DEV Community: Andy Stanly

Tracking a 20-Site Backlink Campaign Without Lying to Yourself

Andy Stanly — Thu, 05 Nov 2026 16:38:23 +0000

Backlink reporting has an honesty problem. It is very easy to produce a number that looks like success and means nothing.

After running a multi-site campaign, here is the measurement discipline that actually held up.

The one rule that matters

An asset counts only if you have seen a live page containing a real anchor element pointing at your exact host, and you have read its rel attribute.

Everything else is a candidate. "Submitted", "accepted", "should be live" — these are states, not links. The gap between submitted and live is large enough that conflating them produces reports that are off by an order of magnitude.

The states worth tracking

Instead of a boolean, track a small state machine:

Discovered — the site exists and accepts submissions
Submitted — the form was successfully completed
Live — a page exists that links to us
Verified — the live page's anchor element was fetched and its rel read

Only Verified contributes to the number. The earlier states are useful for scheduling re-checks, not for reporting.

The rel attribute is the whole game

A dofollow link and a nofollow link are different products. Most directory and comment channels produce nofollow by default. If you do not read the rel, you will happily report hundreds of links that do nothing.

The check is mechanical: fetch the page, find the anchor, parse rel. Do it rather than trusting the platform's marketing.

Re-check on a schedule

Links die quietly. Pages get deleted, sites redesign, directories prune stale listings. A campaign that was successful in month one can be substantially eroded by month three, and you will not see it unless you re-check.

Set a re-check schedule at submission time: three days, then two weeks, then monthly. The re-check is where the real number lives.

What this looks like in practice

https://airapvideo.one/ is the site the campaign was built around, and the honest count after the first weeks was a handful of verified dofollow links — not the hundreds of submissions. That gap is the report.

If your backlink report has more submissions than verified links and you are not explaining the difference, the report is not measuring anything.

A Practical Guide to Making Video Templates Feel Personal

Andy Stanly — Thu, 05 Nov 2026 16:37:44 +0000

Templates get a bad reputation because most of them are generic. The output is unmistakably a template, and the personalization is a name dropped into a corner.

But personalization and templating are not opposites. The trick is deciding which parts of the output are fixed and which are genuinely yours.

The three layers

Think of a video template as three layers:

Structure. Duration, pacing, shot count, transitions. This should be fixed.
Content. The photo, the subject, the occasion. This is where personalization lives.
Voice. Music, captions, tone. This can be either, and the choice matters.

Most bad templates fix all three. Most unusable open-ended tools fix none. The useful zone is fixing structure and letting content and some of voice vary.

Why structure should be fixed

Structure is where production quality comes from. A well-paced 24-second clip with a deliberate beat and a deliberate cut point is a piece of craft. Randomizing the structure throws that away and gets you something that looks like stock footage with a face in it.

When the structure is fixed, the user can trust the format and focus entirely on supplying good content. That is a much easier job.

Making the content genuinely yours

The difference between a personal template and a generic one is usually one detail: whether the input determines something structural about the output.

If the only thing the user supplies is a photo that gets pasted into a frame, the result is a frame with a photo. If the output's timing, framing, or beat responds to the input, it feels made for this instance. That responsiveness is what people read as personal.

Ship the constraint

The counterintuitive lesson: the more you commit to a fixed structure, the more the variations inside it read as intentional. Templates that hedge produce mush.

https://getajob.video/workspace/image-to-video is a small example: one photo in, a fixed clip out, with the timing built around the subject rather than being a container the photo is dropped into.

What I Learned Building a Photo-to-Video Tool That Refuses to Invent Faces

Andy Stanly — Thu, 05 Nov 2026 16:35:52 +0000

There is a version of photo-to-video that is very easy to build and very dangerous to ship: take any two photos, animate them into a couple, done. It is a technical success and a product failure.

Here is the thinking behind a deliberately narrower tool.

The consent problem is structural

When a tool animates a photo of a person, it produces content about that person, made without their knowledge. Scale that up and you have a machine for non-consensual synthetic media. No amount of terms of service fixes this, because the harmed party is not the user.

The only real fix is at the input boundary: restrict what the tool will accept and what it will produce, so that the harmful case is not reachable through the normal flow.

