Concise Deep Research

An agent skill that produces short, evidence-led research briefs. Every standard run searches through at least three distinct search providers, reads the sources it cites, labels each claim as supported, partly supported, disputed or unverified, and delivers a brief of about 1,200 to 1,800 words with a separate evidence appendix. Cheap models and a System One classifier do the searching and sorting; the strongest model plans, challenges and writes.

It follows the open SKILL.md format, so one folder works in Claude Code, Codex CLI, OpenCode and any other host that reads agent skills.

Install

You install the skill once. After that, Claude uses it when you ask for research. Setup takes about five minutes.

First, which Claude are you using?

Claude comes in three forms. The steps are different for each, so find yours first.

You use Claude...	That is	Go to
In a web browser at claude.ai, or in the Claude app, and you type messages in a chat	Claude chat	Claude chat
In the Claude desktop app, in the Cowork tab, where Claude works on files in a folder on your computer	Cowork	Cowork
In a terminal window, where you type claude to start it, or in the Code tab of the desktop app	Claude Code	Claude Code

Not sure? If you only chat with Claude, you use Claude chat. Start there.

You need a paid Claude plan (Pro, Max, Team or Enterprise). The free plan cannot use skills.

What is a skill, and what is a plugin?

A skill is a folder of instructions. It teaches Claude how to do one job. This skill teaches Claude how to do careful research.

A plugin is a package that holds one or more skills. It makes the skill easy to install and update. Cowork and Claude Code install plugins. Claude chat uses the skill file directly.

You do not need to understand the files inside. Follow the steps for your app.

Claude chat

Use this if you chat with Claude at claude.ai, or in the Claude desktop or mobile app.

1. Download the skill file. Click this link: concise-deep-research.zip. Your browser saves a zip file, usually to your Downloads folder. Do not open or unzip it.
2. Open Claude at claude.ai on a computer. You cannot upload skills from the mobile app.
3. Turn on code execution. Go to Settings, then Capabilities. Turn on Code execution and file creation. Skills need this setting.
4. Open the skills page. In the left sidebar, click Customize, then Skills.
5. Upload the file. Click the + button, then Upload a skill. Choose concise-deep-research.zip from your Downloads folder.
6. Check that it is on. The skill now shows in your list. Make sure its switch is on.

The skill now works in every new chat, on the web, the desktop app and the mobile app.

Do not use the green Code > Download ZIP button on this GitHub page. That zip holds the whole project, and Claude will reject it. Use the link in step 1.

Cowork

Use this if you use the Cowork tab in the Claude desktop app.

1. Open the Claude desktop app and click Customize in the left sidebar.
2. Click Plugins.
3. Click Add marketplace. A marketplace is a list of plugins that someone shares from GitHub.
4. Type the address futureformed/concise-deep-research and confirm.
5. Find Concise Deep Research in the list and click Install.
6. Start a new Cowork task. Cowork loads new plugins when a task starts, not during one.

Plugins you add here are saved to your Claude account. They also show in Claude Code on any computer where you sign in.

If you cannot add a marketplace (some work accounts turn it off), download concise-deep-research.zip and use the steps for Claude chat instead.

Claude Code

Use this if you start Claude by typing claude in a terminal.

1. Open a terminal. On a Mac, open the Terminal app. On Windows, open PowerShell.

2. Copy and run this command. It tells Claude Code where to find the plugin.

   claude plugin marketplace add futureformed/concise-deep-research

3. Copy and run this command. It installs the plugin.

   claude plugin install concise-deep-research@concise-deep-research

4. Start Claude Code again. Type claude and press Return. If Claude Code was already open, type /reload-plugins instead.

You can also do steps 2 and 3 inside Claude Code. Type /plugin, press Return, and follow the menu.

In the Code tab of the desktop app: click the + button next to the message box, then Plugins, then Add plugin. Add the marketplace futureformed/concise-deep-research and install Concise Deep Research.

Other agents (Codex, OpenCode, Cursor and others)

Run this in a terminal. Add -g at the end to install it for all your projects.

npx skills add futureformed/concise-deep-research --skill concise-deep-research

Or copy the folder skills/concise-deep-research into your agent's skills folder, for example ~/.claude/skills/ or ~/.agents/skills/. Keep the references, agents and scripts folders inside it.

Check that it works

Start a new chat or session and type:

What skills do you have for research?

Claude should name concise-deep-research. If it does, the install worked. Next, connect three search providers. The skill cannot do research without them.

