Routing Copy Tasks Between a Small and a Large Model

The rule should fit on one page and be readable by a non-engineer, because the person who writes the briefs should be able to predict the route. Adopt this and edit it.

Send a task to the small model only when all four conditions hold. The brief asks for fewer than 150 words. It contains at most two factual slots (product name and price, say). An approved reference draft exists in the same format, channel and tone. The task needs no original argument, point of view, or claim the brand has not already made.

Send everything else to the large model: any task over the word limit, any task with three or more constraints (a headline carrying a benefit, a number and a legal qualifier), any task with no approved reference, and any task that must take a position.

The "approved reference" condition does the most work. A small model has less room to infer an unstated format, so a finished example of the exact output shape replaces inference with pattern matching. Few-shot examples are a standard prompt section for exactly this reason; a worked example of ideal behavior is one of six required sections in a production prompt structure the Vapi prompting guide.

The "original argument" condition catches where small models fail most visibly. A piece that must explain why a category is changing needs several claims held in tension and a choice about which to keep. That is reasoning depth, and reasoning depth is the first thing to check before routing down.

The last row matters. A static rule is usually the best first implementation because it is easy to understand and evaluate the PrompTessor routing guide. Give it an escape hatch anyway. If small model output fails a cheap structural check (wrong length, missing slot, banned phrase), escalate that task to the large model rather than regenerating small. That is cascade routing, and cascades cut blended cost substantially when most traffic resolves in the cheap tier the TrueFoundry routing writeup.

One caution. Learned routers that predict the best model can generalize poorly when deployment traffic differs from training data, and a simple rule-based alternative has held up better across domains in published comparisons the task-aware routing paper. For a workload with a handful of stable task types, a written rule is more transparent and more robust. Reach for a classifier only when task types are too many to enumerate or briefs arrive untagged.

def route(task):
    if task.legal_review_required:
        return "large"
    if task.words > 150 or len(task.factual_slots) > 2:
        return "large"
    if not task.has_approved_reference or task.needs_original_argument:
        return "large"
    return "small"

Keep the route in your production trace next to the output, so a bad draft shows which path produced it. Tracking the chosen model is a standard debugging recommendation the PrompTessor routing guide.

The rule decides who writes. The voice test decides whether the cheap writer stays. Both models are scored on the same rubric over the same tasks.

Build a 20 item rubric from your style guide: four dimensions, five items each, scored 0, 1 or 2, maximum 40.

• Tone markers. Does the draft use the signals your guide lists (direct address, contractions, a specific rhythm) and avoid the banned ones?
• Banned phrases. Does it avoid your blocklist? AI-assisted drafts overuse a recognizable word set, and readers notice even when they cannot name it Ruben Hassid's banned words list.
• Sentence length spread. Does length vary, or does the draft settle into a uniform rhythm?
• Specificity. Does it use the facts it was given, or retreat into generalities? "Brushed aluminum" in the brief must not become "high quality materials" in the draft.

Score each model on the same 20 tasks, five per task type, with two raters who do not know which model produced which draft. Blind scoring is not optional. Raters who know a draft came from the cheap model will score it differently without meaning to.

Then set the pass gap. The small model keeps a task type only if its mean score is within a fixed gap of the large model's. Ten percent of maximum (4 points of 40) is a reasonable start.

Worked example. The large model averages 34 of 40 on product descriptions and the small model averages 31. The 3 point gap is inside tolerance, so descriptions stay small. On subject lines the large model averages 33 and the small model 26. Seven points is outside tolerance, so subject lines return to the large model even though they are short.

That is the insight most routing tutorials miss. Difficulty and risk are not the same thing. Short copy can be the highest-stakes copy you produce, because it carries the brand in eight words. Length is a proxy for cost, not for risk.

For continuous monitoring, an LLM judge can score against the same rubric. Judges are one of four common evaluation methods alongside multiple choice, verifiers and leaderboards Sebastian Raschka's overview of LLM evaluation. Validate the judge against your two human raters on the same 20 tasks before trusting it alone; use it between audits, not instead of them.

Four numbers: the large model's full price per draft, L; the small model's, S; the share of drafts passing the rule, p; the per-draft cost of the routing decision and quality check, C.

cost per 1,000 = 1000 × (S × p + L × (1 - p) + C)

Worked example with round numbers you should replace with your invoice. The large model costs 3 cents per draft at your typical length; the small model costs 0.3 cents, a 10 to 1 ratio. Sixty percent of drafts pass the rule.

All large: 1000 × 0.03 = $30.00. Routed: 1000 × (0.003 × 0.6 + 0.03 × 0.4) = $13.80. The saving is $16.20 per thousand, 54 percent. Note the blended price is still 0.4 of the large price: the expensive model sets the floor, so a 60 percent easy share buys 54 percent, not 60.

Add checks at 0.05 cents each for routing and scoring, so C is 0.1 cents and routed cost becomes $14.80. The honest saving is $15.20.

The decision rule: routing pays when the easy share times the price ratio is large compared with checking cost. At 10 to 1 you need roughly a 20 percent easy share to cover $1 per thousand in overhead; at 2 to 1 you need about 90 percent.

Then compare tokens to human time, because that is where the money is. A reviewer spending two minutes at a loaded $40 an hour costs $1.33 per draft, or $1,330 per thousand. A $16 token saving is about 1 percent of that. So the question is never which model is cheaper. It is whether the cheap model cuts reviewer minutes without adding rework. A small model that saves $15 per thousand but adds a minute of rework to 20 percent of drafts costs hundreds more. Run the voice test before the cost math, not after.

For teams running several workloads, the natural endpoint is one policy layer holding the rule, rubric and escalation path, with model calls behind a single key so that per-model costs land on one invoice. That is what Quanta Inference is for.

If the real gap is that you have no brief template, references or rubric yet, the bottleneck sits upstream of model choice. A specialist who writes the brief and rubric with you will save more than any router. That is a fair place to bring in Quanta.

Some tasks are excluded regardless of length or slot count. Regulated claims (health, financial, environmental) need evidence behind every statement, and advertising claims must be truthful, non-deceptive and evidence-based the FTC advertising and marketing basics. Endorsement copy has its own rules: a material connection with the brand must be disclosed clearly and conspicuously where readers will see it the FTC disclosures guide. Put a hard filter at the top of the rule: regulated claims, endorsements or claims about real people go to the large model and then to a human.

Separately, write your AI-disclosure policy down before a campaign needs it. Obligations vary by jurisdiction and medium, and a rule applied by only one of two models is a rule you will break by accident.

Pull 50 routed drafts at random each month and score them on the same rubric. If the small model's gap widens past tolerance on a task type, stop routing it until you know why. Drift usually comes from changed briefs or from reference drafts being replaced with small model output. That second failure is common and subtle: if the approved reference for a task type becomes a published draft the small model wrote, you are teaching it to imitate itself. Keep references human-edited.

Log the route with every rejected draft. After a few weeks you have a precise list of task types to move up the ladder. And re-run the voice test whenever either model changes: you are choosing a pairing, not a model, and keeping the classifier separate from the model calls lets you swap one side without rewriting the rule.