How I build creative assets, from first principles to final execution.

The Meta Ad Library is an open book. Most teams skim it. I catalog it.

• Audit 200 to 500+ competitor ads across Meta and TikTok, covering every competitor on the same spectrum and scale level, well past the two obvious rivals
• Treat long-running ads as proof of performance. Recency tells you what they are testing; longevity tells you what is working
• Build a profile per competitor
  • Positioning claim and price point
  • Product use cases
  • The personas and ICPs their creative is visibly targeting
  • The acquisition funnel behind the click
• Tear down each competitor's creative in depth
  • Angles and awareness coverage
  • Hooks and the first three seconds
  • Script structure and copy patterns
  • Format mix and variation families
  • Offers, CTAs, and landing destinations
• Quantify their format mix (UGC vs static vs educational split) and read testing velocity from active ad volume
• Map the category conversation: what everyone says, what nobody says, which claims are saturated and which are unclaimed

Output: a white-space map. In the MitoQ case study this pass surfaced the one position no competitor could occupy, prevention at the source, and the entire campaign hangs off it.

Before a single verbatim gets collected, the first question is where these people actually congregate. Every audience has rooms they talk in, and the rooms differ by category: a subreddit with 40k members, one Facebook group run by a moderator with strong opinions, the comment section under a specific creator, a forum that has been running since 2009. Find the rooms first, because language harvested from the wrong room produces an ad aimed at nobody.

The goal after that is being able to speak as the customer, from inside their head.

• Analyze 100 to 200+ reviews per product for pain language, desire language, and objection patterns
  • Product and Amazon reviews for purchase-moment language
  • Reddit and Quora for unguarded peer talk
  • YouTube and TikTok comments for reaction language
  • Community forums for long-tenure vocabulary
• Extract exact phrases, never marketer paraphrases
  • Cliches and buzzwords, documented per ICP
  • Running jokes and shared references
  • Register and tone: how they talk when no brand is listening
• Harvest the metaphors they reach for unprompted. "The pharmacy on the counter." "Wired but scattered." These become creative raw material
• Build the failed-solution autopsy: the frustration, what they already tried, why it failed them. The gap between what they bought and what they wanted is where the next ad lives
• Weight behavioral signals over stated preferences. What they screenshot, save, and send matters more than what they claim to want

The verbatims feed hooks directly. "I'm not focused, I'm just caffeinated" came out of this pass and opens a NeuroGum VSL unchanged.

• Calculate TAM by awareness level, per ICP and persona, as a distribution: what share of the market is unaware, problem aware, solution aware, product aware, most aware, and what each stage needs from creative
• Turn the distribution into a targeting decision instead of a demographics deck. For PetHonesty, Product Aware plus Solution Aware held 55 to 65% of the market, so the rotation leads there
• Assess market sophistication separately from awareness, using Schwartz's five stages. A stage-two market still buys claims. A stage-four market needs a new mechanism. A stage-five market buys identity. Pitching a mechanism to a market that needs identity is how good products run dead creative
• Build two to three named ICP profiles per brand (the Anxious Caregiver, the Skeptical Optimizer, the Partial Results Prisoner), each with specific pains, desires, objections, and buying triggers
• Identify the underserved segments competitors are ignoring, then check whether the product can credibly serve them before a single concept is written

• Develop two unique mechanisms. The problem mechanism explains why the pain exists and persists despite everything the customer has tried. The solution mechanism explains why this product resolves what the alternatives structurally cannot. Ancient Nutrition's Collagen Type Deficiency is the problem mechanism; the full-spectrum activation stack is the solution mechanism. One creates the itch, the other is the only scratch
• Understand the mechanism at a technical level first, then translate it into customer-facing language without flattening it
• Calibrate the mechanism to the persona. The importance of sleep reads differently to a 60-year-old than to a 30-year-old. Same science, different stakes, different vocabulary, different proof
• Map every proof point: the evidence, its source, and what precisely it proves. Claims that cannot be traced to receipts get cut
• Position the mechanism as differentiation rather than a feature list. The mechanism is the reason to believe, and the reason the cheaper alternative stays in the cart

The creative closes what the offer opens, so both get researched together.

• Price framing is mapped before scripts are written: per-day math, stack-replacement math, cost-of-inaction math. $85 a month meets resistance; $0.85 a day next to a five-product stack reads as relief
• Risk reversal is inventoried as creative material: guarantees, trials, and subscription terms decide how hard a CTA can close
• Subscription mechanics shape the asset list. The 30-day churn window is why timeline creatives exist in three of the four case studies
• The click destination is chosen per awareness stage: advertorial pre-sell for cold education, product page for warm intent, and the ad is written to its lander so both make one argument

Under every stated desire there is an identity purchase.

