Reasoning budget

• Match reasoning to step difficulty and bias toward acting; for mechanical / IO / format steps, if a provided script already covers the task, run it — don't re-derive it.
• Save deep, extended reasoning for the one genuinely hard judgment a script can't make for you.

Working style

• Understand first, then decide. Read this skill's SKILL.md and understand the scripts it ships before you act. Then plan accordingly, such as run a script as-is when it fits, change a script when it's close, or write extra scripts to complement — based on what the scripts actually do, not a guess.
• Plan the whole path up front, then chain. Outline the full sequence of steps before running anything, batch independent steps into one turn, and pipeline the whole plan in as few turns as possible. Don't do things that can be pipelined into one call turn-by-turn.
• Inspect an input ONCE. To learn a file's structure (sheets/columns, pages, form fields, keys), dump it once — ideally to a file you then grep — never re-open the same file field-by-field or retry it with several libraries.
• Don't repeat work. Do not rerun a command when its inputs and relevant state are unchanged, and do not reread an unchanged file, script, or SKILL.md already in context. After a tool or command may modify a file, reread the affected content before relying on it.
• Write code once and reuse. If a step needs code, write it once as a small script (paths/params as CLI args) and call it; don't paste near-duplicate inline python across turns. Keep it terse — no comment banners or narration in inline scripts.
• Keep outputs small. Don't put large tool results and outputs into the context, instead write them into a file and use tools to inspect them. If there is no tool available, you should write your own scripts to inspect the file.
• Don't do anything unnecessary. Don't call tools, read files, or put results into context unless they're immediately needed.

UiPath Maestro BPMN