Update or remove

App	To update	To remove
Claude chat	Download the zip again. In Customize > Skills, delete the old skill and upload the new one.	In Customize > Skills, open the skill's menu and delete it.
Cowork	In Customize > Plugins, open the plugin and click Update if it shows.	In Customize > Plugins, open the plugin and click Uninstall.
Claude Code	claude plugin update concise-deep-research@concise-deep-research	claude plugin uninstall concise-deep-research@concise-deep-research

If something goes wrong

• "Invalid skill" or "SKILL.md not found" in Claude chat. You uploaded the wrong zip. Download it from the link in step 1 of Claude chat, not from the green Code button.
• No Skills page in Customize. Your plan may not include skills, or your organisation has turned them off. On a Team or Enterprise plan, ask your admin.
• claude: command not found. Claude Code is not installed on this computer. Install it from claude.com/claude-code, or use the Claude chat steps.
• Claude does not use the skill. Start a new chat or session. Then name it in your request: "Use concise-deep-research to research ...".
• The skill says it has fewer than three search providers. The install worked. Connect more search services, as the next section shows.

Menu names in the Claude apps change from time to time. If a button has a different name, look for the nearest match. These steps were checked on 6 October 2026.

Connect search providers

The skill needs three search services that return real results. Four of these work with no key for light use, so you can try the skill before you spend anything. The last column says what each provider asks for before your agent can use it. The section after the table says where a key goes.

Provider	Get a key	Free allowance	Page text?	What you need
Exa	dashboard.exa.ai	$10 credit a month; keyless MCP	Yes	Nothing to start; a key for more use
Tavily	app.tavily.com	1,000 credits a month	Yes	A browser sign-in the first time
Brave Search	api-dashboard.search.brave.com	$5 credit a month	Snippets	A key
Parallel	platform.parallel.ai	5,000 requests a month; keyless MCP	Excerpts	Nothing
Firecrawl	firecrawl.dev	1,000 credits a month; keyless with limits	Yes	A browser sign-in the first time
Linkup	app.linkup.so	4,000 queries	Yes	A key
Perplexity	console.perplexity.ai	None documented	Snippets	A key
You.com	you.com/platform	100 queries a day keyless	Yes	Nothing to start; a key for more use

Prices, one-line connect commands for Claude Code and Codex, and notes on which services count as distinct providers are in search-providers.md.

The easy start: three free providers in Claude chat or Cowork

A search provider joins Claude as a connector. A connector is a link that lets Claude use another service. These three need no key and no payment:

Provider	Connector address
Parallel	https://search.parallel.ai/mcp
Tavily	https://mcp.tavily.com/mcp/
Firecrawl	https://mcp.firecrawl.dev/v2/mcp-oauth

Add each one like this:

1. In Claude, click Customize in the left sidebar, then Connectors.
2. Click +, then Add custom connector.
3. Type the provider's name, for example Parallel, and paste its address from the table.
4. Click Add. If a sign-in page opens (Tavily and Firecrawl do this), make a free account and sign in.
5. Do the same for the next provider.

Connectors you add are saved to your Claude account, so they work in Claude chat and Cowork. Start a new chat after you add them. On a Team or Enterprise plan, an admin may need to add custom connectors for you.

In Claude Code, use the one-line commands in search-providers.md instead.

Where your API keys go

An API key is a long string of letters and numbers that a provider gives you when you sign up. It works like a password for a program. Each time your agent searches through a provider, it sends the key with the request. The provider checks the key, runs the search and counts the cost against your account. Anyone who has your key can spend your credit, so treat it like a password.

You never type the key into a chat, and the skill never asks you for it. The key goes into one place in your agent host, once, before your first run. After that, the host sends it with every search and you do not touch it again.

When. Do this once per provider, before you run the skill for the first time. The skill checks for working providers at the start of every run and tells you if one is missing.

How.

1. Sign up with the provider and create a key in its console. The links are in the table above.
2. Copy the key. Most consoles show it once. If you lose it, create a new one.
3. Put it in the one place for your host. The table below says where.
4. Start a new session, so the host loads the new connection.
5. Ask your agent to run one search with that provider. A real result means the key works. Installed is not connected.
6. Save the key in your password manager and clear it from your clipboard.

Where.

Host	Where the key lives	What you do
Claude Code	Claude Code's own config file, ~/.claude.json, outside your project	Run the provider's claude mcp add command from search-providers.md. The key is part of the command, as a --header or --env option. Claude Code stores it for you.
Codex CLI	~/.codex/config.toml, under [mcp_servers.<name>]	Run the provider's codex mcp add command, or add the block by hand with the key in its env table.
Claude Cowork, claude.ai, ChatGPT	The provider's connector settings in the app	Add the provider as a connector. Paste the key into the field the connector asks for, or sign in when it opens a browser page.
Any host, for the classifier	A shell environment variable, TYPESAFE_API_KEY	Add export TYPESAFE_API_KEY="..." to ~/.zshenv on macOS or ~/.bashrc on Linux, then open a new terminal. Details in model-routing.md.

