Scope preservation — verbatim, no add, no drop

rule Fri Jul 03 2026 00:00:00 GMT+0000 (Coordinated Universal Time) rulesagentpromptsexecution

Scope preservation — verbatim

Rule

When user gives a multi-requirement task:

• Do not drop requirements even if they seem "covered by" another.
• Do not add requirements the user didn't request.
• Do not paraphrase requirements in ways that lose specificity.
• Preserve exact wording in the output's task list where possible.

Why

• Silent drops = silent failures. The 8th requirement in a 12-item spec is exactly where an agent's summarization instinct nukes work.
• "Helpful" additions bloat scope. Every extra requirement is code to write, review, and maintain.
• Paraphrase drift. "Add error handling for network failures" → "add error handling" loses the "network" specificity, which changes the fix.

Application

• Extract requirements verbatim into the TASK-x.y checklist per [[task-oriented-execution-model]].
• If two requirements look like duplicates, mark both — they may differ in ways you don't see yet.
• If a requirement is unclear, ASK (via MCQ) rather than paraphrase-and-assume.
• If you complete extra work "along the way", flag it explicitly: "Also did X, not in original scope — remove if unwanted."

Anti-patterns

• ❌ "I combined requirements 3 and 7 since they overlap"
• ❌ "I added item 13 because it seemed necessary"
• ❌ "Requirement 5 said 'X in production' but I did X everywhere for consistency"
• ❌ Silent paraphrase — user can't diff their prompt against your interpretation

Exceptions

• Ambiguous or contradictory requirements — surface as MCQ per [[grill-me-default]], don't silently pick.
• Missing prerequisite — flag it ("blocked by "), don't guess.
• User invites expansion ("also cover…") — then add explicitly, still verbatim.

Cross-refs

• [[task-oriented-execution-model]] — the checklist mechanic that enforces this
• [[karpathy-guidelines]] — surface uncertainty; don't silently pick
• [[terse-issues-less-hallucination]] — related: cut noise, not signal
• Upstream pattern: prompts.chat "Agent Role" shared header — "Preserve scope exactly as written; do not drop or add requirements"