Note This white paper describes Agency689's internal AI writing systems. It is a methodology overview and architecture sample. Specific prompts, brand-knowledge structures, evaluation rubrics, and client-confidential details remain proprietary.
AIW
Agency689 · AI Writing Systems
White Paper · Architecture · Methodology
Draft · 03.15.26
Creative Operations · Brand Voice · Campaign Production · Proposals Agentic Processing Layer · A689
Sunset Marquis · Agency689 · Traveler Guitar By Adam R. Cagle

Building Human-Reviewed
AI Workflows for Brand Voice, Campaign Production, Proposals, and Creative Scale

Architecture · Voice · Workflow · Approval · Recursive Learning
00

Executive Summary Overview

Most organizations are experimenting with generative AI as a faster way to produce words. Agency689 has taken a different approach, using AI to build repeatable creative production systems that preserve brand voice, widen the range of viable first drafts, accelerate campaign throughput, and keep humans in control of the decisions that matter.

The Agency689 AI Writing Systems began as a practical response to recurring creative demands across long-running client relationships. For Sunset Marquis, the system supports high-frequency email, banner, and promotional copy production inside a brand voice that took years to build and cannot be allowed to flatten. For Agency689 itself, it supports communications and proposal generation by transforming discovery notes, project requirements, and past precedent into structured, editable business documents. For Traveler Guitar, the same systems thinking extends beyond copy into visual ideation, helping create surreal campaign worlds where the product can live without losing the brand's sense of adventure.

These are not one-off prompts. They are project-based AI workflows built from curated brand context, retrieval-augmented source material, persona and SKILL.md governance, multi-pass drafting, standards review, human approval routing, and recursive workflow learning.

Operational outcomes to date
  • AI writing and marketing workflows are used daily across Agency689 client work.
  • The Sunset Marquis writing workflow alone runs roughly three times per week during active campaign cycles.
  • The systems cut production time by approximately half.
  • They generate three or more viable first drafts in less time than one first draft previously required.
  • They are used by Adam, the Agency689 team, and in some cases by clients.
  • Nothing publishes automatically. Outputs pass through standards review and human approval.
The goal is not to replace copywriters, art directors, or strategists. The goal is to move the creative team past the blank page faster, with more options, better context, and a stronger first pass.
01

Creative Teams Do Not Need More Generic Copy The Problem

Generative AI makes producing language cheap. It does not automatically make producing good, brand-specific, commercially useful language easier.

For a creative agency, the problem is not simply volume. The problem is that recurring work often lives at the intersection of tight deadlines, deep brand histories, highly specific voice rules, campaign continuity, factual constraints, channel-specific formatting, and the need for fresh thinking under pressure.

A monthly promotion is not just a monthly promotion if the brand is Sunset Marquis. A proposal is not just a scope of work if it must sound like Agency689, reflect the actual value of the engagement, and pull intelligently from years of precedent. A guitar advertisement is not just an image prompt if it must hold the emotional territory of The World's Most Adventurous Guitar while still showing a real product credibly and beautifully.

Two common failure modes
  1. It produces competent beige. Grammatically fine copy that could belong to anyone.
  2. It over-indexes on stylistic mimicry without understanding the strategic logic beneath the style.

Agency689's systems were built to solve both problems.

02

AI as a Creative Production System Design Principles

The Agency689 approach treats AI not as a chatbot that occasionally helps with drafting, but as a structured production layer inside creative operations. The system rests on six design principles.

// 01

Retrieval before generation

The model works from brand books, prior approved copy, campaign references, product facts, tone rules, and active briefs before it writes a word.

// 02

Voice rules are operational data

A brand voice is not a mood board. It translates into explicit rules, preferred constructions, forbidden clichés, approved registers, and examples of what good actually sounds like.

// 03

Breadth before refinement

The first value of AI is not perfection. It is the ability to produce multiple credible directions fast enough that humans choose the best strategic path sooner.

// 04

Critique before approval

A second-pass standards layer evaluates drafts before humans spend time on them. It checks for tone drift, factual error, repetition, generic filler, policy issues, and brief drift.

// 05

Humans remain the publication gate

The system can draft, revise, score, and route. It does not publish. Human judgment remains explicit and mandatory.

// 06

Every approval improves the next cycle

Approved outputs, rejected tropes, editorial corrections, revised instructions, and campaign decisions become operating memory for future work.

