Festivals and moments coming up, with a character ready for each.
We read the product name, description and price from the page. Or type them below.
Castly's AI makes the reels. You set the rules. Brands rent your character by category, market and quarter, and you take 60% of every fee.
Send the source, a few reference pictures and your rules. Pick the categories that stay closed.
We write how your character looks, speaks and never behaves. Nothing is made or sold until you approve it.
Brands lock your character for a category. Every reel is checked against your rules before it ships, with a plain version to prove it works.
60% of every lock and licence fee, on a quarterly statement you can check against the ledger.
The owner licenses behaviour, not pixels. This spec is enforced rule by rule on every generation. Draft it from canon, stress-test it on awkward briefs, then save. Weights and source art never leave the vault.
Five awkward briefs run against the spec as written, saved or not. Each rule is a yes-or-no check.
Ads that failed a rule, or every ad when the owner chose review mode.
If every brand can use Sherlock, Sherlock stops working for anyone. Nobody can own Sherlock, but a brand can own Sherlock-in-Indian-tea for a quarter on the platform where production, clearance and proof live. For a public-domain character the lock is Castly's promise not to sell its rendition to a rival in that cell; it cannot stop a rival drawing its own. Owner-licensed characters carry exclusivity the owner grants.
Cells marked example are seeded to show the mechanic.
Every campaign goes out as character ads plus a no-character holdout, each with its own link. Castly counts visits and purchases itself, so the lift is yours to check.
Every step from first visit to paid order, counted on the server without cookies or personal data, plus pilot requests and sign-ups.
A model can check rules, overlaps and sameness; it can't be trusted to judge whether an ad is good. Thumbs can. Every campaign ships with a matched holdout: same offer, same format, no character. The gap is the character's lift, and across campaigns it builds the Character Index.
Each figure was read at its source on 17 Nov 2026. The caveat under each one matters as much as the number.
The best-remembered ads were 6.01 times more likely to feature a character. Using a celebrity consistently scored 2.84 times. Characters appeared in only 5% of ads.
Ipsos, 2020: 2,015 US video ads, top versus bottom third on branded attention. About brand-owned characters; a pre-launch memory measure, not sales.
Ipsos, The Power of YouLong-term campaigns with a recurring character reported a market share gain more often (39.2% vs 29.7%), and a profit gain more often (34.2% vs 26.2%).
Orlando Wood, Lemon (IPA, 2019), from the IPA Effectiveness Databank. Read in a secondary source; to be checked against the book before we quote it outside. Brand-owned characters, award entrants.
The Grocer, citing LemonThe share of long-term campaigns using a character or slogan device fell from 41% in 1992 to 12%. Most brands stopped using the thing that works.
System1 and the IPA, 492 long-term campaigns. Two secondary sources agree; the primary deck was not opened.
The Grocer, citing System1A famous face alone gives no average lift in long-term brand building. Fit is what matters, which is why Castly checks each character against a spec and tests against a holdout.
Kantar LINK database, June 2025. Roughly a third of Indian TV ads already carry a celebrity (TAM AdEx, H1 2024).
Kantarsay they would be less likely to buy from a brand that uses AI-generated ads; 73% would trust such an ad less. People spot generic AI and discount it.
Harris Poll with 4As and Infillion, June 2026: stated attitudes, not behaviour. NielsenIQ (2024, 2,000+ people) found AI ads leave weaker memory traces. This applies to Castly's output too: it is why every ad is labelled and built on a character people already know.
Marketing Brew NielsenIQChildren aged 4–6 said the same food tasted better when a licensed character was on the pack. Characters are powerful with children, so Castly does not sell junk food to them and addresses parents, not kids.
Roberto, Baik, Harris and Brownell, Pediatrics, 2010. 40 children; packaging, not ads.
