Writing Plans

One step teaches one thing, and changes one thing from the step before it.

That is the whole discipline, and it is harder than it sounds, because the finished behavior is obvious to the person writing and the intermediate versions are not. See Splitting and Lumping.

A step needs four pieces of text, and each one has a job:

• Setup — where the dog and the handler start. A step that does not say this gets run five different ways.
• Details — what the handler actually does. Short. This is read on a phone, one-handed, with a dog waiting.
• Pass — what success looks like, observably.
• Fail — what it looks like when it does not happen, observably.

The commonest authoring error is a step that changes two things at once: a longer duration and the handler moving away, a new room and a new cue. When your dog fails it, nobody can tell which half was too hard, and neither can the engine.

The second commonest is a step that is really a set of instructions for a whole training session. If your details field needs a paragraph, you have written three steps.

The pass and fail fields are the measuring instrument. Everything the engine does is computed from which one the handler tapped, so they carry more weight than any other text in the plan.

Two rules, and the first is not negotiable.

Fail is a measurement, never an instruction. Tapping fail means nothing happens to the dog: no treat, reset, go again. If a fail field tells the handler to act on the dog — walk them back, take something away, say "ah-ah", apply any consequence at all — the step is designed wrong, however mild the consequence sounds. Managing the situation inside a fail branch is different and fine: covering the food, keeping the leash short so the dog cannot get to the thing, stopping, so that pulling does not get your dog to the thing they were pulling towards. Preventing access is not applying a cost.

Both must be observable by a stranger. "Your dog ignores you" is not a fail criterion, it is a theory about a dog who may not have heard, may be sniffing something enormous, or may simply be slower than you expected. Write "your dog does not come to you." The fail should mirror the pass, and both should be countable by someone watching who has never met the dog.

Get these two right and an inexperienced handler can judge honestly. Get them wrong and every number the app produces afterwards is decorative.

Each step is judged by one of five sets of rules, and the choice is a claim about what could go wrong.

• Prompting — the default, and right for most steps. You ask, the dog does or does not, you score it.
• Shaping — timed, and it tracks two criteria at once: the version your dog can already do, and the better version you are watching for them to offer. Most often that is a verbal cue being added in front of a known hand signal, but the same rules fit any shaping ladder where you are paying the current approximation and shopping for the next one. See Jumping the Prompt.
• Capturing — for behaviors nobody can prompt: a yawn, a shake off, actually relaxing. No fail button, and it can only ever push.
• Classical — for how the dog feels rather than what they did. One fail drops it.
• Errorless — for steps arranged so failing is nearly impossible. Timed, one fail drops it, and the fail button is disabled until the timer ends.

The choice that matters most: anything to do with fear, comfort, grooming, handling or a dog who has learned to tolerate something should be Classical or Errorless. Those two back off after a single bad rep. Writing a fear step as Prompting means a dog can fail twice and keep going, and by the second one you are usually past the point where the work was doing any good. See Threshold.

The mirror-image error is writing an ordinary obedience step as Classical because it feels gentler. It is not gentler, it is just twitchy: a single unlucky rep sends a dog backwards down a staircase they were climbing fine.

One practical limit on Shaping as it stands. On these steps the Player replaces the check and the X with a verbal cue icon and a hand signal icon, because adding a word is what they are mostly used for. That is exactly right for cue work and slightly wrong for anything else. On any other kind of shaping ladder the judging is still correct and the buttons are simply asking the question in the wrong words. Better labels are on the list.

Two kinds of step sit off the main path, in opposite directions.

A split step is easier than the step it sits in front of, and hidden. Dogs only ever see one if they need it: the engine reveals it after a second drop in a row on the same step. Write them for the rungs you expect to be too tall for somebody — the first time the food leaves the luring hand, the first repetition with the handler out of the room, the first one outdoors. They cost nothing when nobody needs them.

A proof step is harder, and visible. It is not part of the first climb; it is replayed later, when the dog already has the behavior and you want it to survive a new room, a new person, a bit of noise. See Proof.

A good plan usually has more split steps than an author's instinct suggests. Splitting is cheap: an extra rung costs two minutes of somebody's evening. A dog stuck failing a rung that is too tall costs weeks, and costs the dog confidence in the whole exercise.

