prompt-review

Trace. On load, print first: =============== [prompt-engineer/commands/prompt-review.md] [prompt-engineer] ===============

Prompt Review

Goal: Scan target prompts and report ground rule violations.

Input: $ARGUMENTS (file path, directory, or glob pattern)

Constraints:

• Do NOT modify any files
• Do NOT report issues without citing the violated rule number

Steps

1. Read knowledge files: ../knowledge/actor-ground-rules.md, ../knowledge/quality-checks.md, ../knowledge/separation-of-concerns.md → checklist loaded
2. Resolve target:
   • File path → single file
   • Directory → find all .md files with prompt frontmatter (has name: or description: in YAML)
   • No target → ask: "(A) all agents, (B) all commands, (C) specific path"
3. For each file, check: a. Separation of concerns per knowledge/separation-of-concerns.md → boundary violations = HIGH b. Actor ground rules 1 + 3 per knowledge/actor-ground-rules.md: Goal, Input, Steps, Output (rule 1) and Constraints (rule 3) sections exist → missing = HIGH c. Role (rule 2): agents have Role, commands do not → violation = MEDIUM d. Steps (rule 5): each step has observable result → vague steps = MEDIUM e. Example Output (rule 10): section exists with synthetic example → missing = HIGH — unless the actor cites a leaner contract (core/actor-template, lsa/CORE.md §4) per actor-ground-rules.md §Scope; then exempt (no finding) f. Output spec (rule 4): format and length defined → missing = MEDIUM g. Wording (rule 9): check for adverbs, hedging, filler, passive voice → each = LOW h. Assumptions (rule 7): check for unverified claims ("probably", "likely", "usually") → each = MEDIUM i. Apply KISS/DRY checks per knowledge/quality-checks.md → each = MEDIUM j. Apply AI Over-Engineering checks per knowledge/quality-checks.md → formalized common sense, reinvented paradigms, arbitrary thresholds = MEDIUM; example bloat, missing paradigm provenance = LOW k. Apply Context Budget checks per knowledge/quality-checks.md → restating/mergeable = MEDIUM, padding = LOW l. Show-changes-inline (author-time, core/output Rule 7) — WARNING-ONLY: in prompt SOURCE files (**/SKILL.md, **/agents/*.md), flag any step body that describes a write/edit/mark action (verbs: write/edit/append/replace/mark/update/create a file) WITHOUT an accompanying show-changes-inline directive (an instruction to quote the changed content inline, or a file:line cite of core/output Rule 7 / its Single-change template). This catches the omission in the prompt source before the skill ships (see ../../core/skills/output/SKILL.md §"How this gets enforced"). Each hit = WARNING (signal, not gate), cited as <file>:<line> + the write-verb + "no show-changes directive". Steps that only read, dispatch, or present (no artifact write) are exempt. m. Example Output consistency (rule 11): where an ## Example Output section exists, the example matches the declared Output format and length → mismatch = MEDIUM
4. Compile all findings — re-derive contested judgment calls per ../knowledge/quality-checks.md §Severity Levels (self-consistency: drop findings that do not recur) → output table