How-to
Send a coding agent digging through public data
A coding agent can sift a public dataset for you and sound sure of every find. This research brief makes it plan first, log its guesses and check each claim against known records before it reports.
Pietro Miotti, who says he is not an archaeologist, wrote in a post that spread on 11 November that his AI agents had found a Maya site in Mexico in NASA lidar data, and that Mexico’s National Institute of Anthropology and History (INAH) had no registered site at that spot.
He says the whole arc, from a random thought while travelling to INAH’s reply, took less than 72 hours. The same day he added an update: archaeologists had already mapped it in Estrada-Belli et al. 2025, a paper the agents missed while searching the literature, so in his words the agents “rediscovered” the site.
His method is worth copying, and the miss shows exactly which check to build in.
The prompt: a research brief for a coding agent
Paste it into Claude Code, Codex or a similar agent, opened in an empty project folder, and fill in the set-up block first.
You are running a research job on a public dataset for me. Work like a careful research assistant who has to defend every claim to an expert in the field. If you report something as new when it is already on record, or invent a result, I could embarrass myself in front of the people who keep those records and waste their time. If you are too cautious, I miss a real lead. So be exact about what you checked and what you did not. MY SET-UP (I fill this in once; any line can say "not sure") - My question: [e.g. "are there regular, human-made shapes in this lidar area that could be old structures", "which bridges in this city's open inspection data were rated poor and never re-inspected"] - What counts as a hit: [e.g. "straight edges and right angles, at least 20 m across"; or "not sure, propose criteria"] - What to reject: [e.g. "natural hills, sinkholes, roads, field edges"; or "not sure"] - The dataset: [name, link, and version or date] - Its licence or terms: [paste the wording from the download page, or "not read"] - Known records to compare against: [e.g. "the official register at [link]", "published papers on this area", "not sure"] - A known example to test on: [e.g. "a registered site at [location]", "a bridge I know was rated poor", or "not sure, pick one from the known records"] - Size of this run: [e.g. "one area of about 50 km2 first", "the first 1,000 rows"] - Is any of this sensitive: [e.g. "site locations that could attract looters", "personal data", "no"] - What I will do with the result: [e.g. "send it to the body that keeps the register", "write a blog post", "just for me"] RULES FOR THE WHOLE JOB - Never invent a record, source, citation, number or location. If something is missing, write NOT STATED. If you need something from me, write [NEED X] and wait. - Never call a find "new" or "undiscovered". The strongest label you may use is "not found in the sources I searched", followed by the list of those sources. - Once I approve the plan, keep a file named assumptions-log.md in this folder, starting with any assumptions in the plan. Add a line every time you assume something (a unit, a threshold, a missing value, a file format, a coordinate system), with why, and how I could test it. - Stay within the size of this run. Ask me before you download more, run anything that costs money, or send anything outside this folder. - If the data is sensitive, keep exact locations and personal details only in a separate file named restricted-details.md, and leave them out of the main report. - Keep progress.md listing every item done and skipped, so a fresh session can pick up where this one stopped. STAGE 1. CHECK THE SET-UP AND THE LICENCE Read my set-up. Treat any line that is blank or still shows the example text as blank. If my question, the dataset or the licence line is blank, ask for them once and wait. If the licence says "not read", tell me to read it and wait. If I say the page has no licence, tell me that public does not mean free to reuse, and ask before going on. If the licence looks like it bars what I said I will do with the result (for example commercial use, sharing the raw data, or publishing work built on it), quote the words and stop until I reply. If "what counts as a hit" or "what to reject" says "not sure", propose three to five concrete hit criteria and three reject rules, and wait for my yes. If "is any of this sensitive" says "not sure", treat it as sensitive. STAGE 2. PLAN, THEN WAIT Write a plan and change nothing yet. Cover: the files you will fetch and from where; how you will turn them into something you can examine (tables, images, slices); the exact hit criteria and reject rules, written so every item is judged the same way; how you will split the work; what you will save and where; and roughly how much data and time you expect. List what could produce a false hit in this kind of data (for example tree cover, gaps, duplicates, processing artefacts, a wrong coordinate system). Name the test slice that contains my known example. If my known example says "not sure", pick one from the known records and say where you found it. If you cannot find a known example in the records, say so, mark the method UNTESTED in the plan, and ask whether to go ahead. End with "Approve this plan?" and wait. If I change anything, rewrite the plan and ask again. STAGE 3. TEST ON THE KNOWN EXAMPLE, THEN WAIT Once I approve, run the criteria on the test slice only. Report whether they found my known example, and what else they flagged in that slice. If they missed it, propose changes to the criteria and wait for my yes before going further. If I said to go ahead UNTESTED, skip this stage and carry the UNTESTED label into the report. STAGE 4. RUN Work only inside the approved plan. Apply the same written criteria to every item. For each candidate, save: an ID, where it is, which criteria it meets, which reject rules you checked, a confidence of high, medium or low with one line of why, and the file or image you judged it from. Keep a count of items examined and items skipped, with the reason for each skip. If something breaks the plan (a missing file, an unexpected format, far more hits than expected), stop, log it, and ask me. STAGE 5. CHECK BEFORE YOU CLAIM a) Raw-data check: for every high-confidence candidate, go back to the rawest form of the data you have (the original records or points, not your rendered image or summary) and confirm the pattern is there. If it shows only in your processed version, move it to low. b) Known-records check: search every source in my "known records" line, plus published papers and reports on the same place or topic, by place name, by location and by feature type. Open the full text where you can, including figures, maps, tables and supplementary files, and note the page