← Back to blog

Why 134 Standing Rules Beat Permission Prompts for AI Work

Stop clicking Approve forty times a day. Replace permission prompts with 134 standing rules for trustworthy AI without constant supervision.

Why 134 Standing Rules Beat Permission Prompts for AI Work

Most teams supervise AI agents the same way: the agent proposes an action, a human clicks "Approve," and everyone calls that a safety system. It is not. It is a doorbell that rings forty times a day, and after the tenth ring, nobody is really listening anymore.

Over four and a half months, we replaced permission prompts as the primary control mechanism with 134 written standing rules — and rebuilt how we manage AI work around them. This article covers the nine things we now do differently, and why each one matters in business terms: fewer interruptions, fewer silent failures, and AI output that can be trusted without being babysat.

Why permission prompts fail as a management system

A permission prompt asks a human to make a judgment call with almost no context, under time pressure, dozens of times per day. Research on workplace interruptions by Gloria Mark at UC Irvine found that after an interruption, it takes on average over 23 minutes to fully return to the original task. An agent that interrupts a senior engineer even ten times a day is not being supervised — it is destroying the supervisor's working day while producing the feeling of oversight.

There is a second, quieter failure mode: approval fatigue. When most prompts are routine ("read this file?", "run this test?"), people stop reading them. The one prompt that genuinely matters — the destructive one, the outward-facing one — gets the same reflexive click as the other thirty-nine. This is a general industry pattern, and it mirrors what security teams have known about alert fatigue for years: a control that fires constantly is a control that gets ignored.

The skepticism is warranted, by the way. In the 2025 Stack Overflow Developer Survey, 46% of developers said they actively distrust the accuracy of AI tools — up from previous years, even as usage keeps climbing. The answer to that distrust is not more clicking. It is better management structure.

Put simply: permission prompts are a manager approving every keystroke. Standing rules are an operations manual. No serious business runs on the first model with human employees, and there is no reason to run AI that way either.

The nine things we do differently

1. Safety lives in written rules, not in approval clicks

Every recurring judgment call gets converted into a standing rule the moment it appears twice. "Never send an email without explicit sign-off." "Deploys go through the safe-deploy path, never raw file sync." "Shared registries are mutated only through the audited change path, never directly."

Real numbers: it took four and a half months to accumulate 134 of these rules. That sounds like a lot of writing. It is roughly one short rule per day — a few minutes each — in exchange for eliminating the corresponding prompt forever. A rule written once replaces hundreds of future interruptions.

The analogy that made this click for us: permission prompts are airport security — a checkpoint that inspects everything and slows everyone down. Standing rules are building codes — the safety is designed into the structure, so the inspection only happens where it actually matters.

2. Rules capture the why, not just the what

A rule that says "don't repoint server IDs via SQL" gets followed until the first ambiguous case, then broken. A rule that says "don't repoint server IDs via SQL, because a naive repoint silently disables security audits" gets followed and generalized. The agent (or the new team member) can now handle cases the rule's author never anticipated, because they understand the failure the rule prevents.

This is the difference between an employee handbook that says "no personal deliveries to the office" and one that explains the mailroom got buried last December. One produces compliance; the other produces judgment.

3. Irreversible actions are fail-closed by default

The rulebook divides all actions into two classes. Reversible ones — reading, analyzing, drafting, editing files under version control — run freely. Hard-to-reverse ones — sending emails, deploying to production, deleting data, spending money — stop and wait for a human, always, no matter how confident the agent is.

This single distinction removes the hardest judgment call from the moment of action. Nobody decides in real time whether a deploy is "probably fine." The category was decided months ago, in writing, when nobody was in a hurry.

4. Confirm the problem is real before fixing it

The most expensive AI failure mode is not a wrong answer — it is a confident fix applied to a problem that does not exist. So a standing rule requires reproducing or verifying every reported problem before any change is made. Two questions gate every fix: what exactly is broken (the thing), and what test proves it is broken (the checker)?

Industry data suggests this discipline matters more with AI than without it. A randomized controlled trial by METR in 2025 found that experienced open-source developers were actually 19% slower when using AI tools on their own codebases — while believing they had been faster. A significant share of that gap comes from reviewing and correcting plausible-but-wrong output. Verification before action is how that tax gets cut.

5. Verify outcomes, not intentions

Reviewing an AI's plan feels like oversight but proves nothing — plans are exactly what language models are best at making look good. So review moved to the other end: the diff, the test run, the live behavior. Every task must define what "done and verified" looks like before work starts, and the check runs after.

The management parallel is direct. Competent managers of human teams do not attend every working session; they define acceptance criteria and inspect deliverables. AI management converges on the same shape, just faster and with tighter loops.

