Change the rules, not the prompt

In 1986 the sports scientist Karl Newell proposed that skilled movement takes the shape of three things: the performer, the environment and the task itself [1]. Change one of the three and behavior changes with it. No instruction required. Coaches spent the following decades turning that idea into a method: redesign the practice itself instead of asking a player to comply [2].

1. Cutting the corners

Thomas Tuchel took over as head coach of a Bundesliga football club in 2009, four days before the first match of the season, having never played or coached at that level before [7]. His team kept losing the ball the same way: wide, along the sideline, easy for an opponent to defend. Tuchel wanted the ball to move diagonally through the middle instead. He tried shouting the correction first, then stopped:

"I didn't want to be the one blowing the whistle on every long ball down the line, screaming: no, play diagonal! That wears thin." [7, translated]

So he changed the practice pitch instead. He had the corners cut off, shaping it into a diamond:

"We cut the corners off our training pitch. We trained on a diamond-shaped field. Why? Because diagonal, flat play was our principle. We forced the players into that principle, and inside those guardrails they kept their full creativity." [7, translated]

A wide ball down the line became harder to play. Nobody had to repeat the instruction. The space a wide ball needed no longer existed.

2. Why the field beats the whistle

Sport science research now confirms what Tuchel found on his own training pitch. A 2020 review of coaching feedback names constraint manipulation as the coach's main tool. Detailed verbal instructions can narrow what a performer notices and tries [3]. Controlled and professional-level studies back this up. Ten-year-old girls trained through changed equipment and rules found more different ways to hit a working tennis stroke than a group drilled through repetition [4]. Elite Australian-rules football players shifted measurable tactical behavior the moment a single numerical rule changed in training [5]. The pattern holds outside training design too. In 2026 the WNBA team Portland Fire dropped pregame shootarounds and built practice around changing the game itself, going on to beat the league's defending champion [6]. At the 1984 Olympics 17-year-old Jon Sieben stood 28 centimeters shorter than favorite Michael Gross. His coach had him swim the final in Gross's wake and attack only in the last lap. Sieben won 200-meter butterfly gold in world record time. That constraint worked as race tactics, not just as practice design [13].

The same idea has a name outside sport too. In 2007 the Royal Swedish Academy of Sciences awarded the Nobel Memorial Prize in Economic Sciences to Leonid Hurwicz, Eric Maskin and Roger Myerson "for having laid the foundations of mechanism design theory" [11]. The field asks economics the mirror-image question of game theory: instead of predicting behavior from fixed rules, design the rules so that self-interested behavior produces the outcome wanted. Myerson's founding result showed a system can be built so that telling the truth becomes each participant's own best move [12]. Coaches, economists and now software teams keep arriving at the same answer from three different directions.

3. What spec coding draws from this

Every point below is a real, hard limit built into the system, not a request. A prompt is only a suggestion. An agent can misread it, forget it, or talk itself out of it. None of the following can be argued with.

1. Write access only to the files the task needs. An agent fixes a bug in one part of the program. It can only change files in that one folder, nowhere else. That limit is a container mount: the system exposes only that one folder to the agent, nothing more. The spec file, the shared interfaces and another team's code stay out of reach. The filesystem simply will not allow it.

2. A wrong import gets blocked automatically, not caught by a person. A tool called a linter reviews every change on its own. The rule is fixed: no direct access from a billing module into another module's database. The agent can still try. The build then fails automatically. A person sees the broken change only afterward, if at all.

3. A fixed interface for the agent. Type signatures and API contracts set the rules between two parts of the program. These rules live on a read-only mount: the agent can read them but never write to them. It builds its own code against them. The compiler checks this automatically. A compiler cannot be argued with the way a prompt can.

4. Network access only when needed, otherwise none at all. A coding agent runs inside a sandboxed container. This workspace is sealed off from the rest of the machine. It reaches no network by default. It cannot install a new dependency mid-task. It cannot reach an outside service on its own. This is the same boundary least privilege draws around every tool an agent holds [8].

5. A machine-checked result instead of the agent's own word. A durable check does not ask the agent for its opinion. It measures the result directly, the same way a smoke detector is not asked whether it works: real smoke is used instead. One tool throws unusual, deliberately awkward inputs at the code. It checks whether the code holds up. A second tool plants a small bug in the code on purpose. If the tests catch it, the check passes. If they miss it, the tests are worthless, no matter what the agent claims. All that reaches the agent in the end is a simple result: passed or failed, with an exact error report. There is nothing to spin.

6. How strict the gate is matches the task. The same coaching framework sorts practice difficulty into three bands [2]:

   • Too easy. An agent changes one small thing in a stable file. A strict check blocks nothing new here, it just slows things down.
   • Hard enough. An agent builds a new feature. A strict type checker forces it to actually handle a rare edge case, not skip it.
   • Too hard. A check demands full test coverage, even for trivial one-line getters. Even a correct change gets stuck.

