The Shot You Cannot Miss: Practice for Reliability, Not One Success
One clean run proves a skill is possible, not reliable. How to practice for reliability: measure your success rate under pressure, fatigue, and variation.
How many clean runs does it take before a skill is yours? A presentation, a tricky deploy, a free throw, a hard conversation. Most people stop practicing the first time it goes well, and most people are surprised when it falls apart on the day.
The honest answer is that nobody has a number. No study gives a count of repetitions after which a skill becomes reliable, because the answer depends on the skill, the person, and what is riding on it. What we do have is some plain statistics, a few decades of learning research, and the way engineers already treat software they cannot afford to have fail. Together they point one way. To practice for reliability, stop aiming at the first success and start measuring your error rate under the conditions you will actually face.
One demo, six rehearsals, one clean run
An engineer has eight minutes at a company all-hands to demo a feature her team spent a quarter building. She rehearses at her desk the afternoon before. (This is an illustrative composite, not a measured case.)
On the day, the question at minute four knocked her off the script. She jumped ahead to recover time, skipped the slide the rest of the demo depended on, and the live demo stalled on the venue network while she filled the silence. Nothing went wrong that had not already gone wrong in rehearsal. Five of her six runs had failed in some way. She had treated the sixth as the truth and the first five as warm-up.
That is the mistake this post is about. She did not under-practice in hours. She stopped measuring at exactly the moment the measurement started to matter.
The model: one success is a sample of one
Here is the model in one sentence: a skill is a success rate under a set of conditions, and a single success tells you almost nothing about that rate.
The arithmetic is old and unglamorous. If you have done something once and it worked, the data cannot rule out that you would succeed only one time in twenty. Ten clean runs in a row still leave room for a skill that fails one time in four. Statisticians compress this into the rule of three: after zero failures in n attempts, the failure rate is probably below 3 divided by n. To claim a 1 percent miss rate with any confidence, you need about three hundred clean attempts.
So the first success answers one question: is this possible for me? That is worth knowing. It is a different question from the one the day will ask, which is how often it works when the room is loud and you are tired. Four findings follow from taking the second question seriously.
Finding 1: engineers already distrust a single pass
Software teams learned this the expensive way. A flaky test is one that passes and fails on the same code, and Google’s testing team wrote in 2016 that about 1.5 percent of all its test runs produced a flaky result, with almost 16 percent of its tests showing some flakiness. Nobody at that scale trusts a single green check. They rerun suspicious tests many times and track each test’s pass rate over history.
The same instinct separates a demo from production. A demo is one success, on a known path, with the presenter’s hands on the keyboard. Production is the same code run millions of times by people who click in the wrong order. Reliability engineering exists because the second number is the one users feel. A service is described by its success rate over a window, never by the fact that it once answered a request.
Your skills deserve the same suspicion. A clean rehearsal is a green run. The interesting question is what the other 199 look like.
Finding 2: practice past the first success holds, then fades
Psychologists call practice beyond the first success overlearning. Driskell, Willis, and Copper pooled the studies on it in a 1992 meta-analysis in the Journal of Applied Psychology and found that it does improve retention, with the size of the benefit depending on how much extra practice was done, the kind of task, and how long people waited before being tested. For cognitive tasks, the benefit shrank as that wait grew. Physical tasks held up better.
Doug Rohrer and colleagues tested the decay directly in a 2005 paper in Applied Cognitive Psychology. Students who kept drilling geography facts or word definitions past the point of first getting them were far ahead at one week. At longer delays the advantage shrank sharply.
This is the honest limit on “just do more reps.” Extra repetitions crammed into one session are a loan against a few weeks. If the event is three months out, reliability comes from spacing the work: practice past the first success, then come back to it again and again, so the rate is high on the day and not only on the day you trained.
Finding 3: practice that varies looks worse and holds up better
John Shea and Robyn Morgan ran the classic version of this in 1979, in the Journal of Experimental Psychology: Human Learning and Memory. People learned three quick arm-movement patterns, either in blocks (all of one, then all of the next) or in a random order. The blocked group improved faster during practice. The random group improved more slowly, and by the end of practice the two were close. On the later retention test, and on a transfer test with a new and harder pattern, the random group did better.