Important distinction. This is not a claim that the foundation model's weights are retrained after every project. The learning happens at the workflow and retrieval level. Better source material, better rules, better examples, better evaluation prompts, and better routing of human decisions.

03

System Architecture Seven-Layer Stack

The Agency689 AI Writing Systems use a modular architecture that adapts to different brands and output types. Each layer has a defined responsibility, and the stack runs end-to-end inside every project.

ABrand
Brand Knowledge Layer

A curated project knowledge base assembled per client and updated continuously.

  • Brand books
  • Style guides
  • Persona files
  • SKILL.md
  • Approved headlines
  • Email history
  • Promotions
  • Service descriptions
  • Product facts
  • Do/don't lists
  • Prior proposals
  • Scope language
  • Performance notes
  • Visual references
  • Deprecated copy
BRetrieval
Retrieval Layer

A RAG-supported process surfaces the most relevant source material for the brief at hand. Retrieval is selective, not generic.

  • By client
  • By campaign type
  • By channel
  • By season or event
  • By product or offer
  • By historical similarity
  • By recency
CIntake
Brief Intake Layer

Rough inputs convert into a structured working brief with missing-information flags where needed. Inputs may include email requests, promotional calendars, limited client notes, seasonal offers, event details, product facts, proposal requests, discovery notes, or visual campaign directions.

DDraft
Draft Generation Layer

The system produces multiple first-draft options aligned to the intended channel.

  • Subject lines
  • Preheaders
  • Email body
  • Display ads
  • Banners
  • Social posts
  • Landing blocks
  • Concept territories
  • Proposal narratives
  • Scope sections
  • Follow-up emails
  • Image prompts
  • Video concepts
EStandards
Standards AI Layer

A second AI pass audits outputs before human review and can trigger revision before the work reaches an editor.

  • Voice drift
  • Forbidden language
  • Factual error
  • Channel mismatch
  • Overclaiming
  • Repetition
  • Weak CTA logic
  • Poor hierarchy
  • Brief contradiction
  • Creatively dead copy
FReview
Human Review Router

Outputs route for review through Telegram or an equivalent approval path. Humans may approve, request a revision, select a direction, reject a premise, or add new factual context. Edits are not noise. They are signal.

GLearning
Recursive Workflow Learning Layer

Approval outcomes, revision patterns, and standards critiques improve future behavior through updated prompt rules, new brand examples, revised SKILL.md guidance, campaign-family memory, preferred offer framing, channel learnings, and sharper evaluation criteria.

04

Architecture at a Glance System Flow

A brief enters from the left and reaches publication only after passing every gate in the system. The learning loop closes back into Brand Knowledge, where the next brief begins with stronger context than the last.

// 04 · SYSTEM FLOW A689 / AIW 01 · INPUT CLIENT / TEAM BRIEF rough inputs · notes · calendar 02 · INTAKE STRUCTURED BRIEF normalized · flag missing info 03 · RETRIEVAL BRAND + CAMPAIGN RAG docs · examples · rules 04 · GENERATE MULTI-DRAFT GENERATION copy · proposals · art prompts 3+ directions per assignment 05 · CRITIQUE STANDARDS AI REVIEW tone · facts · hierarchy may trigger revision loop 06 · APPROVAL HUMAN REVIEW ROUTER Telegram · team review publication gate 07 · OUTPUT FINAL PACKAGING publish-ready materials 08 · LEARN RECURSIVE WORKFLOW LEARNING approval signals update rules · examples · SKILL.md · retrieval priors revise FEEDBACK LOOP FORWARD FLOW · BLACK LEARNING RETURN · ACCENT HUMAN GATE · STEP 06 Eight stations. One loop. Nothing publishes without a human.
Eight-Phase Flow · Standards Review Closes Before Human · Learning Returns to Retrieval
05

Sunset Marquis Campaign Production System One

5.1 Why Sunset Marquis Was the Right Test Case

Sunset Marquis is not a forgiving environment for generic hospitality copy. Agency689 has helped steward the brand for more than two decades. The voice is distinct, established, and continuously evolving. It must be cool without trying too hard, exclusive without sounding stiff, atmospheric without dissolving into fluff, and accurate down to the specific restaurant, villa, spa, or seasonal offer being promoted.

The workflow supports three especially recurring production categories.

