Loading...
Loading...
Determine where a photo was taken (photo location guessing / internet sleuthing / GeoGuessr-style play / infer capture time). Given one or several photos, a single command first performs metadata, OCR, and reverse image search (`intake.py`), records clues and candidates on the candidate board (`board.py`) for script ranking and next-step suggestions; lookup-table clues use `clues.py`; satellite and street view both follow the principle that the machine ranks first and humans only look at the top hits (`sat_scan.py`, `match.py`); every conclusion is checked against real data and output with coordinates + error radius, evidence montage, and tiered confidence. Geolocate or chronolocate a photo with tool-verified reasoning — EXIF, OCR, reverse image search (Baidu/Yandex), lookup tables (plates, area codes, calling codes, driving side, territories), candidate board with likelihood ranking, sun and shadow math, OSM Overpass, CLIP-ranked satellite scan, DINOv2+SIFT-ranked street view matching, DEM skyline rendering. Use when the user shares a photo and asks where this is / where it was taken / help me geolocate this photo / internet-sleuth this / what time this was taken / where was this taken / geolocate this.
npx skill4agent add oldcircle/geo-sleuth geo-sleuthin commands refers to the directory containing this SKILL.md (same for commands in references). Claude Code replaces it automatically; for other agents, first run${CLAUDE_SKILL_DIR}(if the shell does not preserve variables across commands, put it before every command), or replace it with that path directly.export CLAUDE_SKILL_DIR=<absolute path to this directory>
| Layer | Who does it | Tool |
|---|---|---|
| Decision: what candidates exist, how evidence is scored, what can be excluded, where to scan next | Scripts (candidate board) | |
| Perception: reading text, lookup tables, finding targets on satellite imagery, street-view comparison | Scripts compute and rank first; humans only look at the top hits | |
| Judgment: extracting clues from the scene, hypothesizing when lookup tables have no answer, adjudicating among the machine-ranked top hits | You | — |
board.pyboard.pypose.py project --horizon <horizon row>terrain.py elevsun.py compassgeo.py bearingsboard.py childrenclues.py lookup territories <A国> --continent <洲>board.py reportboard.py excludegeo.py frameterrain.pyevidence --againstboard.py add --from <output of poi.py --out / osm.py geom> --level area/roadevidence --againstcheckexclude--covers lat,lon[:lat,lon]board.py nextuv run ${CLAUDE_SKILL_DIR}/scripts/intake.py photo.jpg --out-dir intake/ [--box x0,y0,x1,y1 ...] [--proxy socks5://127.0.0.1:10808]
uv run ${CLAUDE_SKILL_DIR}/scripts/board.py init --photo photo.jpgintake.pyrev/intake.mdintake.mdedges/references/observe.mdboard.py clue "<text>" --kind <type> --status observed|read|inferred|computed --file <crop>clues.py lookup <kind> <value>board.py apply --kind plate --value 渝G --file <crop>rev/<name>_baidu_similar.jpgsimilar[i].frompoi.py "<name>" --city <city> --out pois.jsonboard.py add --from pois.json --level areareferences/search.mduv run ${CLAUDE_SKILL_DIR}/scripts/board.py children <上级行政区名> # 直辖市 → 全部区县;国家 → 一级行政区(gazetteer.py 查 OSM,带 bbox)
uv run ${CLAUDE_SKILL_DIR}/scripts/board.py evidence --clue K1 --for <候选A>:5 --for <候选B>:2 --against <候选C>:0.3 --why "…" --file <比对图>
uv run ${CLAUDE_SKILL_DIR}/scripts/board.py rank
uv run ${CLAUDE_SKILL_DIR}/scripts/board.py nextnexttiles.py fetch --zoom 13terrain.pyrevimg.py --query "<城市> <颜色> 公交"baidu_pano.py sample --bbox <built-up area> --n 24nextboard.py urban <candidate>scan-bboxboard.py falsify <candidate> --text "…"references/clues/| What you have | What to do | Read |
|---|---|---|
| Unique anchor (building, statue, tower, scenic-area structure) | Anchor geometry: line-of-sight intersection, alignment lines, tangents, filter buildings by camera height | |
| Distant mountains, skyline | Render candidate camera positions with | |