Three things to know:

• The connect commands write $LINKUP_API_KEY and similar. That is a placeholder for a shell variable. Either replace it with your key, between the quotes, or set the variable in your shell first.
• Providers marked "browser sign-in" (Tavily, Firecrawl) open a web page the first time your agent uses them. You log in there and the host keeps a token. You never see or store a key.
• In Claude Code, claude mcp add --scope project writes a .mcp.json file inside your project, which git will commit. Do not put a real key there. If a team needs a shared .mcp.json, write ${LINKUP_API_KEY} in the file and set the variable in each person's shell.

Where a key must never go: a chat message, a prompt, SKILL.md, the run folder, or any file in a git repository.

Use

First decide where the results should go: a project folder, a synced Google Drive or OneDrive folder, or wherever you keep research. Name it in the request. If you do not, the skill saves to a research/<date>-<slug>/ folder in the current working directory and tells you so before it starts.

Ask your agent:

Use concise-deep-research to research [question]. The audience is [audience], the decision is [decision], and the scope is [scope]. Save the results in [folder].

The skill discovers the search tools your host exposes, runs the first query as a readiness test, and asks you to connect more providers only if fewer than three work. It also tells you, before spending anything, if a classifier key or a search provider is missing and how to add it.

What you get

Two files, in a run folder:

• brief.md: the answer, about 1,200 to 1,800 words. A 2–3 sentence conclusion, up to five cited findings, up to three implications, the uncertainties that could change the decision, and one coverage line naming the providers and date. Each finding carries the figure, its denominator, the period, the method and the main caveat, so you can judge it without opening the appendix. This is the file to read.
• evidence.md: the receipts. Open it only to check a claim: each source ID in the brief leads to the URL, the passage location, the reader's status and the challenger's verdict. It also holds the provider log, the remaining gaps and a model and cost log.

The run also writes a work/ subfolder: search registers, classifier input and output, full claim ledgers, the draft findings and the challenge record, plus run notes. It is the audit trail and the checkpoint a resumed run picks up from. It is not meant to be read, and you can delete it once you accept the brief.

How long it takes

A standard run takes roughly 15 to 40 minutes of wall-clock time, longer on a slow host. The skill searches three providers, reads the sources it cites rather than trusting snippets, labels every claim, and then runs a separate adversarial pass that tries to overturn the findings before anything is written. Most of that time is reading and cross-checking, not searching. If you need an answer in two minutes, this is the wrong tool; ask your agent a plain question instead.

What a System One model is

A language model writes text. A System One model decides. You give it a passage and a question with a fixed set of allowed answers, and it returns the answer and a probability. It does not write prose or reason in steps, so it answers in about a tenth of a second at a few cents per million tokens, roughly a thousandth of a language-model call.

This skill uses one for the clerical half of research: ranking results, labelling source types, spotting sources that share one origin, pre-checking citations, flagging paywalls and injected instructions, and deciding which claims need the expensive challenge pass. Reading, challenging and writing stay with language models.

TypeSafe's Jev is one option, and the bundled script targets its API. Cloudflare's Clef, launched in October 2026, is another, with open weights. Rerank and classify endpoints from Voyage, Cohere and Jina cover parts of the job, and a small language model with a JSON schema is the fallback when nothing else is configured. Set TYPESAFE_API_KEY to use Jev; without it the skill runs the same steps on the cheapest language model and says so in the appendix. The options and their prices are in model-routing.md.

Contents

Path	Purpose
skills/concise-deep-research/SKILL.md	The skill: six stages from scope to delivery
skills/concise-deep-research/references/provider-readiness.md	How providers are discovered, counted and activated
skills/concise-deep-research/references/search-providers.md	Eight search providers: keys, prices, connect commands
skills/concise-deep-research/references/deliverables.md	Templates for the brief, the appendix and run notes
skills/concise-deep-research/references/model-routing.md	Which model tier and which classifier runs each stage, with dated prices
skills/concise-deep-research/references/delegation-templates.md	Sub-agent contracts and agent files for Claude Code and Codex
skills/concise-deep-research/scripts/typesafe_sort.py	Runs the classifier steps over JSONL rows, standard library only
scripts/build-release-zip.sh	Builds the zip for Claude chat to attach to each GitHub release
METHODS.md	Why the method is shaped this way
APPROACH.md	The agreed research method in prose
EVALUATION.md	Behavioural acceptance scenarios
assets/	The pipeline diagram and a square version for social posts

Limits

The skill is instructions. It does not install providers, hold keys or buy subscriptions, and it orchestrates whatever search tools the host exposes. Package validation checks structure and links; research quality needs a live run. Prices and model names in the references are dated and will go stale.

Licence

MIT. See LICENSE.