>
<img src="https://assets.kitploit.com/production/public/readmes/57233/9310f1aac99eb2ad7f28b13bcdfbf651656eb0de6d8ea896cfa0f29ed5ddba01/b10c721a3ae78626a1f00a1e37ac44934ecc23dfd74af151c4075f10bb39a892-display-v1.webp" width="600" alt="Evidence image: W 60th Ave 100 block, camera on the north sidewalk facing east, street trees from the city data"><br>
<sub>Evidence image. Pink rings are large cherries from the city's data, green rings are saplings.</sub>
</div>
**Blind re-run.** The table above comes from a blind run: a new agent running the current skill, given the photo and no hint. It finished in 53 minutes, **3 m from the confirmed camera position**.
<details>
<summary>Reproduction conditions and checkpoints</summary>
<br>
- **What this record is**: the table is the blind run itself (Claude Opus 5.5, a new agent, the photo and no hint): 154 tool calls, 53 minutes.
- **Input**: the original 1440×1920 photo. The images in this README were cropped and then downscaled to about 1/16 of the original pixels, and are not suitable as input.
- **Checkpoints**: 184,517 street trees with coordinates in the city's data → 1,095 ornamental cherries with trunk diameter ≥ 40 cm → 97 street segments with at least three → 72 segments where the cross slope could be measured → a few left after matching which side the large trees stand on and which way the slope falls.
- **Street-view check**: the whole-image `match.py` ranking was poor (peak bloom, and the photo was taken from the sidewalk while the street-view car drove down the middle), so the run compared only the tree region with `--query-box 0,900,1440,1600`, then checked 2009 and 2024 street view for stable features: stone walls, steps, lamp posts. The true spot had only 0–5 SIFT inliers and was not rejected on that basis. In `match.py` output, empty `inliers` (`refined: false`) means the candidate was not refined, not that it had 0 matches.
</details>
## Installation
geo-sleuth is a standard [Agent Skill](https://agentskills.io): one folder holding `SKILL.md`, `scripts/`, `references/` and `data/`. Install it with the [`skills`](https://github.com/vercel-labs/skills) CLI, or copy the folder yourself.
**All six agents, user-wide, one command:**
```bash
npx skills add Oldcircle/geo-sleuth -g -a claude-code -a codex -a cursor -a gemini-cli -a opencode -a github-copilot -y
```
**By hand:**
```bash
git clone https://github.com/Oldcircle/geo-sleuth
mkdir -p ~/.agents/skills ~/.claude/skills
cp -r geo-sleuth/skills/geo-sleuth ~/.agents/skills/ # Codex, Cursor, Gemini CLI, OpenCode, GitHub Copilot
ln -s ~/.agents/skills/geo-sleuth ~/.claude/skills/geo-sleuth # Claude Code
```
`~/.agents/skills/` is read by Codex, Cursor, Gemini CLI, OpenCode and GitHub Copilot, so one copy there covers all five. Each agent's own folders, from its docs:
| Agent | User-wide | Per project |
|---|---|---|
| [Claude Code](https://code.claude.com/docs/en/skills) | `~/.claude/skills/` | `.claude/skills/` |
| [Codex](https://developers.openai.com/codex/skills) | `~/.agents/skills/` | `.agents/skills/` |
| [Cursor](https://cursor.com/docs/skills) | `~/.cursor/skills/` or `~/.agents/skills/` | `.cursor/skills/` or `.agents/skills/` |
| [Gemini CLI](https://geminicli.com/docs/cli/skills/) | `~/.gemini/skills/` or `~/.agents/skills/` | `.gemini/skills/` or `.agents/skills/` |
| [OpenCode](https://opencode.ai/docs/skills/) | `~/.config/opencode/skills/` or `~/.agents/skills/` | `.opencode/skills/` or `.agents/skills/` |
| [GitHub Copilot](https://docs.github.com/en/copilot/concepts/agents/about-agent-skills) | `~/.copilot/skills/` or `~/.agents/skills/` | `.github/skills/` or `.agents/skills/` |
Any other agent that reads `SKILL.md` and runs shell commands works the same way: put the folder where it looks for skills.
## How it works
The work is split into three layers. Scripts decide, scripts perceive and rank, the model only judges among the top few.
```mermaid
flowchart LR
A["photo"] --> B["intake.py<br/>EXIF · OCR · reverse image search"]
B --> C["board.py<br/>candidate board: clues, likelihood ratios, ranking, next step"]
C --> D{"which branch?"}
D --> E["sun.py · terrain.py · osm.py · pose.py<br/>shadows, skylines, OSM corridors, camera pose"]
D --> F["sat_scan.py · match.py · gsv.py · baidu_pano.py<br/>CLIP-ranked satellite tiles, DINOv2+SIFT street view"]
E --> G["board.py check · report"]
F --> G
G --> H["evidence.py<br/>coordinates ± radius · evidence image · graded confidence"]
```
| Layer | Who | Tools |
|---|---|---|
| **Decide**: which candidates, how evidence scores, what can be excluded, where to scan next | scripts (the candidate board) | `board.py` |
| **Perceive**: read text, look up tables, find targets in satellite tiles, compare street view | scripts rank first, a person looks at the top few | `intake.py` `ocr.py` `clues.py` `sat_scan.py` `match.py` `geo.py` |
| **Judge**: pull clues from the frame, propose hypotheses, pick among the ranked few | the model | `SKILL.md` + `references/` |
Every conclusion has to point at a command that actually ran in the session and the file it produced. Exclusions need read or computed evidence; observations and guesses can only lower a candidate's weight.