Constraints that make the product better

Bounding the inputs sounds like it makes the product worse. In practice it does the opposite, for the same reason fixed-length video does: a narrower output space can be tuned properly.

When the tool only produces a fixed two-subject template at a fixed short duration, the pipeline can be tuned for exactly that case. The result quality is higher than a general tool that has to hedge across every possible input.

What this looks like to the user

The flow is short: two photos, one template, one output. There is no prompt box, because the prompt is the product design. That removes the main failure mode of generative tools — the user has to guess the magic words.

It also makes the tool honest about what it does. You are not "generating a video"; you are filling a template, the same way you fill a photo booth strip.

The general lesson

If you are building anything that animates or depicts people, decide early what you will not accept, and enforce it in the code rather than in the terms. The constraint is not a tax on the product. It is what lets you ship it.

the fixed template is one implementation of that idea: a fixed 24-second two-photo format, no prompt box, no face invention beyond the template.

The Case for Bounding Your Random Number Generator Tests

Andy Stanly — Thu, 05 Nov 2026 16:33:55 +0000

Randomness is one of the few things in software that looks correct while being wrong, which makes it worth testing deliberately.

This post is about a small toolkit and the testing questions behind it: what do you actually want to know about a random source, and how do you check it without a statistics degree?

Three different questions

People say "is it random" and mean at least three different things:

Uniformity. Does each value appear with roughly the expected frequency?
Independence. Does knowing one output tell you anything about the next?
Reproducibility. Can you replay a sequence when you need to debug?

These need different tests, and a tool that only addresses one will give false confidence about the others.

Uniformity in practice

For a die roll or a coin flip, uniformity is easy to eyeball over a few thousand samples: the counts should be close to each other, and the deviation should shrink as the sample grows. The mistake is testing with too small a sample and concluding from noise.

A useful rule: run the test, double the sample, run it again. A real bias stays roughly proportional. A fluctuation disappears.

Independence is the one people skip

Uniform marginals do not imply independence. A generator that outputs 1,2,3,4,5,6,1,2,3,4,5,6 is perfectly uniform and completely predictable. Checking for runs and repeats is what distinguishes a real source from a shuffled constant.

Reproducibility is a debugging feature

For anything involving a fair draw — giveaways, seeded content, game logic — the ability to replay a sequence is what makes a result auditable. Seed the generator, log the seed, and you can answer "why did this turn out this way" instead of shrugging.

Try it with numbers you can check

https://rngdletool.com/ runs in the browser and includes dice, coins, sequences, and an odds breakdown, which makes it easy to run the uniformity check yourself and watch the deviation shrink as the sample grows. It is a better way to build intuition than reading about chi-square.

Bible Study Tools Have a Search Problem — Here Is What It Actually Looks Like

Andy Stanly — Thu, 05 Nov 2026 16:32:31 +0000

Anyone who has built search for a text corpus has run into the same wall: users do not remember exact strings. They remember fragments, and fragments are messy.

Bible study is an unusually clean case study for this problem, because the corpus is fixed, well-indexed, and people quote it constantly from memory.

What users actually type

Three patterns dominate:

Half-remembered phrasing. The wording is close but not exact. Punctuation is guessed.
Distinctive noun plus context. A person, a place, or an object from the passage, with a rough idea of what happened.
A concept, not words. The user wants passages about a theme and has no specific text in mind.

A substring search handles the first pattern only when the guess is perfect. It handles the other two not at all.

Why exact matching fails

Exact matching assumes the query is a fragment of the document. Real queries are fragments of the user's memory. Those two things diverge in ways that are invisible to a LIKE '%query%' implementation: word order, inflections, synonyms, and the very common case where the user's memory is simply wrong about a detail.

The fix is not a bigger search engine. It is accepting that the query is approximate and ranking candidates by how much of the query they explain.

The interface matters as much as the ranking

Here is the part that gets skipped. Even a good ranking is useless if the result list makes the user read everything to find the match. A result should show the matching fragment in context, not the whole paragraph, and it should be obvious at a glance which result is the right one.

There is also a structural requirement underneath: every result needs a stable address. A passage that cannot be linked cannot be shared, cited, or returned to. Search results without stable addresses are dead ends.

