AI CODING AGENT GUIDE.
SETUP. SCOPE. REVIEW.
A short set of protocols for keeping an AI-assisted coding session focused, inspectable and under your control.
[00]ORIENTATION
Open a working folder and identify the task. Start with inspection, not a broad instruction to change everything. Confirm the build and available tools before relying on a capability.
[01]SCOPE OF WORK
State the desired outcome, which paths may change and what must remain untouched. Mention existing work that belongs to someone else. Ask for a plan before consequential operations.
[02]MODEL CONNECTION
The reference configuration supports OpenAI-compatible and Anthropic API connections. Enter the required endpoint details and credentials, choose an available model and test a small request. Do not paste secret credentials into example prompts.
[03]TOOL ACCESS
Connector settings determine which external tools are available. Review their command, dependencies, credentials and permissions. Enable only the connections required for the task. An MCP connection is not a blanket guarantee about a tool.
[04]REMOTE TARGET
Verify the host, user, directory and environment. Begin with read-only status checks. Explicitly prohibit restarts or edits when investigating an issue. Establish a backup and rollback procedure before a production change.
[05]REVIEW & VERIFICATION
Inspect the actual diff and the output of relevant checks. Separate verified results from suggestions and assumptions. If a check could not run, treat that as an open item rather than a pass.
[06]DATA BOUNDARY
A desktop process may send context to a remote model or external connector. Retention and logging depend on those services. Local inference requires your own suitable server; an all-local workflow also needs local tools.
[07]RELEASE STATUS
This site presents the direction for a separate product. The screenshots are temporary shared references. The Windows installer is published independently; other platform builds and final specifications are not announced here.
Objective: Relevant files or environment: What may change: What must not change: How to verify the result: Inspect first. Report uncertainties. Keep unrelated changes out of scope.
An explicit order produces a clearer review
Before an agent session, write down the objective, relevant paths, permitted changes and verification criteria. This is especially useful when a task touches a shared repository or a remote environment, where plausible actions can still be outside the request.
After the task, distinguish the produced work from the evidence supporting it. An unrun test remains unrun even when the proposed fix looks sensible. Keep open issues visible so the next person can evaluate the result without reconstructing the whole conversation.
