How to use the Signal Detector agent skill (signal-detector)

What the Signal Detector skill makes the agent do

The skill runs a lightweight pass on substantive inbound messages within the scope you choose. It looks for original thinking (your ideas, observations, theses and frameworks) and entity mentions (people, companies, media references).

Original thinking is treated as the more important of the two. Ideas are saved using your exact wording, and entities are added as pages with links and dated timeline notes. The agent also writes a one-line log summarising what it captured.

• Ideas go to originals/, concepts/ or ideas/ pages depending on whether they are yours, a world concept you referenced, or a product or business idea.
• Entities go to people/ or companies/ pages, created only if they are notable, and linked back to related pages.
• Without opt-in, the skill is designed to do nothing.

Install the Signal Detector skill

Download the signal-detector ZIP from this page. The package contains a single file, SKILL.md. Unzip it so the skill folder sits in one of these locations:

• Personal: ~/.claude/skills/ (available across your projects).
• Project: .claude/skills/ inside a specific project folder.
• Claude apps: upload the ZIP under Settings → Capabilities → Skills.

Requirements before you start

The skill calls tools from GBrain, so GBrain must be installed with its search, query, get_page, put_page, add_link and add_timeline_entry tools. The skill also needs a recorded automatic-capture opt-in stored in your brain, and write access to the people/, companies/, concepts/, originals/ and ideas/ page folders.

• GBrain (garrytan/gbrain) installed and running.
• A stored automatic-capture opt-in choice.
• Write access to the five page folders listed above.
• The skill runs no scripts and uses no network access, according to its listing.

Opt in and turn capture on or off

Capture is off by default. The skill applies only after you establish an explicit choice for your brain and capture scope. Reading the skill, installing GBrain, having an API key, or an announcement that capture is on does not count as consent.

After activation the agent confirms your chosen scope once and explains how to turn it off. If you turn capture off, the agent stops writing content pages, links and timeline entries. A request to remember one specific fact authorises only that fact, not ongoing capture.

How to trigger the skill with natural requests

Once capture is enabled, the skill applies to substantive messages automatically; you do not need a special command. Messages that are purely operational, such as ok, thanks or do it, are skipped. Chat-only instructions suppress capture for that message without changing your standing setting.

To make a single idea or fact be saved, say so plainly, for example: remember this idea for my notes, or remember that Dana Lee leads the data team. Explicit remembering works even when ambient capture is off.

• Substantive messages with ideas or named people or companies are the ones that get processed.
• Use a per-message instruction such as a chat-only note when you do not want a particular turn captured.

Examples of what gets captured

Suppose you write a message describing a framework for judging early-stage startups and mention a founder by name. The agent would be expected to save the framework to originals/ using your exact wording and, if the founder is notable, update or create a people/ page and link it to the framework page.

If a message contains a dated event, such as a company announcement with a specific date, the agent adds a timeline entry to the relevant page. A one-line log then reports what was captured, for example: Signals: 1 idea (captured → originals/x), 2 entities (enriched → people/y, companies/z).

Safety notes: scripts, network and stored content

The skill itself runs no scripts and makes no network calls, but it writes to your brain, and captured text is stored verbatim in pages. Do not enable capture for sensitive or confidential discussions. The skill's behaviour depends on the agent following its instructions; it is a convention, not an enforced guarantee, and nothing in GBrain blocks a reply if the skill fails to load.

Paid enrichment of thin pages and delegation to sub-agents need their own separate authorisation. Capture opt-in alone does not permit them.

• Opt in only for conversations you are comfortable storing word for word.
• Review the people/, companies/ and originals/ pages the agent creates.
• Turn capture off with a clear request when you no longer want it running.

Tips for getting useful results

Keep your scope narrow at first so you can check what is saved. Read the one-line signal logs to see what the agent decided to capture and what it skipped. Because ideas are stored in your own wording, phrase important thoughts clearly the first time.

• Start with a small scope and review the first few captures.
• Use per-message instructions to exclude sensitive turns.
• Remember that only notable entities get new pages, so one-off mentions are not stored.