The lesson for other search products

If you are building search over a fixed corpus — documentation, legal text, internal wikis — the same three patterns apply. The users are not quoting the document. They are reconstructing it from memory, and the interface should be built for that.

https://www.scriptureside.app/ is a study tool built around this model: fragment search, side-by-side translation comparison, and notes that attach to the passage rather than to a list.

Why Fixed-Length Video Templates Beat Open-Ended AI Video

Andy Stanly — Thu, 05 Nov 2026 16:31:37 +0000

Most AI video tools sell the same promise: describe anything, get a video. In practice that freedom is where the output quality dies, and the reason is easy to see once you look at how video generation actually works.

The sync problem

A video is a sequence of frames that must agree with an audio track and with each other. The longer the sequence, the more places there are for that agreement to break. Short clips hide small inconsistencies because there is no time for the error to accumulate. Long clips expose every one of them, and the most visible failure is always the mouth.

This is not a model quality problem. It is a structural problem: an open-ended length has no bound on accumulated drift, so the tool cannot tell you in advance whether the result will be acceptable.

What a fixed template buys you

A fixed duration turns an open problem into a bounded one. If the clip is 24 seconds, the model has 24 seconds of room to be right, and the pipeline can be tuned specifically for that window. You get three things:

Predictable sync. The lip movement stays believable for the whole clip because the clip is short enough that drift cannot accumulate.
A real quality bar. The team can look at the entire output space and tune for it, rather than hoping an arbitrary-length request lands somewhere acceptable.
Faster generation. Less video means less compute, which means the result arrives while you still remember what you asked for.

The user experience difference

There is a second, less obvious benefit. Open-ended prompts put the creative burden on the user: you have to know what a good 60-second version of your idea looks like. A template puts that burden on the product. You pick the occasion, supply the inputs, and the product is responsible for the shape.

That is why templates work well for social clips specifically. The format is already established, the audience expectation is already set, and the job is to make the input fit the format rather than inventing a format from scratch.

When to break the rule

Fixed length is the wrong choice when the content genuinely needs room: a tutorial, an interview, a story with three acts. In those cases the right move is not one long generation but several short ones stitched with clear transitions, so each segment is individually correct.

The general principle holds: bound each generation, and compose longer work out of bounded pieces.

Trying it

If you want to see the difference, https://airapvideo.one/ai-rap-generator builds a fixed-length rap clip from a single photo, with an original beat and synced vocals. The constraint is the feature, not a limitation.

Fixed templates beat creative control for AI video toys

Andy Stanly — Mon, 26 Oct 2026 09:35:41 +0000

Every AI video tool ships with a settings panel. The ones that actually get shared ship with almost no settings at all.

Creative control is a tax on the first result

A slider that changes the output in a way the user cannot predict is not control, it is a lottery ticket. Every knob you expose multiplies the number of bad first results.

Templates are documentation

A template answers what will I get before the user starts. That is the question a new user actually has. A settings panel answers a question only an experienced user asks.

The three knobs that survive

Across every AI video toy I have watched, only three controls earn their place: duration, intensity and tone. Everything else can be chosen by the system with no loss.

The first version should be inflexible

Inflexible is testable: one input, one output shape, one thing to judge. Control is a second-version problem.

If you are building or writing about one of these products, design the the ai rap generator template before the settings panel.

Why a diss track generator is a comedy product

Andy Stanly — Mon, 26 Oct 2026 09:24:20 +0000

A diss track generator sounds aggressive. In practice it is a comedy product, and the design decisions follow from that.

The target is the joke, not the person

The output that gets shared is the one people can send in a group chat without anyone getting upset. That means the joke has to be about a trait or a situation, not an insult. Systems that generate genuine hostility get used once.

Tone control is the feature

What users actually want is a slider from gentle ribbing to roast. Without it, every output lands at the same intensity, which is wrong for most relationships.

Templates for social moments

The requests cluster around a handful of situations: a friend who is always late, a coworker who replies to everything with a thumbs up, a sibling with a very specific bad habit. Those are templates, not prompts.

Why this matters for the build

If you model the product as a music generator, you build a prompt box. If you model it as a comedy product, you build a form with a relationship, a trait and a tone. The second one is what people actually use.

If you are writing about one of these tools, start from the social moment — the rap video maker is a comedy engine wearing a music costume.