  1. Email campaigns
  2. Banner and digital ad copy
  3. Promotional copy blocks

5.2 Inputs

The system may receive monthly promotional calendars, seasonal campaigns, holiday menus, room or villa updates, restaurant or spa promotions, event schedules, occupancy-oriented booking pushes, historical campaign references, and any factual restrictions tied to renovations, availability, or service offerings.

5.3 Retrieval Context

The Sunset workflow retrieves from a curated brand memory that may include approved email history, website language, house style rules, campaign-level tone rules, CTA patterns, promotional phrasing that has been approved before, vocabulary to avoid, and channel-specific copy examples.

The system can be tuned to understand rules such as:

  • Do not use generic hospitality language.
  • Do not flatten the brand into interchangeable luxury-hotel copy.
  • Preserve the brand's cool, confident, slightly mysterious tone.
  • Keep language experience-led rather than over-descriptive.
  • Maintain the editorial discipline Agency689 expects.

5.4 Outputs

A single promotional brief can produce subject line options, preheaders, three-paragraph email body copy, hero banner copy, secondary banner copy, paid display variants, web promotional blocks, and optional social or ad extensions.

The workflow produces multiple credible directions rather than one answer. This shifts the human reviewer from blank-page authoring to editorial selection and elevation.

5.5 Standards Review

Before human review, a standards pass evaluates whether the copy sounds like Sunset Marquis, whether it accidentally introduces unsupported facts, whether it repeats phrases that have been overused, whether the CTA fits the promotion, whether the copy is sufficiently differentiated between email and banner contexts, and whether the work has drifted toward generic hotel language.

5.6 Human Approval

Approved or critique-ready outputs are routed for human review. Adam or the team may select the winning direction, request a fresh angle, revise a headline, adjust factual nuance, or update the voice system when the critique reveals a repeat issue.

5.7 Operational Result

The Sunset Marquis workflow runs roughly three times per week during active campaign cycles. Across Agency689 more broadly, AI writing and marketing systems are used daily.

The result is not merely faster copy. The result is more strategic options earlier, faster first-pass execution, stronger brand consistency, more efficient reuse of approved context, and less energy wasted rebuilding a brand voice from scratch every time a new promotional need arrives.

06

Communications and Proposal Generator System Two

6.1 The Problem

Creative agencies write a surprising amount of consequential business material that is not campaign copy. Proposals, scopes of work, outreach emails, pitch follow-ups, internal summaries, capability narratives, client-facing process explanations, and strategic positioning documents.

These materials often begin from fragmented notes, emails, meeting takeaways, or rough bullets. The challenge is not simply wording them. The challenge is turning incomplete inputs into a clear, persuasive, appropriately scoped business document that reflects how Agency689 actually works.

6.2 Proposal Generator Inputs

The proposal workflow can draw from client discovery notes, email chains, prior proposals, scope templates, service descriptions, Agency689 positioning language, standard payment and deliverable structures, example timelines, and precedent from similar engagements.

6.3 Outputs

Depending on the assignment, the system can generate polished outreach emails, proposal introductions, executive summaries, project objectives, phased scopes of work, deliverables tables, timeline language, payment terms language, process narratives, follow-up correspondence, and tailored language for specific prospects or categories.

6.4 Why It Matters

Proposal writing is often where agencies either sound clear and senior or diffuse and generic. A strong proposal generator helps preserve tone, institutional memory, commercial precision, and the ability to move quickly without sacrificing judgment. The system also reduces the chance that a strong opportunity receives a weak first draft simply because the team is busy.

6.5 Suggested Next-Phase Enhancements

  • Scope gap detection. Identify where a proposed deliverable set lacks implementation, revision rounds, training, or handoff language.
  • Pricing precedent retrieval. Surface comparable past work without exposing confidential information.
  • Risk flagging. Call out ambiguous asks, missing decision-makers, or unclear approval paths.
  • Proposal-to-contract consistency. Ensure that what is promised in the proposal is reflected in later agreement language.
  • Tone modes. Formal, friendly, decisive, or founder-led depending on prospect type.
  • Before-and-after memory. Capture what edits Adam or Zander tend to make and use those patterns to sharpen future drafts.
07

Traveler Guitar Surreal Art Generation System Three

7.1 The Creative Challenge

Traveler Guitar's long-running brand idea, The World's Most Adventurous Guitar, creates an unusually fertile campaign territory. The work must make the product feel portable, iconic, and alive with possibility. Agency689's recent Traveler work pushed that territory into surreal visual storytelling. Guitars played in impossible places, natural scenes expanded into dreamlike worlds, and the product placed inside environments that feel as adventurous as the headline claims.

