I read all three files closely against each other. Overall they’re in good shape — the cross-referencing between stages (Research owns the steps, Writing owns wording, Edit owns cleanup) is clear and mostly conflict-free. Here’s what I found, ordered by importance.
- Internal contradiction in EDIT-INSTRUCTIONS (Pass 4, Part B). The hard example says “The watch shop sells luxury timepieces” is wrong. But the “Fix by” list a few lines later says the business may be rewritten so it “produces, supplies, offers, manages, provides, coordinates, stocks, sells, displays, serves, or supports the result.” So “sells” is both forbidden and approved in the same pass. A medium-effort model resolves this by context; a low-effort model may not. Remove “sells” (and consider “displays”) from the acceptable-verb list, or align it with the Writing file’s list.
- The two acceptable-verb lists don’t match. Writing says a business can “provide, offer, operate, specialize in, be structured as, require, accept orders, or serve a market.” Edit uses a different list. Harmonizing them to one identical list means the writer and editor judge sentences by the same standard, which reduces edit churn.
- Dead-weight currency/date rules in EDIT-INSTRUCTIONS. The Language and Style Standard includes “Currency: $1,000” and “Always use numerals for money.” Since your system bans dollar amounts entirely, these rules never apply — and worse, they could signal to a weaker model that dollar figures are expected content. I’d cut the currency line or replace it with a one-line reminder that articles should contain no dollar amounts.
- The 25-word cap makes the 2-sentence allowance nearly moot. Two sentences within 25 words must average ~12 words each. That’s fine if intended, but stating both limits in four places (Writing rules, Writing audit, Edit checklist, Edit audit) costs tokens and gives a low-effort model two rules to reconcile instead of one. You could simplify to “25 words max, 1–2 sentences” stated once per file.
- Number style exists only in the Edit file. “Spell out one through nine” isn’t in the Writing file, so the writer produces inconsistent numbers the editor must fix. One line added to Writing closes that gap.
- Leftover “opening readiness” phrasing. Research still uses “opening readiness” (Context Rule) and “opening-stage readiness” (Startup Scope). Since Research is internal-facing this may be deliberate, but if the pre-opening phrasing decision was meant to be system-wide, those remnants could leak into research output and then into articles.
- For the low-effort problem specifically: the files are long, and weaker runs miss buried rules. Adding a short “Hard Rules — Never Violate” block of 5–7 lines at the top of each file (paragraph limits, no dollar amounts, human-actor rule, output format) would give a cheaper model the critical constraints up front even if it skims the rest. This is probably the single highest-leverage change for your token-cost goal.
Nothing here is broken enough to be causing bad output on Sonnet 4.6 at medium effort — items 1 and 3 are the only true conflicts. Let me know which items you want, and I’ll rewrite the affected files.