• Uncover identity-level aspirations, emotional end-states, and the social validation fantasies driving consideration. The MitoQ buyer is not buying CoQ10; she is buying "present and active," the version of herself her grandkids remember
• Run the dimensionalization ladder on every benefit: feature, functional benefit, emotional payoff, identity claim. Copy that stops at the functional rung reads like a label
• Locate the identity gap: the tension between how the customer currently sees themselves and who they are trying to become. Every strong close sells the second self

Whitespace is the position nobody in the category is standing on. It is the output of all the research above, and it is the only part of the research that converts directly into leverage.

Everything up to here maps what exists. This pass maps what is missing, and the gaps show up in five places:

• Creative. Which formats is the category refusing to make? If eleven competitors run polished studio UGC, a hand-held phone recording with no lighting reads as the only honest thing in the feed
• Mechanism. Which explanation has nobody claimed? Most categories sell outcomes and leave the reason unowned, which is why "upstream vs downstream" was available to MitoQ despite the science being public for years
• Messaging. Which true sentence is nobody saying out loud? Usually it is the one that sounds slightly against interest, which is exactly why it is credible
• Format. Which surface is uncontested? Long-form native static, advertorial pre-sell, two-person conversation, and podcast-shaped video are all cheap to enter when the category has standardised on 30-second UGC
• Offer. Which structure, guarantee, or price frame is the category too timid to run?

Two rules keep this honest. Whitespace is only valuable where the product can credibly stand, so the gap gets checked against what is verifiably true before anything is written. And absence is ambiguous: sometimes nobody is running an angle because it does not work. The check is whether the gap is unclaimed or already tested and abandoned, and the customer research usually answers it.

This is where the edge is. Matching the category on their terms means competing on production budget. Standing where nobody else is means competing on the one thing a bigger budget cannot buy.

• All research converges into one master brief: positioning messages, ICP profiles, awareness and sophistication calls, both mechanisms, the verbatim bank, the proof map
• The positioning messages act as the anchor for every creative decision that follows. If an asset cannot be traced back to the brief, it does not get made
• Every messaging recommendation grounds in documented evidence, so 20 to 30 assets cohere as one campaign instead of scattering as 30 opinions

• One reframe, held everywhere. Upstream vs downstream for MitoQ. Energy you can steer for NeuroGum. Expensive because, not expensive despite, for Ancient Nutrition
• The anchor claims the white space the competitive map exposed, phrased in language the customer research proved the audience already thinks in
• Own the language as well as the position. "Patching holes vs preventing them" is engineered to be repeated by the buyer to themselves after the ad is gone

• Every campaign is built around Schwartz's awareness spectrum: unaware, problem aware, solution aware, product aware, most aware
• Each stage gets its own entry point, message, format, and CTA. Stories for the unaware, mechanism education for the solution aware, direct comparison for the product aware, reinforcement for the most aware
• Most DTC brands overinvest in stages four and five and underinvest in one through three, which is exactly where the cheapest attention and the biggest creative opportunity sit

The unit of planning is the intersection: persona × angle × offer, with format last, because format is the cheapest variable to swap.

• Campaigns target an awareness stage with a goal. Ad sets target an ICP intersection with specific creatives. No asset runs without a defined place in the system
• The same creative can hold two jobs. NeuroGum's VSL #3 attacks coffee in Campaign 1 and teaches the pairing in Campaign 2. Same asset, different strategic function, decided by audience
• The architecture is what makes the right person see the right message at the right stage: no impressions wasted educating believers, no selling to people who do not yet see the problem

These are structural, load-bearing before the first line exists. Grouped by the job each does:

Attention. Pattern-interrupt hooks that stop the scroll in the first one to three seconds. Curiosity gaps that only the next line can close. Open loops that hold the viewer through the mechanism.

Emotion. Storytelling before selling. Agitation that names the pain sharper than the customer can. Relatability: talking in their own language, their cliches, their register. Dimensionalization that turns a feature into a lived scene.

Belief. Mechanism-first positioning: the why underneath the what. Proof stacking: clinical, authority, social, in that order of scrutiny. Intensification: one claim escalated with evidence beats five claims stated once. Neurochemical sequencing: curiosity and rapport before education, because logical delivery alone does not close.

Action. Desire manufacturing before the offer, so the CTA harvests instead of pitches. Future pacing that lets the prospect live the after-state. Urgency and scarcity only where they are true. Irresistible offer framing plus risk reversal. One CTA, one action, no decision paralysis.