The visual system supports campaigns built around lines such as:

  • Next Stop, Greatness
  • Sonic Bloom
  • Endless Strummer
  • A Force of Nature
  • Hey Jungle, Welcome to Me
  • Flying Solo

7.2 AI's Role

The role of AI is not to make an ad. It is to accelerate visual exploration, widen compositional options, and help art directors test campaign worlds at a speed that would otherwise be impractical. The system can support surreal setting ideation, mood and material exploration, environmental storytelling, photographic reference generation, background generation, fantastical natural elements, product-scene prompt construction, and refinement passes that help human designers move faster into final compositing.

7.3 Human Art Direction

Human designers remain central. Final campaign images are refined, corrected, composited, and made credible through human art direction. The system is valuable because it helps the creative team explore more possibilities before locking into a final direction.

7.4 Potential Workflow

// 07 · VISUAL WORKFLOW · TRAVELER GUITAR A FORCE OF NATURE 01 Campaign line selected 02 Visual territory defined 03 Prompt families setting · light · mood 04 Image territories explored at speed 05 Art director critique pass 06 Prompt refinement targeted second pass 07 Designers composite final campaign art 08 Visual language saved for reuse Eight stations from line to locked campaign world.
Linear Workflow · Human Critique at Station Five · Designer Finishing at Station Seven

7.5 Suggested Next-Phase Enhancements

  • Visual prompt libraries organized by Traveler tagline or campaign territory.
  • Style-memory boards that preserve successful lighting, surrealism level, and composition logic.
  • Product accuracy checks to reduce distortion of guitar proportions, hardware, and branding.
  • Art direction critics that evaluate generated images against campaign goals before they reach designers.
  • Multi-format generation logic for magazine ads, billboards, social crops, and trade-show materials.
  • Asset lineage tracking so the team can see which prompt family produced which final approved piece.
  • Creative novelty controls that prevent future Traveler campaigns from recycling the same fantasy-world tropes.

7.6 Why This Belongs in a Writing Systems White Paper

Because the deeper system is not limited to text. It is a creative operations architecture. Brand context, strategic rules, promptable execution, standards review, human approval, and recursive accumulation of what works. Traveler Guitar demonstrates that the same architecture that helps write better copy can also help create stronger visual campaign territories.

08

Why One Model Should Critique Another The Standards AI

One of the key distinctions in Agency689's workflow is the use of a second standards AI before human review. The drafting model is encouraged to explore. The standards model is encouraged to interrogate.

// 08 · TWO MODELS · ONE TENSION A689 / AIW 01 · DRAFTING AI Explores. Generates breadth. Multiple credible angles. Strategic territory. Tone variation. Permission to take swings. 02 · STANDARDS AI Interrogates. Did this follow the brief? Does it sound like the brand? Did it accidentally overclaim? Would a senior editor reject this? Permission to be ruthless. REVIEW The first pass expands. The second pass disciplines. The human reviewer receives a real reviewable draft, not raw model exhaust.
Drafting Encouraged To Explore · Standards Encouraged To Interrogate · Human Receives Reviewable Draft
The standards layer asks
  • Did this follow the brief?
  • Does it sound like the brand?
  • Did it accidentally overclaim?
  • Did it use language the brand avoids?
  • Does it repeat old phrasing too closely?
  • Is the hierarchy clear?
  • Is the CTA appropriate?
  • Would a senior human editor reject this immediately?

This creates an internal tension that improves quality. The first pass expands. The second pass disciplines. The human reviewer receives something closer to a real reviewable draft rather than raw model exhaust.

09

Recursive Workflow Learning Memory · Curation

The phrase recursive learning can sound grandiose if it is not defined. In the Agency689 system, it means something precise.

// 09 · LEARNING CYCLE CURATED · NOT PASSIVE 01 brief Brief produces outputs 02 critique Standards AI critiques 03 approve Humans approve / revise / reject 04 capture Decisions captured 05 update Brand guidance + prompts updated 06 stronger Next brief begins stronger Recursive Workflow Learning curated improvement
Six Stations · Closed Loop · Each Pass Sharpens The Operating Context

This is not passive accumulation. It is curated improvement.