A goal step is a milestone. The dog earns it by pushing off that step, it turns green on the goals page, and from then on the app can call that behavior in a proofing game alongside goals from other plans.

Put one wherever something has genuinely been achieved. That is a judgement about the dog's progress, not a position in the list, and it happens in three kinds of place:

• A criterion reaches its climax. The full distance, the full duration, the real distraction. Those earn the distance, duration and distraction goals.
• A cue arrives, or changes hands. The dog responds to the word, or to the hand signal, or starts offering the behavior unasked, or begins doing it automatically in its context. Those are the verbal, hand, offered and automatic goals.
• The dog is comfortable with something they were not. That is the comfort goal, and it is the one most often written too late.

Goals are not only for the end of a plan. A comfort plan frequently earns one on step one, where the dog has developed a good feeling about the mere sight of the object. Nothing has been done to the dog yet and nothing has been handled, but a real thing has changed and it is worth recording as an achievement.

The restraint worth keeping: a goal should mean something. A plan with a goal on every step teaches its owner that goals are decoration, and a goal on a step that is not a milestone puts a green tick against a skill the dog does not have.

The numbers, and what they actually do. Every one of these has a sensible default, and most plans should leave all of them alone.

• reps — how many repetitions make a set. Defaults to 5, which suits most teaching steps: short enough that a dog does not sour, long enough to mean something. It is a default, not a rule. Fear and comfort work often wants more, because those step types drop on a single fail anyway, and a longer set is what makes a push mean the dog was genuinely settled rather than lucky. Expect this advice to sharpen as there is more data to read.
• passRatio — the share of completed reps that must pass to push. Default 0.61, rounded up, so a five rep step needs four.
• failRatio — the share that must fail to drop. Also 0.61, but rounded down, so a five rep step drops at three. The asymmetry is on purpose: easier to go back than to go forward.
• bailRatio — how much failing ends the set early. Default 0.5, so a five rep step stops once three reps have failed rather than making the dog finish.
• duration — for timed steps, in milliseconds. Sixty seconds is 60000.
• interval — for timed steps, whether every rep is the target length or the lengths vary. See Interval.
• repsPerMinute — Shaping only. With duration, it is what the expected rep count is derived from.

Move a ratio only when you can say why in a sentence. The one legitimate reason is a step where a single fail matters more or less than usual, and even then the step type is usually the better lever: if you want one fail to drop, that is what Classical and Errorless are for.

The number worth thinking about hardest is not a ratio at all, it is duration on the timed steps, because it sets the rate of reinforcement. A sixty second hold paid once a minute is thin going, and a variable interval is usually the fix.

A new or edited plan is a draft until it is committed. That is deliberate: it means a plan can be written, run once with a real dog, and fixed before anybody else ever sees it.

Committing does two things at once. It saves a named checkpoint, so a plan has a history you can read and roll back, and it publishes, which is what puts the plan on the home screen and offers the update to anyone already using it. Write the checkpoint message the way a trainer would note a change: "shortened the hold to three seconds", not "update".

Publishing to the wider world is separate again, and the rule is worth knowing before you write anything you would mind strangers reading. Until you claim a public handle, your plans are private to you and to the co-trainers of the dog they belong to. Once you claim a handle, new plans default to public, and you set individual ones back to private in the plan's settings.

A public plan is listed on your profile at pushdropstick.com/u/your-handle, is searchable inside the app, is exposed to search engines and to AI assistants through our public catalog, and can be forked by anyone. A fork is a copy that belongs to the person who made it: your later edits do not reach their copy, and theirs never touch yours. Attribution carries your handle, and carries it onward through their forks.

Two consequences that surprise people. Setting a plan private afterwards removes it from your profile and from search, but copies already forked stay with the people who forked them — you cannot call a plan back. And plans that recommend aversive equipment or methods can be hidden or removed, which is a house rule written into the Terms rather than a matter of taste.

Two things worth knowing before you publish. The website shows every step of a plan, splits included, along with the step types, treat values, rep counts and thresholds, so a public plan is genuinely readable by other trainers. And a plan is a claim about how to train a dog, with your name on it — which is the honest reason to run it with a real animal first.