## Toolbox
Twenty-odd scripts, one job each. The full table with data sources is in `skills/geo-sleuth/references/data-sources.md`.
| What it does | Script |
|---|---|
| Environment checks, browser launch and optional network probes | `doctor.py` |
| EXIF: GPS, capture time, equivalent focal length, heading | `exif.py` |
| OCR on the whole image, zoomed crops and tiles (Apple Vision on macOS, RapidOCR elsewhere) | `ocr.py` |
| Reverse image search on Baidu and Yandex, similar images tiled into a numbered sheet; keyword image search | `revimg.py` |
| Steps 0–3 in one command: metadata, edge crops, variants, OCR, reverse search → `intake.md` | `intake.py` |
| Zoom crops, edge and corner crops, tiling, pixel columns of evenly spaced structures such as piers | `imgprep.py` |
| Lookup tables: calling codes, driving side, dependent territories worldwide; plate prefixes, area codes and admin divisions from region packs | `clues.py` + `data/` + `regions/` |
| Region packs: list them, print a country's card and clue index, lint a pack | `regions.py` |
| Candidate board: candidates, clues, likelihood ratios, exclusion, ranking, scan order, pre-report checks | `board.py` |
| Gazetteer: list sub-divisions with bounding boxes, built-up area extent | `gazetteer.py` |
| Place, compound or shop name → coordinate candidates, every namesake listed | `poi.py` |
| Sun and shadows: latitude band, time of day, street orientation, heading from lit faces, true bearings | `sun.py` |
| OSM Overpass: feature co-occurrence, line-to-point, route corridors, line intersections, street grid templates | `osm.py` |
| Satellite tile mosaics, markers, numbered thumbnail sheets | `tiles.py` |
| CLIP zero-shot scoring of satellite grid cells or candidate points (tracks, factories, silos, dams…) | `sat_scan.py` |
| Baidu panoramas / Google Street View: find points, render headings, thumbnail sheets, historical batches | `baidu_pano.py` `gsv.py` |
| Rank candidate ground-level images against the photo: DINOv2 global similarity + SIFT inliers | `match.py` |
| Elevation: synthetic mountain views, skyline overlays, profiles; linear feature × terrain scan, ridge extraction, batch skyline scoring | `terrain.py` |
| Multi-point camera pose: lat/lon, height, heading, pitch, roll, with error radius | `pose.py` |
| Bearings, distances, line-of-sight intersections, alignment lines, frame/occlusion checks, camera position from evenly spaced structures | `geo.py` |
| Evidence image: satellite tile + camera fan + comparison grid | `evidence.py` |
The steps from case 1 (region scan, batch skyline scoring, camera position from pier spacing) are built into the skill as subcommands: `terrain.py scan / ridge / fit`, `imgprep.py piers`, `geo.py spacing`; case scripts tuned to that photo are kept in `examples/rail-skyline-session/`. The tilt from case 2 is solved by `terrain.py fit` itself (`--roll-max`, default 1°); `examples/desert-roll-fix/` reproduces the before and after.
## Benchmarks
Per-operator measurements:
| Script | Test | Result |
|---|---|---|
| `match.py` | 8 cases: a historical Baidu panorama batch rendered as the photo, panoramas within 150 m as candidates (Shenzhen) | ground truth ranked 1/2/4/1/1 and 5/1/6, all in the top 6, half at #1 |
| `sat_scan.py` | 4×8 km, 364 cells at z17, 40 OSM-tagged running tracks as ground truth, multi-scale (Shenzhen) | recall@20 17/40, @30 22/40, @100 32/40, median rank 23 |
| `terrain.py scan / fit` + `geo.py spacing` | bounded re-run on the case photo above | true cluster ranks #1, final position about 2 m from ground truth |
| `terrain.py fit --roll-max` | 8 synthetic skyline cases (random heading, focal length, ±1° roll) | median position error 324 → 238 m (untilted cases 324 → 89 m, tilted 1,695 → 265 m) |
| `terrain.py fit --roll-max` | the 12 best sites from the desert case | true site #8 at 12.0 px → #1 at 6.2 px (runner-up 9.9 px) |
| `clues.py` | 6 tables, 9 values spot-checked | 9/9 correct |
End to end, blind re-runs with a fresh agent that had only the photo (and the hint shown):
| Case | Hint | Result | Time |
|---|---|---|---|
| [Desert skyline](#2-a-desert-skyline-one-degree-off) | "it is in China" | 87 m from the confirmed camera position | 32 min |
| [Cherry street](#3-a-cherry-tree-on-a-vancouver-street) | none | 3 m from the confirmed camera position | 53 min |
Both photos had been solved before and lessons from them are in the skill, so these are regression checks, not accuracy on unseen photos.
The method comes from breaking down 14 videos by online-geolocation creators, 22 puzzles and a set of real runs, then turning what works into rules and scripts. v2 moves every rule that can be code into `board.py`, so the rules get executed, not just read.
## Requirements and setup
You need Python 3.10+, [`uv`](https://docs.astral.sh/uv/), `curl`, and an agent that can run shell commands. Each script declares its own dependencies; use `uv run`, which installs them on first use. Keep the complete skill folder, including `data/` and the helper modules in `scripts/`.
Reverse image search requires **Google Chrome or Playwright Chromium**. The scripts try Chrome first and automatically fall back to Chromium. If neither is installed:
```bash
uvx playwright install chromium
```
---
[Read more](https://github.com/oldcircle/geo-sleuth)