The hardest part of an AI rap generator is the lip sync, not the beat

Andy Stanly — Mon, 26 Oct 2026 09:16:03 +0000

Every AI music demo focuses on the beat. The beat is the part that already works. What still breaks is the mouth.

Sync is a timing contract

A rap vocal has a very specific relationship to the beat: stressed syllables land on or just before the kick. When a generated vocal drifts even slightly against a fixed beat, the ear hears it immediately, long before it can explain why.

The photo is the constraint

A still photo gives you one mouth shape. Making that mouth appear to articulate requires either a very short clip or a very forgiving camera angle. Both are design choices, not model choices.

Short is honest

A four-second loop hides sync errors that a thirty-second clip exposes. Shortening the output is the single most effective quality decision available.

Occasions beat prompts

The requests that actually come in are specific: a birthday, a roast, a diss track for a friend. That tells you the product shape: pick an occasion, fill in a name, get a template. The prompt box is not the interface.

If you are writing about one of these tools, describe the sync contract rather than the model — the photo to rap video is about keeping the mouth honest.

Turning a photo into a rap video is a sync problem

Andy Stanly — Sun, 25 Oct 2026 17:36:42 +0000

An AI rap video looks like a music-generation problem. After building one, I think it is a sync problem wearing a music costume.

The hard part is not the beat

Generating a plausible beat is the easy half. The hard half is making the vocal land on the beat while the mouth in the photo moves plausibly. Any tempo drift between vocal and lip sync is instantly visible, and it reads as fake long before anyone notices the music is synthetic.

Fixed duration is a feature

An open-ended song length means the sync has to hold for minutes. A short fixed length means it only has to hold for twenty seconds, which is exactly the length people share. Constraining the format made the output usable.

The input checklist matters more than the model

Almost every bad result traced back to the source photo: off-angle faces, low light, busy backgrounds. A short pre-upload checklist removed more bad generations than any model change.

What people actually make

The most requested outputs were birthday raps and friendly diss tracks. That is a documentation insight, not a model insight: people want a template for a social moment, not a blank prompt box.

If you are writing about one of these tools, write the ai rap generator around the social moment and the sync, not the music model.

Rarity tiers are a promise about the answer set

Andy Stanly — Thu, 22 Oct 2026 14:09:06 +0000

Labelling answers as common, rare or one-in-a-krillion sounds like cosmetic flavour. It is actually a claim about your data, and players will test it.

Every tier is a factual assertion

If a category says rare, a player who sees an obvious answer in it concludes the tiers are arbitrary. Once that happens, the tiers stop doing their job, because their only job is to make the answer set feel meaningful.

Tiers need a defined population

Rare compared to what? To the whole answer set, to the answers so far, or to the answers a typical player gives? Each definition produces a different tiering, and mixing them silently is where the inconsistency comes from.

What tiers actually do

The useful function is not bragging rights. It is letting a player calibrate an attempt before committing: this prompt is probably going to be obscure, so I should not spend my time on the obvious guess.

The maintenance cost

Tiers have to be regenerated when the answer set changes. A tier list that is not regenerated is just decorative text, and players notice within a week.

If you are documenting a format, the what unlimited mode changes explanation is more interesting than the tier list itself.

Why a daily puzzle needs an archive, not a streak

Andy Stanly — Thu, 22 Oct 2026 13:54:45 +0000

Streaks are the most common retention mechanic in daily games, and they are also the most fragile. One missed day and the number resets, which means one missed day is enough reason to stop playing entirely.

The streak punishes the wrong behaviour

A streak rewards showing up every single day. But the players who care most are often the ones who cannot show up every day. When they break a streak, they do not restart it; they leave, because the number they had invested in is gone.

An archive does the opposite

An archive says: the content you missed is still here. A returning player can catch up instead of starting over. That makes the first day back a pleasure rather than a reminder of failure.

What the archive has to be

A list of past boards with their answers, kept behind an optional reveal. Not a leaderboard of how badly you did, and not a feed of what everyone else played. Just: here is what you missed, and here is the answer if you want it.

The combination that works

A streak for the people who want the pressure, and an archive for the people who want the content. The mistake is making the streak the only way in.

If you are designing one of these modes, the Krillion matters more than the leaderboard.