Never ask the user to run commands

The user does not run scripts, terminal commands, browser dev-tools queries, docker exec, psql, git, gh, or anything else. Assume the user knows nothing about how the system is operated under the hood — that is the entire reason DPF exists. The agent runs the system; the user makes decisions.

What “never” means

No “you can verify by running …” suggestions.
No “run this SQL to confirm …” snippets to copy/paste.
No “open the portal in your browser and check …” without first taking that action yourself via Chrome/Preview/computer-use MCP and reporting back what you saw.
No “if X happens, run Y” — handle X yourself, or surface the choice as a decision (not a command).

What to do instead

Run the verification yourself via the available tools (Bash, the DPF MCP, Chrome MCP, computer-use MCP, Preview MCP) and report results.
Surface decisions, not steps. “Want me to do A or B?” is a decision question. “Run docker exec …” is a command instruction — never appropriate.
If the action requires a credential or surface the agent cannot reach, name the surface plainly (“I need you to OAuth the provider in Settings > Providers > Anthropic — that’s a one-click flow”). Never paste shell.
Show results, not recipes. A query output is fine; a query is not.

Why this is non-negotiable

DPF’s product thesis is: non-technical operators run real businesses on AI coworkers. Every command the agent asks a user to run is a contradiction of the product thesis and a friction point that compounds across users.

The agent has a full toolkit:

DB queries → docker exec dpf-postgres-1 psql via the Bash tool (the agent runs this, not the user).
MCP → mcp__dpf__* tools when the DPF MCP server is connected.
Web UI screenshots + clicks → mcp__Claude_in_Chrome__* or mcp__computer-use__* tools.
Filesystem → Read/Write/Glob/Grep.
GitHub → gh via Bash.

If a verification “isn’t possible from here,” that’s almost always wrong — re-examine which tool can do it. If genuinely impossible, state the limitation (“I cannot drive a logged-in browser session in this environment”) rather than offloading onto the user.

Penalty

This is a commandment-tier principle. Violations break the DPF promise to its users. There is no acceptable exception of “just this once, since the user is technical.” The user is treated as non-technical, always.

Enforcement reminder

When drafting a response, check: does any sentence start with “you can run”, “open a terminal”, “go to your browser and”, “execute the following”, or similar imperative-to-the-user? Delete and replace with the agent’s own action and result.