GuideThread
Back to implementation guides

Implementation guide

Claude Code Agent Build System

Structure a Claude Code project, add explicit reliability and evaluation controls, then extend it with a practical authorized-data integration workflow.

3 ordered steps

  1. Step 1

    Project structure

    Foundation

    CLAUDE.md Build Playbook

    Created by Lynn O’Sullivan

    Use this to establish a Claude Code project with reusable CLAUDE.md instructions, conventions, modular guidance, and project context. Carry that structured project into the reliability work in Step 2.

    How to use it

    1. Map the project context, conventions, and reusable instructions.
    2. Add modular CLAUDE.md guidance at the appropriate project levels.
    3. Confirm the project structure before carrying it into Step 2.

    Open resource View original source post

  2. Step 2

    Agent reliability contract

    Reliability

    Sonnet 5 Agent Upgrade Kit

    Created by Balázs Turán

    Use this to apply objective, context, tool-use, state, recovery, acceptance, evaluation, escalation, and human-review rules to the Step 1 project. Treat model-specific behavior as changeable rather than guaranteed, then carry the approved contract into Step 3.

    How to use it

    1. Adapt the run specification to the Step 1 project and available tools.
    2. Define recovery, acceptance, escalation, and human-review checkpoints.
    3. Run the evaluation gate and carry the approved contract into Step 3.

    Open resource View original source post

  3. Step 3

    Authorized data integration

    Integration

    Site-to-API

    Created by Tom Crawshaw

    Use this as one optional extension of the Step 2 contract: build an authorized website-data workflow that returns repeatable JSON or CSV. Protect credentials, cookies, sessions, and secrets; authenticated Browserless use may require paid infrastructure.

    How to use it

    1. Choose an authorized site and define the JSON or CSV output contract.
    2. Configure Browserless without exposing credentials or session secrets.
    3. Run, verify, and document the workflow against the Step 2 eval gates.

    Open resource View original source post