devflow engineer persona

The problem

Standard AI assistants often suffer from "helpfulness drift," offering bloated refactors when you just need a bugfix or apologizing for code they haven't even read yet. They prioritize agreeableness over technical rigor, leading to scope creep and untested regressions.

What it does

• Enforces a "read-before-write" policy to ensure all code edits are based on current session inspections.
• Mandates a test-first approach, requiring a failing test case before any bugfix is implemented.
• Strictly decouples refactoring from functional changes to prevent hidden regressions and PR bloat.
• Adopts a staff-level engineering persona that leads with technical verdicts rather than conversational fluff.
• Maintains strict security boundaries, refusing to handle secrets or execute destructive commands without explicit confirmation.

Why this beats prompting it yourself

Writing "act like an engineer" in a system prompt is not enough to stop LLMs from drifting into generic assistant behavior during long sessions or tool loops. This skill uses specific anti-drift mechanisms, including forbidden phrase lists and periodic identity reinjection, to ensure the agent stays disciplined and terse even 30 turns deep into a complex debugging session.

Use cases

• Debug production issues by identifying the smallest possible fix that satisfies a failing test.
• Conduct disciplined code reviews that prioritize severity and technical reasoning over politeness.
• Navigate complex tool-calling sequences without the agent losing its professional context or engineering voice.
• Execute safe, incremental changes in sensitive codebases where destructive actions must be gated.