Examples of recursive learning signals
  • Headline types that repeatedly survive review.
  • CTAs that consistently feel wrong for a brand.
  • Clichés or filler words frequently removed by humans.
  • Offer framings that preserve brand tone under promotional pressure.
  • Proposal clauses Adam routinely adds back in.
  • Art-generation prompts that create useful Traveler worlds versus dead-end imagery.
  • Channel-specific length or structure corrections.

Over time, the system becomes more useful because the agency becomes more explicit about its own standards.

10

What Makes This Different Generic AI vs A689

Generic AI Use Agency689 AI Writing Systems
One-off promptReusable project workflow
No durable contextBrand knowledge base and retrieval layer
One answerMultiple strategic directions
Human edits after raw outputStandards AI before human review
Tone described vaguelyVoice rules, examples, persona files, SKILL.md guidance
No memory of approvalsRecursive workflow learning
Mostly text-onlyCopy, proposals, and visual ideation
Speed is the main goalSpeed, quality, consistency, and adoption
11

What Remains Human Authorship · Accountability · Taste

The Agency689 systems are deliberately designed around the belief that the most valuable creative decisions remain human.

Humans still own
  • The brief.
  • The strategic leap.
  • The final headline judgment.
  • Whether a draft actually has taste.
  • Whether a campaign territory is worth pursuing.
  • Whether a proposal says enough or promises too much.
  • Whether an image feels alive or merely generated.
  • Whether the final work should exist at all.
The AI system expands the field of options and compresses repetitive production. It does not replace authorship, accountability, or taste.
12

Measured and Observable Results Live Snapshot · 03.15.26

DAILY
Use Across Agency689 Client Work
3×/wk
Sunset Marquis Workflow Cadence
~50%
Reduction in Production Time
3+drafts
In Less Time Than One Previously Required

Beyond the headline numbers, the system has produced increased consistency across recurring promotions and campaign families, faster proposal and communication drafting inside Agency689, broader visual exploration for Traveler Guitar campaign worlds, and team adoption beyond the person who built the systems.

13

Next-Phase Build Opportunities Roadmap

// 13.1

Evaluation Dashboard

First-pass acceptance rate, average human revision depth, time from brief to approved draft, standards-AI pass/fail, options per assignment, and recurring rejection reasons.

// 13.2

Brand Freshness Governor

A retrieval filter that favors recent approved work when the brand has evolved, while still allowing older canonical copy to be retrieved intentionally.

// 13.3

Novelty Guardrail

A critic that identifies when AI drafts are too similar to previous approved campaigns and requests fresh conceptual territory.

// 13.4

Proposal Risk Checker

Flags missing scope, ambiguous deliverables, unaddressed implementation needs, mismatched payment terms, or client responsibilities that should be stated explicitly.

// 13.5

Visual Accuracy Critic

For Traveler Guitar. Identifies distorted product proportions, implausible placement, inconsistent brand treatment, and environmental concepts that are visually interesting but strategically off-brand.

// 13.6

Campaign Outcome Memory

Where reliable performance data exists, connect approved creative approaches to observed outcomes without overstating causality. The point is not to chase the metric. The point is to make creative history queryable.

// 13.7

Unified Creative Brief Compiler

A layer that accepts messy source inputs and outputs a structured brief used by copy, visual, and proposal systems alike. Reduces repeated translation of client asks across departments.

14

Creative Operations Design Conclusion

The Agency689 AI Writing Systems were not built to prove that AI can write. That proof is trivial and no longer interesting.

They were built to answer a harder question.

Can AI be structured so that it helps a creative team move faster without becoming less itself?

For Agency689, the answer is yes, when the system is grounded in real brand memory, retrieval before generation, explicit voice governance, critique before approval, humans as the publication gate, and recursive workflow learning that turns repeated work into institutional intelligence.

Sunset Marquis demonstrates how a long-standing brand voice can be extended into high-frequency campaign production without flattening it. Agency689's proposal and communications workflows show how internal business writing can become faster and more coherent without losing senior judgment. Traveler Guitar shows that the same architecture can expand into visual campaign ideation, producing more adventurous creative territory while keeping art direction human.

This is not content automation. It is creative operations design.

And for agencies, marketing teams, and brands facing an ever-growing demand for quality at speed, that distinction matters.