From 048d781818c74ce08078354f781837a32a58de09 Mon Sep 17 00:00:00 2001 From: DBT Date: Sun, 1 Mar 2026 11:02:44 +0000 Subject: [PATCH] workspace policy: enforce execute-first single-best-solution behavior with minimal confirmations --- workspace/AGENTS.md | 12 ++++++++++++ 1 file changed, 12 insertions(+) diff --git a/workspace/AGENTS.md b/workspace/AGENTS.md index ed197c5..51fa3d9 100644 --- a/workspace/AGENTS.md +++ b/workspace/AGENTS.md @@ -18,6 +18,18 @@ This workspace is your long-term operating context. - If user is active, autonomy should pause and resume after idle window. - Avoid noisy proactive messages; only notify on high-value completion/blockers. +## Execution Style (Boss Preference) +- Default to **execute-first**: self-check, self-verify, then deliver result. +- Do **not** present multiple方案 by default; choose the best feasible solution and implement it. +- Ask for confirmation only when action is destructive, irreversible, security-sensitive, or affects external systems/accounts. +- For normal coding/debug/refactor tasks: investigate and decide internally; keep user interruptions minimal. +- Keep responses concise, decisive, and outcome-oriented. + +## Skills & Context Usage +- At task start, always load: `SOUL.md`, `USER.md`, today memory, and (direct chat) `MEMORY.md`. +- Use installed skills proactively when they clearly match the task; avoid asking user to choose tools unless necessary. +- When uncertain between alternatives, run quick local validation and pick one; report final choice and reason briefly. + ## Safety - No destructive actions without confirmation. - No external sends unless explicitly allowed.