The finding has a name, the contextual interference effect, and a practical sting: the practice that feels best is often the one that predicts the day worst. Blocked repetition gives you a run of successes that are partly borrowed from the previous attempt. Mixed practice makes every attempt start cold, which is what the real attempt will do.
For the demo engineer, mixing means rehearsing from a random slide, with the network off, with someone interrupting at a time she did not pick. Her rate will drop at first. That drop is information about the day.
Finding 4: pressure is a condition you can train in
Sian Beilock and Thomas Carr studied choking in a 2001 paper in the Journal of Experimental Psychology: General. Pressure hurt golf putting, a skill that runs largely on autopilot once learned, while it did not hurt a step-by-step arithmetic task. Their reading was that pressure makes people watch the mechanics of an automated movement, and watching it breaks it.
The training result is the useful part. In one experiment they trained 54 novice golfers in three ways: plain putting, putting with a distracting second task, or putting in front of a camera after being told golf pros would review the recordings. Under a later pressure test, the first two groups got worse. The group that had trained while self-conscious improved.
That is one lab study with novices and a single skill, so hold it loosely. It does say something specific: the feeling of being watched is a condition, and a condition can be part of practice. A rate measured in private describes private performance.
Finding 5: some steps should not depend on practice at all
The last finding is the one practice alone cannot reach. Alex Haynes and colleagues reported in the New England Journal of Medicine in 2009 on a 19-item surgical safety checklist introduced in eight hospitals in eight cities. Complications fell from 11.0 percent to 7.0 percent, and inpatient deaths fell from 1.5 percent to 0.8 percent. It was a before-and-after study rather than a randomized trial, so some of the change may have other causes. The surgeons did not learn new surgery. They stopped trusting memory for steps like confirming the patient and the site.
Aviation runs on the same idea. Highly practiced pilots still run written checklists, because a skilled person’s miss rate on a routine step is low but not zero, and some steps cannot carry any miss rate. Practice drives the rate down. A checklist covers what is left.
What this gets wrong
A success rate is only as honest as the conditions behind it. Ten clean runs at a desk say little about a stage, and you cannot rehearse every condition the day will throw at you. The model asks you to guess the important ones and add them on purpose, and a bad guess produces a confident number about the wrong thing.
Repetition also locks in whatever you repeat. A high rate on a flawed technique makes the flaw harder to remove later, which is why the deliberate practice question of what you are practicing comes before the question of how often it works.
The cost curve is steep too. Going from rarely succeeding to usually succeeding is cheap compared with going from 90 percent to 99. Most skills do not need the last points. A weekly status update can miss sometimes. The step that deletes a production database cannot. Part of the work is deciding which is which, the same way good judgment is a set of checks applied where the cost of a miss is high.
And the research is narrower than this post’s reach. The putting study used novices in a lab, the overlearning studies used facts and word lists, and the checklist study was not randomized. Applying them to a sales pitch or an on-call rotation is reasonable extrapolation, not a measured result.
Practice for reliability: count the misses
Start keeping score. Most skills will tolerate a few misses, and you only find out which ones by counting.
- Keep a tally. A spreadsheet with four columns is enough: date, condition, attempt, clean or miss. Compute the rate over your last twenty attempts and watch that number, not your best run.
- Record and review the misses. Set a phone on a shelf and film the rehearsal or the drill. Watch only the attempts that failed, and write one line on what broke.
- Space and mix. For knowledge, a spaced-repetition app such as Anki schedules each card at growing intervals so the review lands just before you would forget. For performance skills, practice in a shuffled order and return to the skill across weeks, not in one long session.
- Add pressure on purpose. A countdown timer, one colleague watching, an attempt at the end of a long day. Each is a condition the real attempt will probably have.
- For engineers, run it many times. Rerun a suspicious test a few hundred times before trusting it (Go’s test runner takes a count flag, and Bazel has a runs-per-test option). Load test with tools like k6 or Locust. Chaos engineering, which Netflix popularized with Chaos Monkey randomly terminating production instances, measures how often the system survives the failures you would rather not wait for. Rehearse incident runbooks in a game day, so the first time someone follows the runbook is not at 3 a.m.
- Write the checklist. For the steps that must never fail, take them off memory entirely.
The engineer with the demo did not need more confidence. She needed a count. Seven clean runs out of ten, under a timer with a colleague interrupting, would have told her the truth the day told her anyway, two days early.