Creativity is conceived inside constraints. The variables:

• Product and brand truths: what is verifiably true about this specific offer
• Competitive research: what is working, what is saturated, what is being ignored
• The unique mechanisms: the functional reason it works differently, problem side and solution side
• ICP definition: who specifically, never "anyone who needs more energy"
• Awareness and sophistication stage: unaware needs stories, solution aware needs comparisons, jaded markets need identity
• The target market's beliefs, norms, cliches, and buying triggers
• The hook, which is a constraint rather than a finishing touch. A concept that cannot be opened in three seconds is not a concept yet, so the opening gets stress-tested at the idea stage instead of being written last
• Input from live campaigns, when it exists, refining the model over time

You cannot write a winning VSL for an unaware audience using solution-aware messaging. The constraints are the scaffolding that makes volume possible without drift.

On top of the constraints sits everything the constraints cannot generate:

• Organic content: the memes, skits, and native formats the audience already rewards with attention. The Autocorrect static is an iMessage thread because that is what the feed already looks like
• The swipe bank: proven direct response assets, classic and current, studied until their structure is muscle memory
• The personal hive mind: my own knowledge base, the books, the cross-domain content I consume, and the lived experiences that float up mid-process. A hook can come from a walking app, a breakup letter, or a receptor pathway, if the mind is stocked

The least talked-about stage in direct response, and the one that separates volume from noise.

• Load the constraints, saturate on the research, then step away. The subconscious keeps working the intersection while the desk is closed
• Ideas surface on walks, between projects, mid-shower, and they surface pre-filtered, because the constraints were loaded first
• This is why the process front-loads research so hard: incubation quality is a function of input quality. A lean, stocked mind produces angles a prompt never will

Run every candidate concept through the full stack of constraints. Most die there. The ones that survive every circle, true to the product, differentiated from the category, aimed at one ICP, matched to their stage, phrased in their language, are the ones worth producing.

The diagram above is the actual test: a concept that clears only two or three circles is a clever line. A concept at the center is a campaign.

AI produces quantity. The quality of that quantity is inherited from whoever briefed it.

A model handed a thin brief returns the category average at speed, because the average is what it was trained on. The same model handed a real positioning anchor, a verbatim bank, a defined ICP, and a mechanism returns work that is on-voice by construction. Same tool, same hour, completely different ceiling, and the only variable that moved was the person operating it.

So the leverage is never in the generation step. It sits in the judgment loaded before the generation step and the taste applied after it. Which means AI raises a strategist's output and cannot raise their level: it multiplies whatever is already there, including the gaps.

The practical consequence runs through everything below. I do not evaluate a tool by what it can produce. I evaluate it by how much of my own judgment it lets me encode, which is why the pipeline is a library of purpose-built skills instead of a folder of prompts.

• Research synthesis: processes hundreds of competitor ads into pattern maps, extracts customer language from review databases at volume, builds positioning matrices and awareness-distribution reads
• Asset generation: VSL script variants, hook alternatives, static ad concepts, advertorial long-form, angle ideation at 50+ concepts per project. Across the four-brand portfolio that meant roughly 85 script blocks, including 23 VSL scripts for Ancient Nutrition alone
• Architecture support: awareness-level creative mapping, A/B variant families, format recommendations per stage

• Define the constraints AI operates within. The output is only as strategic as the brief it is generated against
• Make the merit calls: which angles have strategic weight versus mere novelty, what gets produced, what gets killed
• Edit for customer-voice authenticity, tighten pacing, sharpen hooks, and strip the cliches AI defaults to. AI reaches for tired patterns; a swipe-trained ear catches them
• Verify every claim is defensible and compliant before anything ships

The engine is a library of purpose-built skills inside Claude and Claude Code, each owning one stage of the pipeline:

• Research: a method-actor research system that gets inside the customer's head before any copy exists: psychology maps, content ecosystem maps (what they watch, read, follow, and which rabbit holes they fall down), customer avatars, verbatim language banks, market awareness analysis. Swipe intelligence pulls from the Meta Ad Library, Taboola, Outbrain, and ClickBank, so every concept is grounded in ads already spending money
• Brand context: a brand-brief creator converts raw research into a reusable brand brain: brand context, avatars, and an angle library, packaged so every downstream skill loads it on demand. One pass, and every later script is on-voice by construction
• Scripting: a reference-based copywriting skill deconstructs proven scripts and rebuilds them for the loaded brand, running internal quality checks before anything is delivered. An advertorial copywriter pairs with a lander engine that turns finished copy into production HTML; the three live advertorials in this Library shipped through it
• Verified stats: a maintained swipe file of source-checked statistics. No number ships without a traceable citation