| Clear shadows, lit faces, sun in the frame | | |
| Route number, rail/power-line specifications, large river | Linear corridor + line-change points: | |
| Two or three infrastructure types in one frame | | |
| Linear facility + unrecognizable mountains, no text | Compute the horizon with elevation along the facility, filter the whole large region → numerically score skyline + facility distance → overlay top 3–5 → locate the camera from evenly spaced components: | |
| Street layout clear, no anchor | Filter intersections with | |
| Image search has no near-duplicate, no text, only a set of scene elements (road spec + adjacent features + landform) | Don't hand-pick candidate points first: describe the scene as one English query, let | |
≥4 points in the image with known positions (window views, top-down shots; eye-level shots of distant tower/bridge/waterfront also work, with | Solve camera position, heading, and height with | |
| Want to exclude using “X is not in the frame” | First use | |
| Recognizable facility type (drying tower, feed mill, sugar refinery, cement station) | Look up industry distribution → enumerate large buildings with | |
| Chain-brand sub-brand store | First search opening press releases for addresses; use locators only as a candidate pool → | |
| Window seat / drone top-down | Aerial branch | |
uv run ${CLAUDE_SKILL_DIR}/scripts/sat_scan.py grid --bbox <scan_bbox> --zoom 17 --preset track --multi-scale --top 30 --out sat.json --sheet sat_top.jpg --heat heat.jpg
uv run ${CLAUDE_SKILL_DIR}/scripts/poi.py "<区县> 学校" --city <地级市> --out schools.json # 种子:--seeds schools.json 给附近格子加分
uv run ${CLAUDE_SKILL_DIR}/scripts/sat_scan.py points --points big.json --preset factory --out r.json --sheet r.jpg # osm.py buildings / poi.py 的候选点排序trackstadiumfactorysilodambridgequarrysolargreenhouseport--queryboard.py checkuv run ${CLAUDE_SKILL_DIR}/scripts/baidu_pano.py scan <lat,lon> --radius 300 --out panos.json # 国内;国外 gsv.py
uv run ${CLAUDE_SKILL_DIR}/scripts/match.py rank --query photo.jpg --panos panos.json --toward <地标lat,lon> --spread 15 --refine sift --top 10 --out m.json --sheet m.jpg
uv run ${CLAUDE_SKILL_DIR}/scripts/match.py rank --query photo.jpg --items around.index.json --render gsv --spread-headings -30,0,30 --out m.json --sheet m.jpgmatch.py--spread-headings--withingsv.py sheet --datesheet --road <road_name> --spread 60terrain.py view --photogeo.py intersectgeo.py linepose.pysun.py facing --lit left --shaded camerauv run ${CLAUDE_SKILL_DIR}/scripts/board.py check # 出结论前:排除有文件吗、主答案有核实证据吗、哪些线索没用上
uv run ${CLAUDE_SKILL_DIR}/scripts/board.py report --merge result.json # 主答案=第一名;备选、排除、未用线索自动写进 result.json
uv run ${CLAUDE_SKILL_DIR}/scripts/evidence.py spec.json --out evidence.jpggeo.py convertboard.py report| Level | High | Medium | Low |
|---|---|---|---|
| City | Text, license plate, area code, or confirmed anchor | Multiple independent weak clues agree | A single weak clue or hint only |
| District | The unique facility or landmark matches ≥3 top-down features on satellite imagery (all candidate areas inspected, no second occurrence) | The most plausible place in the candidate area, with uninspected candidates remaining | Inference |
| Road | Ground-level view matches ≥2 unique features | Layout matches with 1 unique feature; when no ground photo exists: district level is high + bearing self-check passes | Satellite layout only |
| Building | Two independent constraints + ≥3 ground-view features | Only one of these | Inference |
| Floor | Two reference types cross-check | Single reference | Not given |
intake.pyboard.py nextsat_scan.pymatch.pyosm.py coveragesat_scan.py gridboard.py reportboard.py checkuv run ${CLAUDE_SKILL_DIR}/scripts/xxx.pymatch.pysat_scan.pyscripts/revimg.pyintake.pyocr.pyGEO_PROXYexport GEO_PROXY=socks5h://...127.0.0.1:10808intake.py --proxysocks5://.geo-cache/board.jsonreferences/data-sources.mdtimeout$varbash -c${=var}