Concise Deep Research: AI research with the receipts attached

If you've ever run AI "deep research" before a pitch, a meeting or an interview, you'll know the shape of what comes back: 3,000 words you won't read, with citations that people rarely check.

Run badly, it's pure slop. It might take 15 minutes to produce (good for you), but it takes the person on the receiving end far longer to read and understand (not good for them). And most of the time it runs the most expensive AI model for the whole job, including the part where it sorts search results into piles.

So I built a research skill for Claude Code and Codex called Concise Deep Research. It can also be used in Claude Cowork, ChatGPT and any other AI tool that reads the open agent-skill format. It's open source, MIT licensed, and the whole thing is on GitHub at github.com/futureformed/concise-deep-research.

The five steps, and which tier of model runs each one

What it does

It searches through at least three separate search providers (Exa, Tavily and Brave, for example), reads the sources it cites, and labels every claim as supported, partly supported, disputed or unverified. If three articles are quoting the same press release, it counts them as one source, not three. If a page is paywalled, it says so rather than pretending it read it.

You get a brief of 1,200 to 1,800 words and a separate evidence file showing what was read, what wasn't, and where the evidence is thin. Each finding carries enough of its evidence that you can judge it without opening the evidence file. Half the length of a typical deep research report, with the receipts attached. The brief opens with the answer, then up to five findings, up to three things to do about it, and only the uncertainties that could change your decision.

Before it hands anything over, a second pass tries to knock the findings down. It looks for counter-evidence, checks whether two sources disagree because they measured different things, and marks anything it can't settle. That pass is the part most research tools skip, and it's the part that stops you walking into a meeting with a confident wrong answer.

Where the money goes

The cost part is pleasing. The expensive model does three jobs: scope the question, try to knock down the findings, and write the brief. Those are the jobs where judgement earns its price.

Cheap models run the searches. They come back with a table of what they found, never the pages themselves, so the expensive model isn't re-reading thirty web pages every time it thinks.

The sorting goes somewhere cheaper still. Is this result relevant? Is this a press release or a paper? Does this passage actually support the claim? Those questions go to a System One model. These are classifier models like TypeSafe's Jev or Cloudflare's new Clef. They don't write anything. You give them a passage and a question with a fixed set of answers, and they pick one and tell you how sure they are, at roughly a thousandth of the price of a chat model. Four cents per million tokens. For a whole research run, that's the cost of a text message.

The rule I've landed on: pay for judgement once. Everything that isn't judgement runs cheap or free.

Try it for nothing

Four of the search providers it supports work without an API key for light use, so you can try it before you spend a penny. The repo lists eight providers with sign-up links, free allowances and the one-line commands to connect each one to Claude Code or Codex. Installing the skill itself is two commands in Claude Code, or one with the skills installer for anything else.

A couple of things to be straight about. It needs a real run on a real question to judge research quality, and the prices in the docs will drift (fair warning). The method is all written up in the repo, including why each rule is there, so you can argue with it.

If you're not totally comfortable with GitHub repos, API keys and the like, drop me a note. Knowing how to use them will boost your AI capability massively, and it's less work than it looks.

What I'd like back

If you run it on something real, tell me what it got wrong. That's the feedback that makes it better. And if you want help getting research like this working inside your business, get in touch for a chat.