• Script to shot plan: every VSL is broken into a per-clip prompting document, and script beats are mapped to planned B-roll before anything is generated, so the edit plan drives what gets generated
• Generation through Kling and Seedance, driven by dedicated frame-to-video prompt skills. Frame consistency is engineered: the reference frame's subject is translated into exhaustive written description (face structure, skin texture, wardrobe, camera characteristics) so the same person holds across separately generated clips, with anti-artifact rules riding in every prompt
• One clip per prompt, reviewed before the next is issued. Volume never outruns judgment
• Voice through ElevenLabs, with a consistency system that keeps delivery uniform across separately generated segments, so a multi-clip VSL reads as one continuous performance
• Assembly, timing, and finishing in Premiere Pro. Output: the five produced videos in this Library
• On the wider bench, cast per job: dedicated prompt skills for VEO 3 and Sora, plus Gemini and Higgsfield

• Awareness-level specific: stories for unaware, mechanism education for solution aware, comparison for product aware
• Length follows the belief gap. The right runtime is however long it takes to move this specific audience from what they currently believe to the belief that makes them buy. A product-aware viewer who already accepts the mechanism needs 30 seconds. A cold viewer who has to be walked through a new explanation, shown proof, and handed an offer can hold for eight minutes, and long-form VSLs run that long profitably every day. Awareness stage, offer complexity, and price all move the number
• What stays fixed is that every second earns the next one. No filler, no preamble, no recap. A script gets cut for slack, never for hitting an arbitrary ceiling
• Two to three hook variants per script for testing, each with its own rationale
• Customer language integrated straight from the verbatim bank
• Format follows emotional register: heavy emotional scripts run UGC confessional, mechanism scripts run educational DR, humor runs skit. No format by default

• Comparison charts, one of the highest-performing static families in DTC health, because they let the viewer self-diagnose without reading a word
• Belief-disruption visuals that attack a category assumption in a single frame
• Native-format statics: iMessage threads, reviews turned into ads, formats that pretend not to be advertising
• Testimonial and social-proof formats built on specific numbers, never vague claims
• Benefit-driven product shots reserved for warm and most-aware retargeting

• Long-form education-to-conversion arc, story-driven: personal transformation or investigation narrative
• Scientific proof layered through the piece, arriving where the reader starts to doubt
• Objection handling woven in as the story unfolds, answered at the moment the doubt lands
• Two registers built and tested: journalist investigation and first-person confession, because authority and vulnerability convert different readers

AI generates volume → I refine for strategy → AI iterates on feedback → I approve final assets.

Ten times the output velocity of traditional copywriting, with strategic coherence intact, because every loop runs inside the constraint set defined upstream. The loop is also where compounding starts: every edit teaches the next generation pass.

Creative ideation compounds when it stays pointed at one offer. Every pass through the same market sharpens the constraint set, and the second campaign for a brand is meaningfully better than the first for reasons that have nothing to do with effort.

Four things accumulate:

• The constraint-based ideation framework gets sharper with each project, because every kill teaches the filter something
• The customer language database grows, so the verbatim bank starts each new brief further along
• The competitive intelligence deepens, and a category audited twice reveals patterns a single pass misses
• The skills themselves get refined. Every edit I make to an AI output is a rule I can encode, which means the next generation pass starts from where the last one ended

This portfolio was built over focused days to months collectively. The compounding visible in it is compounding against my own judgment, with no market feedback in the loop.

Everything in this Library is spec. It was built without a client, without a budget, and without a single impression served, which means the strategy has been pressure-tested against my own reasoning and against nothing else.

I am stating that plainly because the alternative is letting a portfolio imply experience it does not have, and that gets discovered in the first serious conversation anyway.

What I have is the part that transfers: a research process that goes deeper than the brief asks for, a production system that turns one research foundation into 20 to 30 coherent assets, an analytical framework I can defend decision by decision, and roughly 90 more angles and scripts behind what is shown here.

What I am missing is the only thing that cannot be self-taught. Real spend, real data, and someone senior enough to tell me which of my instincts are wrong.

I want to work under someone better than me at this, in an environment where the market gets to be the judge instead of my own conviction.

That means a few specific things:

• Feedback with the varnish off. The fastest way to improve is being told which call was wrong and why, and I would rather hear it in a review than learn it from a dead test. I hold opinions strongly and drop them cheaply, because defending a position in this discipline is billed to someone else
• Reps. Intuition here is a function of at-bats. Most of my ads will not work, and the only way through that is volume of attempts with someone watching
• Real data in the loop. Every framework in this portfolio is currently theory I trust. Contact with spend is what converts trust into knowledge, and I expect a meaningful share of what I believe today to be revised inside six months

I am early, and I am not confused about that. What I am is unusually stocked for it, and moving quickly. Give me the context and the correction and I compound fast, because the systems are already built and the only missing input is the environment.