Make your AI agent read what was published after its training

By Mackenzie Bowes · mackenzie@mackenziebowes.com · @_mackenziebowes on X

do-research is a Claude Code skill that sends an agent to the published record before it solves anything: notes first, then papers, practitioners and code, with exact calls to 14 sources that answer a plain curl with no key. Why we need it, and a one-line install.

Contents

Ask an agent a hard question and it answers from its training. That stopped months ago. People kept publishing. The agent doesn't know what it hasn't read, so it doesn't look.

Here is what looking found for one article this week. We wanted to know whether AI models cheat on the checks you give them. A 2025 study gave models tasks that were impossible on purpose. In one kind of case, GPT-5 gamed the checks 76% of the time rather than admit it was stuck. Given an explicit way out ("this can't be done, get a human"), the cheating dropped from 54% to 9% [1]. That paper came out after most models' training stopped, so an agent working from memory, however clever, would have had nothing to go on but a guess. It wouldn't even know it was guessing.

So we research first. Always.

A shelf of books is split by a post. The books left of it are bright and labelled trained on; those right of it are dark and labelled published since. An agent carrying a lamp walks past the post, and each book it passes lights green. The tallest lifts, labelled: in one kind of case, GPT-5 gamed the checks 76%.

An agent answers from the shelf it was trained on. What was published since sits past a line it will not look beyond unless it is sent.

Why it's a skill and not a habit

Most problems have been studied before. Some solutions worked and some failed, for different reasons. Find that work first and you get two things: you copy what worked and improve on it, and you see what failed before you repeat it.

An agent left alone does one web search, reads the first vendor page and calls it research. I asked for more than that:

I want do-research to be better, still. I want it to outline specific methods and tool-calls that agents can use to get information.

Mackenzie, 2026-10-16 15:09

So the skill names each source with the exact call.

Two lanes. In the first, a marker moves from search to the first vendor page and stops, red. In the second it passes your notes, papers, practitioners, code and an answer page, lighting each green.

Left alone an agent stops at the first vendor page; the skill walks the whole record.

What's in it

Your own notes first. Another agent has probably searched this before. We use qmd to search every note on the machine in about a tenth of a second.
Web search, with a warning. Search results are full of marketing that looks like research. Use search to find dense collections, then read those pages directly.
Reading pages: lightpanda (fast, runs the page's JavaScript, returns Markdown), then a full headless browser for pages that need clicks.
Papers: OpenAlex for search and for newer work that cites a paper, arXiv for the newest preprints, then Semantic Scholar, Crossref, Unpaywall, PubMed, Europe PMC, bioRxiv, DOAJ, OpenReview, Hugging Face papers, Zenodo and Wikipedia. We tested each one on 2026-10-17: all answered a plain curl with no key. The skill also records which didn't (SSRN blocked us, CORE needs a key).
Practitioners: Hacker News through Algolia's free API, and GitHub code search for what people built.
Wrapping up: save every source you relied on, write an answer page (what worked, what failed, the principles, what's still unknown), and record the dead ends so the next agent skips them.

The list is open, and your field probably has its own archive that belongs on it, but the list was never the point. The skill tells the agent what it's looking for: the shape of success and failure, and the principles that made those shapes. Then it can split the task into steps.

do-research tests are solved. do-research works amazing.

Mackenzie, 2026-10-16 23:50

Sixteen tiles light one by one. Fourteen turn green: sources that answered a plain curl with no key. Two turn red: SSRN blocked us, CORE needs a key.

The skill records which sources answered a plain curl with no key, and which did not.

Install it

A page labelled SKILL.md on the site lifts, arcs over, and drops into a folder labelled ~/.claude/skills/do-research. It turns green, and below it reads: new session: /do-research.

The skill is one file, copied into the skills folder.

For Claude Code:

mkdir -p ~/.claude/skills/do-research
curl -fsSL https://www.mackenziebowes.com/skills/do-research/SKILL.md -o ~/.claude/skills/do-research/SKILL.md

Start a new session and say "do research", or type /do-research. Read the file first: /skills/do-research/SKILL.md. It works best beside mbsi-mechanism, which says what the research is for, and dig-apc, for when the work fails anyway.

If your field has an archive that answers a plain request and isn't on the list, tell me and I'll test it and add it. mackenzie@mackenziebowes.com or @_mackenziebowes.

Sources

ImpossibleBench: tasks that are impossible on purpose, and how often models game the tests (arXiv 2510.20270). arxiv.org/abs/2510.20270