6. Durable knowledge goes into a canonical store, not the chat

An AI session's context dies when the session ends. Anything learned in conversation — a gotcha, a decision, the reason an approach was rejected — evaporates unless it is written somewhere permanent. So a standing rule: durable understanding gets captured in a canonical, backed-up store (a database, versioned docs), keyed to the product it belongs to. Session memory is a cache; the canonical store is the truth.

The business test for this rule is brutal and simple: if the machine died tonight and was rebuilt from backups tomorrow, would a fresh AI session be immediately productive? If the answer requires "re-explaining everything we figured out last month," the knowledge was stored in the wrong place. What this means for your project: budget for the writing-down, or pay for the re-discovering — repeatedly.

7. Small, reversible steps — with an audit trail for the shared stuff

Big-bang changes are where AI confidence does the most damage. The working rule is: the smallest change that can be independently verified, then the next one. And for anything shared — configuration registries, records other systems depend on — mutations go through an audited path that logs who changed what, from which value to which value.

This is not bureaucracy; it is the same reason banks use ledgers instead of erasers. When something breaks three weeks later, the audit trail turns a forensic investigation into a five-minute lookup.

8. Reports brief the result, not the worklog

Unmanaged AI output has a signature: five paragraphs of narrated process wrapped around one sentence of actual finding. So reporting is governed by rules too: verdict first, evidence after, risks and gotchas surfaced explicitly, no re-explaining of things the reader already knows. Depth stays in the work; the report is a briefing.

The compounding effect is easy to underestimate. A manager reading ten AI reports a day saves significant time when each one leads with the answer — but the bigger win is that buried risks stop being buried. A skim of a well-structured briefing still delivers everything that matters. A skim of a worklog delivers nothing.

9. Autonomy is earned per-category, not granted globally

The 134 rules are not static — they encode a graduated trust model. Actions with a track record of safe execution get standing permission. New categories of action start supervised. And when something goes wrong, the response is never "supervise everything again" — it is one new rule covering that specific failure, with the why attached.

This mirrors how delegation works with people. A new hire does not get production access on day one, and a senior engineer does not need approval to run tests. Blanket policies in either direction — full autonomy or full supervision — are both management failures. The interesting work is in the middle, and it has to be written down.

Alternatives and trade-offs

Honest take: this approach has real costs, and it is not right for everyone.

Approach Upfront cost Ongoing cost Failure mode
Approve every action None Constant interruptions, approval fatigue The one critical prompt gets reflexively clicked
Full autonomy, no rules None Cleanup after incidents Irreversible mistakes with no warning
Standing rules (this model) Months of rule-writing Rule maintenance A gap in the rules goes unnoticed until it bites

The standing-rules model front-loads effort that prompt-clicking spreads out invisibly. For a team using AI a few hours a week, that investment probably never pays back — the interruption load is low enough that prompts are fine. The economics flip when AI runs daily and unattended: at that intensity, every recurring prompt is a recurring tax, and every unwritten rule is a latent incident.

The model also demands honesty about failures. A rulebook only improves if every incident produces a written rule rather than a shrug. Teams that skip the post-incident write-up end up with the worst of both worlds: no prompts and no rules.

Key takeaway for business:

  • Permission prompts measure activity, not safety. A human clicking "Approve" forty times a day is not oversight — it is theater with an interruption cost attached. Interruption research suggests each context switch costs the better part of half an hour of refocus time.
  • Written rules are an asset; approvals are an expense. Each of the 134 standing rules was written once and works every day since. Every approval click is spent the moment it happens and buys nothing tomorrow.
  • The management structure matters more than the model. The METR study's 19% slowdown happened with capable AI tools — the gap was in how the work was verified and managed. Here is what we recommend: before upgrading models, write down the ten rules you are currently enforcing one approval click at a time. That document will outlast every tool it governs.

This article is based on publicly available sources and may contain inaccuracies.

Related articles

✦SqueezeAI
  1. Frequent approval prompts produce approval fatigue: when routine actions dominate the queue, people stop reading each prompt, so the one genuinely risky action receives the same reflexive click as the thirty-nine routine ones — a pattern identical to security alert fatigue.
  2. Each AI interruption carries a hidden time cost: Gloria Mark's UC Irvine research found recovery from a single interruption takes 23+ minutes on average, meaning even ten daily approval prompts to a senior engineer consumes the working day while producing only the appearance of oversight.
  3. Converting a recurring judgment call into a written standing rule is a one-time investment — roughly one short rule per day — that permanently eliminates the corresponding prompt; at that rate, 134 rules accumulated over four and a half months each replace hundreds of future interruptions.

Powered by B1KEY