   A gate needs the same judgment. A relaxed type checker on a fresh prototype branch sits in the middle band. The same relaxed checker on the release branch slides into the first band: too easy, a gap. An agent finds it eventually.

Tuchel described exactly this shift in his own role:

"I don't interrupt the game anymore. I watch how the player deals with that space. I help them there." [7, translated]

A reviewer works the same way on a spec-coding agent. They watch what the constraints catch. They do not correct what a prompt failed to prevent.

4. An honest limit

A constraint can fail in a second, quieter way. The coaching framework behind this piece names it the over-constraining trap: a rule that forces one specific action, like "two touches only" before passing, does not invite a player's judgment, it replaces it [2]. The rule looks like the method working. It is the method's opposite, a decision the structure made instead of the person, exactly the failure the whole approach exists to avoid. The same trap catches software: a check strict enough to block every legitimate change looks identical in the logs to a check that is actually working, right up until someone needs to ship.

Constraint design is not automatically the better choice either. A 2024 review of the approach argues the opposite point too: traditional, prescriptive coaching still wins when the goal is raw speed of skill acquisition. The review explicitly warns against treating the newer approach as simply superior in every case [10]. A 2025 study of youth football training found a sharper result still: when the measured goal was team cooperation rather than individual skill, coach instruction actually outperformed constraint manipulation. Constraint changes on their own were linked to more individual, less cooperative play [9]. The field is not a universal substitute for a clear word. It's a tool that fits some goals better than others. The honest answer checks which goal is on the table before reaching for it.

5. Conclusion: the shape of the space

A rule you have to say out loud is a rule someone, eventually, won't hear in time. A rule built into the shape of the space doesn't need saying at all.

A constraint built into the structure holds by itself. A rule that depends on being said again and again eventually isn't.

6. Sources

[1] K. M. Newell, "Constraints on the Development of Coordination," in Motor Development in Children: Aspects of Coordination and Control, M. G. Wade and H. T. A. Whiting, Eds. Dordrecht: Martinus Nijhoff, 1986, pp. 341-360.

[2] I. Renshaw, K. Davids, D. Newcombe and W. Roberts, "The Constraints-Led Approach: Principles for Sports Coaching and Practice Design." Abingdon: Routledge, 2019.

[3] F. W. Otte, K. Davids, S.-K. Millar and S. Klatt, "When and How to Provide Feedback and Instructions to Athletes? How Sport Psychology and Pedagogy Insights Can Improve Coaching Interventions to Enhance Self-Regulation in Training," Frontiers in Psychology, vol. 11, art. 1444, 2020.

[4] M. C. Y. Lee, J. Y. Chow, J. Komar, C. W. K. Tan and C. Button, "Nonlinear Pedagogy: An Effective Approach to Cater for Individual Differences in Learning a Sports Skill," PLOS ONE, vol. 9, no. 8, e104744, 2014.

[5] B. Teune, C. Woods, A. Sweeting, M. Inness and S. Robertson, "Evaluating the influence of a constraint manipulation on technical, tactical and physical athlete behaviour," PLOS ONE, vol. 17, no. 12, e0278644, 2022.

[6] EssentiallySports, "Alex Sarama's Bold NBA-Inspired Coaching Move Fuels Portland Fire's Rise Above .500," Aug. 2026.

[7] T. Tuchel, "Der Fußball-Rulebreaker: Wie Leistungssportler das Vergessen lernen," talk, Executive Days, 2b.ahead, 2012.

[8] Betteryields, "Least privilege: the cage around the agent," 2026, internal piece (Beitrag 0016).

[9] J. O'Brien-Smith, M. R. Smith, M. Lenoir and J. Fransen, "Exploring the Effects of Instruction and Game Design on Youth Soccer Players' Skill Involvement and Cooperative Team Behaviour," Research Quarterly for Exercise and Sport, vol. 96, no. 1, pp. 109-115, 2025.

[10] R. Lindsay and M. Spittle, "The adaptable coach: a critical review of the practical implications for traditional and constraints-led approaches in sport coaching," International Journal of Sports Science & Coaching, vol. 19, no. 3, pp. 1240-1254, 2024.

[11] The Royal Swedish Academy of Sciences, "The Sveriges Riksbank Prize in Economic Sciences in Memory of Alfred Nobel 2007," press release, Oct. 15, 2007.

[12] R. B. Myerson, "Incentive Compatibility and the Bargaining Problem," Econometrica, vol. 47, no. 1, pp. 61-73, 1979.

[13] Australian Olympic Committee, "Jon Sieben," athlete profile, olympics.com.au, 2026.