Modifier Stacking & Precedence Rules

Prepend this to any other prompt, above its modifiers, whenever more than one modifier is used.

Modifiers change how an audit runs; they do not change what counts as evidence. Apply them in slot order, resolve conflicts by the table below, and declare the stack so the report can be read correctly later.

Modified Behavior

1. Slot order (apply in this sequence):

   1. Scope — a diff scope that limits the audit to changed files and their dependents.
   2. Depth — a quick scan or a deep audit, never both.
   3. Environment — a live environment that runs the checks against a deployed build.
   4. Post-stages, in this order — findings verification → fix-and-verify → re-audit.
   5. Format — an output format, always last.

   A forensic root-cause debugging mode is not a slot: it replaces the base prompt's checklist with a root-cause investigation and accepts only the Format slot.

2. Valid and invalid combinations:

Combination	Result
Quick scan + deep audit	Invalid — one depth modifier only
Quick scan + fix-and-verify	Valid — fixes Critical only
Diff scope + deep audit	Valid — read-everything and reproduce-everything, but only inside the diff and its dependents
Diff scope + live environment	Valid — only surfaces touched by the diff are driven live
Diff scope + re-audit	Valid — prior findings outside the diff are reported OUT OF SCOPE, not reclassified
Live environment without credentials	Invalid — ask, do not improvise
Re-audit without the prior commit	Invalid — a delta needs an anchor
Fix-and-verify without findings verification	Valid only if each finding is reproduced before its fix, which fix-and-verify already requires
Forensic debugging + any slot other than format	Invalid — it is a different investigation, not a modified audit

3. Precedence: the base prompt's evidence rules always apply — Confirmed still needs tool-produced evidence, severity is still capped at Medium without it, clean is still a valid outcome. A modifier may narrow scope, change depth, add an environment, add stages, or change format. It may never lower the evidence bar, remove the anti-hallucination guard, or authorise an action the base prompt forbids (real-money purchases, production writes, hook bypasses). The repository's own CLAUDE.md and documented conventions outrank every modifier and the base prompt alike.

4. Conflicts inside a slot: when two instructions still disagree after slot order, the more restrictive one wins for scope and writes (smaller surface, fewer side effects) and the more demanding one wins for evidence (reproduce rather than assert).

5. Declare the stack: put one line at the very top of the invocation, in slot order, for example Modifiers: diff scope → deep audit → live environment → findings verification → fix-and-verify. The report's first line echoes it verbatim as Active modifiers: so a reader six weeks later knows what the report does and does not claim. An undeclared stack is reported as Modifiers: none and the modifiers are ignored.

Out of Scope

These rules govern combination only. Each modifier's own behaviour, inputs, and report additions live in that modifier's own text; the base prompt's checklist and output format are unchanged except where a modifier says otherwise. New modifiers join the slot they most resemble and state their slot and conflicts in their own Stacking note.