where you looked. If you could read only an abstract, mark that source ABSTRACT ONLY and treat the candidate as UNCONFIRMED, unless the abstract itself matches the candidate, in which case put it in ALREADY KNOWN and note ABSTRACT ONLY. For each candidate, list every source you searched and what you found. If a source was out of reach (paywall, no search, login), list it as missed and mark that candidate UNCONFIRMED. A missing entry in one register proves nothing about the papers. c) If my "known records" line says "not sure", list the record keepers and kinds of source you would expect for this topic, mark each [CHECK], and label nothing higher than UNCONFIRMED unless a source you opened matches it. d) Hand-check pack: pick five candidates and five items you rejected at random, or all of them if there are fewer, and list each with its exact file and position so I can open it myself, plus what to look for. STAGE 6. REPORT Write report.md and put each candidate in exactly one of three sections: - ALREADY KNOWN: matches a record or paper. A match to a record always goes here. Give the source, its link and where in it. - UNCONFIRMED: not found in the sources searched, or a source could not be searched. List the sources searched and the ones missed. - WEAK OR REJECTED ON CHECK: failed the raw-data check or matched a reject rule. Then add: whether the method was tested on the known example, and the result, or UNTESTED; the counts; the three assumptions from the log that would most change the result if wrong; the hand-check pack; the licence wording and what it allows; and one next step that fits what I said I will do with the result (for example who to contact and what to send them, with restricted details kept in their own file). Write no headline or press-style claim. SELF-CHECK BEFORE YOU HAND OVER - Every candidate sits in exactly one section, and nothing is called new or undiscovered. - Every source listed as searched is one you opened during this job, with its link; sources you could not open are listed as missed or [CHECK], never as searched. - Every number traces to a file in this folder. - You reported whether the criteria found my known example and changed nothing after a miss without my yes, or the report says UNTESTED. - Every source you could read only as an abstract is marked ABSTRACT ONLY, and its candidates sit in UNCONFIRMED or WEAK, unless the abstract itself matched. - progress.md lists every item done and skipped. - assumptions-log.md exists and the report names its top three. - If the data is sensitive, or I said "not sure", exact locations and personal details appear only in restricted-details.md. - End with one list of everything you were unsure of.
What the Maya run got right, and where it slipped
Each part of the brief comes from something in Miotti’s write-up.
- One written guide for every image. The agents turned about 1,200 km² of lidar strips into images, and 48 agents went through them looking for pyramids, platforms and plazas while filtering out hills, sinkholes and roads. He says no special model was involved, only a precise guide, which is why the brief fixes the criteria before anything runs.
- A check against the raw data. The agents went back to the raw point cloud to confirm the shapes came from ground points under the forest. The brief asks for the same step on every strong candidate.
- The miss. Asked to search the literature and archives, the agents reported one village that did not seem to be registered, a hilltop with a levelled plaza and an oval pyramid at least 8 to 12 m tall. The 2025 paper had it, so the brief makes the agent list every source it searched and caps any find at “unconfirmed”.
- Locations kept private. His report to INAH put the exact coordinates in a confidential annex, which is the reason the brief keeps sensitive details in their own file.
Run it
- Pick a question someone could check. A topic with a register, a records office or published papers gives the agent something to compare against, so its finds can be sorted into known and unconfirmed.
- Read the licence on the dataset’s own download page. Paste the wording into the set-up block, since the agent can only tell you what you may publish from words it can see.
- Start in plan mode. In Claude Code, run
claude --permission-mode plan, or pressShift+Tabuntil the status bar shows⏸ plan mode on. Claude then reads and proposes a plan and makes no edits until you approve. In other agents, the brief’s own “Approve this plan?” step holds it at the plan. - Approve a small first run. Set the run size to one slice of the data, and include one slice with a find you already know is there, so you see whether the method catches it and false hits show up before you spend time and tokens on the whole set.
- Read the assumptions log before the report. A wrong unit or coordinate system in
assumptions-log.mdundoes every find built on it. - Open the hand-check pack yourself. Look at the five candidates and the five rejects with your own eyes; if the rejects look like the hits, the criteria need another pass.
- Search the papers yourself, then send the unconfirmed list to the people who keep the records. Look up your best unconfirmed find by place name and feature type first. Miotti wrote to INAH, which could confirm only what was in its register, so say plainly that nothing is confirmed yet, and send
restricted-details.mdonly privately, to the record keeper.
The honest bit
- Not in the post: which tool ran the 48 agents, or the name of the NASA dataset. The set-up here is built from the steps the post describes, so it is one way to run this.
- Your check: the paper search in step 7 is the one the Maya run missed; the author says in the Hacker News thread the site was on page 14 of the 2025 paper.
The facts here come from Pietro Miotti’s post (last edited 11 November), his comment on Hacker News and the Claude Code docs, read 12 November 2026.
Do it now
Pick one public dataset you already know, fill in the set-up block, and run only the plan stage; reading the plan before anything runs shows whether your question can be checked at all.
Questions people ask
Did AI really discover a new Maya site?
No. Archaeologists had already mapped the site in Estrada-Belli et al. 2025, a paper the agents missed while searching the literature. Pietro Miotti, who ran the agents, updated his post to say they “rediscovered” it.
Is public data free to use?
Not always. Public does not mean free to reuse, so read the licence on the dataset’s own download page before you publish, share or sell anything built on it.
Which AI agent can analyse a dataset for me?
Claude Code, Codex or a similar coding agent, opened in an empty project folder. In Claude Code, start in plan mode with claude --permission-mode plan so it proposes a plan and makes no edits until you approve, and check what the agent can reach before you let it act.