Pediatrics 126(1)From docs/revenue.md. The pilot win rate is the number the model rests on.
| Test | Pass | Fail, and what it means |
|---|---|---|
| Company keys added, checkout and invoicing on, plans published on the site | Live in 2 weeks | Not live: nothing else can be tested |
| Pilots started, direct and through 3 performance agencies | 10 pilots | Under 5: the offer is not reaching buyers; fix distribution before price |
| Pilot win rate under the signed rule | Character beats plain in 40% or more of completed pilots | Under 25%: keep the plans, replace the CPA guarantee with an attention or recall measure |
| Pilots and sign-ups converting to a paid plan | 5 paying brands across the tiers | None: the price or the tier contents are wrong; test Starter at a lower price first |
| One owner signed at 60%; one licensed lock offered | One brand pays ₹2 lakh for it | No premium over public domain: hold the licensed tier and learn why |
Fills from reported results. Empty until pilots report; no numbers are invented here.
AI makes production free, so production stops differentiating, and left alone every model converges on the same few looks and lines. What stays scarce is an audience's existing affection for a character, the right to use it safely, and proof that it sells. Castly sells all three as one product.
Rendition, voice, never-rules, protected looks to avoid, blocked categories, approval mode. Drafted from canon, stress-tested on awkward briefs.
Clarifying questions, eight territories, then finished ads: rendered reels and posts in the style and for the audience the brief needs, from 3D animation for kids to ink engraving for adults 55+.
Character × category × market × window. A quarterly fee per lock on top of the monthly plan; 60% of every lock and licence fee to the owner on licensed characters. Plain invoices.
A no-character control ships with every campaign. Real-world lift is the only taste judge that scales, and the Index it builds is the moat.
| Part | On spark-pc | Cost and cap |
|---|---|---|
| Page and API | Node server in C:\roster, reached on the tailnet only | Negligible |
| Character art and ad images | The ComfyUI already running on the GPU, Z-Image Turbo, 8 steps | About 15–25 s of GPU per image, one at a time, cached; capped at 60 an hour. A finished reel campaign is about 13 images. |
| Copy (questions, territories, ads, checks) | An OpenAI-compatible text model; by default the newsroom's own qwen2.5 14B on the CPU | Shares the CPU with the Fourth Estate editor; capped at 20 calls an hour, one at a time. A reel script is 1 call, streamed as it is written (about a minute); the advanced brief is 3. |
| Video reels | MiniMax H3 through the same ComfyUI: each shot is moving picture and the character's own voice in one pass, with the licensed art as the reference so the face holds; ACE-Step music beds; cut, captioned and mixed by ffmpeg | About 3 to 5 minutes of GPU per 5-second shot; a reel and its holdout are 5 shots. Capped at 10 shots an hour and 4 reels per company per day, one shot at a time |
| Accounts, campaigns, tracking | Work-email sign-in, orders by invoice, a launch pack per campaign, and a first-party script that counts visits and purchases per ad | Negligible. Razorpay checkout and one-click Meta publishing are built and switch on when the company's keys are added; email delivery waits on a sending domain |
| Saved state | Specs, locks, ledger and results stay in each browser for now | Moves to the server when pilots start |
| Lesson | Change in Castly |
|---|---|
| You get out what you put in; brand context is the durable asset | Brand kit plus Character Spec go into every prompt, whatever model runs underneath |
| Ask for ten options, mix, go again | Eight territories first; develop up to three or ask for more like the ones picked |
| People love being asked questions up front | Every brief starts with clarifying questions; answer rate is a pilot metric |
| Tell the model its failure modes | Known AI-ad clichés are named in every prompt, with a slop check and a sameness check after |
| Yes/no beats 1–5; models can't grade taste | Each owner rule is a yes/no check with a reason; taste is left to people and the holdout |
| Point at the problem | Notes on one ad, then revise that ad alone |
| People see one thing at a time and need time to read | Reels play one shot and one line per beat; lines too long for their beat are flagged |
| Build skills from real examples, test on boring prompts | Draft a spec from canon, stress-test it on five awkward briefs |
| Models collapse to one look | Every character has a native style (cinematic, 3D, pixel art, comic, claymation, watercolour, painterly, engraving) and a target audience; both can be changed per brief and steer the copy, the checks and the art |
| Build standalone first, then merge in | Castly starts as its own studio; the end state is a connector agents call mid-campaign |
Clearance notes are a starting point for counsel, not legal advice. Character art is Castly's own rendition, made for this prototype.