Skip to main content

​
Multi-Pass Editorial Review Without the Bottleneck

What this solves: Strategic documents get three independent quality audits in parallel instead of one slow sequential pass, catching errors that any single reviewer would miss.

​
The Problem

Reviewing a strategic document — a positioning brief, an executive one-pager, a competitive analysis — typically means one person reading it once and catching whatever they happen to notice. Or worse, running it through multiple review passes sequentially: first structure, then claims, then voice. Each pass takes the full read time, and by the third pass, you’re fatigued and skimming. The reviews I cared about most (accuracy, clarity, AI pattern detection) were independent of each other. An unsupported claim doesn’t depend on a structural clarity issue. A buried lede doesn’t depend on hedging language. These audits could run at the same time if I had three reviewers who each knew what to look for.

​
How It Works

The system detects document stakes automatically, then dispatches the right level of review. Each agent gets the full document and a specific audit protocol. The clarity agent runs structural sweeps — sentence length, section density, whether the lede is buried. The accuracy agent inventories every factual claim and checks which ones have evidence. The voice agent scans for AI writing patterns using a taxonomy of 25+ tag categories (hedging, corporate language, stacked adjectives, the telltale rule-of-three). The synthesis layer does the reconciliation. When two agents flag the same passage — say, a hedging phrase that’s both a voice problem and a clarity problem — the system deduplicates and attributes to the primary domain. Findings get ranked into critical (fix before sharing), important (fix if time allows), and minor (nice to have). Each dimension gets a 1-10 score. For high-stakes documents (anything going to leadership or external audiences), the system offers a follow-up strategic interview — Socratic probing of assumptions and evidence gaps that pattern-based review can’t catch.

​
In Practice

A typical medium-stakes review — a positioning brief, maybe 800 words — takes about 90 seconds with all three agents running in parallel. The same review done sequentially (clarity pass, then accuracy pass, then voice check) used to take 4-5 minutes. The time savings compound when you’re reviewing multiple documents in a session. More importantly, the parallel approach catches things sequential review misses. On one executive one-pager, the accuracy agent flagged an unsupported performance claim while the voice agent independently flagged the same sentence for promotional language. Neither issue alone would have stopped the document from shipping. Together, they revealed that the strongest-sounding sentence in the piece was also the least substantiated — exactly the kind of thing that erodes credibility with a skeptical audience. The stakes-based routing matters more than I expected. Running three parallel agents on a Slack message draft is overhead that doesn’t pay off. The five-check quick sweep handles low-stakes work in seconds. Reserving the full orchestration for documents that actually need it keeps the system from feeling heavy.

​
What I Learned Building This

The model selection for each agent turned out to be a real design decision, not just an optimization. The clarity and voice agents do pattern matching — they’re looking for structural signals and known AI tells. They run well on lighter models. The accuracy agent needs to reason about whether a claim is actually supported by its context, which requires more capability. Matching model weight to task type cut costs without sacrificing quality on the dimension that mattered most. Claude originally built the voice agent to both detect and rewrite AI patterns. That was a mistake. Rewriting mid-review contaminates the accuracy audit — suddenly the accuracy agent is checking claims in text that didn’t exist in the original. Separating audit from edit (detect everything first, fix after the report) was a small architectural change that fixed a whole category of bugs. The scoring system (1-10 per dimension) seemed gimmicky when I added it, but it turned out to be the fastest way to decide whether a document needs more work. A document scoring 8/clarity, 4/accuracy, 7/voice tells you exactly where to spend your time without reading the full report. The number does what three paragraphs of qualitative assessment couldn’t.