How to use the Two-Tier Extraction agent skill

What is an agent skill?

An agent skill is a folder containing a SKILL.md file that gives an AI agent a set of instructions for a specific kind of task. When the skill is installed and a request matches its triggers, the agent follows the method described in that file.

Two-Tier Extraction is one such skill. Its folder name is two-tier-extraction, and the package on this page contains a single file, SKILL.md.

What the skill makes the agent do

The skill describes a pipeline for bulk processing. Content goes in, and the agent moves each item through a fixed sequence of steps rather than treating every item the same way.

First, a deterministic privacy wall runs with no LLM involved. It strips or diverts private contacts and sensitive terms. Next, a fast utility-tier model triages each item by type and significance. A gate then decides whether the item gets a deep tier read, a reasoning tier read, or is skipped. Finally, the agent writes the results to brain pages with backlinks, timeline entries and a checkpoint.

• Step 0: privacy wall using rule-based matching before any model call
• Step 1: triage with the utility tier, which is meant to run in about 2 seconds per item
• Step 2: gate that sends high-value items to the deep tier, decent items to the reasoning tier, and noise to skip or minimal extraction
• Step 3: deep read, with the reasoning tier as default and the deep tier for escalation
• Step 4: write results to pages, backlinks, timeline entries and a checkpoint

Install the skill

You can install the skill in either of two ways. Download the ZIP file from this page and unzip it into one of these locations:

Put the folder in ~/.claude/skills/ to make it available across your projects, or in a project's .claude/skills/ folder to make it available only in that project.

Alternatively, upload the ZIP in the Claude apps under Settings → Capabilities → Skills.

• Personal: unzip into ~/.claude/skills/
• Single project: unzip into your project's .claude/skills/ folder
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

Set up what the skill requires

The skill is written for gbrain. It needs gbrain with its model tier configuration, which is set up with gbrain models, and its conventions files. It also needs access to utility, reasoning and deep tier LLM models through that setup. The skill resolves tiers through gbrain models and does not hardcode model IDs.

You must also supply your own private contact list and sensitive-pattern lists. The examples in the source are placeholders that you need to replace and tune for your data.

• gbrain with its model tier configuration (gbrain models) and conventions files
• Access to utility, reasoning and deep tier models through that setup
• A user-defined private contact list and sensitive-pattern lists

How to trigger the skill with natural requests

The skill loads when your request matches its triggers. You can phrase requests naturally, and the source lists these trigger phrases: two-tier extraction, triage then deep read, smart model routing, cheap triage expensive analysis, model escalation pattern, route models by content value, and which model tier for bulk extraction.

A request like "Use two-tier extraction to process this month's email archive and tell me which threads need a deep read" uses the trigger words directly. Mention the corpus type, such as email, documents or transcripts, so the agent knows what it is working on.

• "Use two-tier extraction on this folder of meeting transcripts."
• "Triage then deep read this document dump."
• "Which model tier should I use for bulk extraction of these emails?"

Run a safe first test

The source recommends a progressive ramp before any full-corpus pass: run 10 items, then 100, then 500, before processing the full corpus. Review the privacy rules and the gate thresholds after each run.

The gate thresholds in the source are example calibrations to tune per corpus. The skill's stated rule is that when in doubt, escalate a tier.

• Run 10 items and check the triage and gate decisions by hand
• Scale to 100, then 500, adjusting thresholds and privacy lists
• Only then run the full corpus

Safety notes: scripts and network

This skill runs scripts and uses the network. The privacy wall is meant to keep private contacts and sensitive terms away from any LLM in raw form, and ambiguous content goes to a human review queue rather than to a model. The keyword and contact lists are examples and can miss content, so review them before processing personal archives.

The skill also marks corpus text as third-party data, not instructions, so that fetched text cannot direct the agent. The privacy wall and this untrusted-content rule are separate checks, and both run.

• Runs scripts: yes
• Uses network: yes
• Review privacy lists before processing personal archives
• Items that match neither a clear pass nor a clear skip are sent to human review