andurilcode/limit-thinking

Limit Thinking

Core principle: Don't ask "what is the value at this point?" — ask "what does this converge to as we push the variable toward its extreme?" Most planning failures happen because people evaluate a snapshot instead of a trajectory. Limit thinking reveals the destination that incremental progress is silently heading toward.

The mathematical concept of a limit — studying what a function approaches rather than where it is — is a powerful reasoning tool far beyond calculus. It exposes asymptotes, phase transitions, and convergence traps hiding inside seemingly linear plans.

---

How to Execute This Skill

STEP 1 — Identify the Variables

For the system, decision, or strategy being evaluated, identify the key variables that could be pushed toward an extreme.

LIMIT ANALYSIS SETUP
System: [what's being evaluated]
Variables to push:
  - Variable 1: [what is it?] → Push toward: [0 / ∞ / 100% / some boundary]
  - Variable 2: [what is it?] → Push toward: [0 / ∞ / 100% / some boundary]
  - Variable 3: [what is it?] → Push toward: [0 / ∞ / 100% / some boundary]

For each: What does intuition say happens? (This is the naive expectation to test.)

Good variables to push include:

• Adoption rate → 100% ("what if everyone uses this?")
• Automation level → total ("what if no human is in the loop?")
• Scale → orders of magnitude ("what if we 10x or 100x this?")
• Time → long horizon ("what does this look like in 5 years?")
• Cost → zero ("what if this becomes free?")
• Speed → instant ("what if latency drops to zero?")
• Volume → extreme ("what if input grows without bound?")

Selecting Variables: The 2-3 Rule

A vague or ambitious prompt ("scale AI across the org", "go all-in on microservices") can generate 5+ candidate variables. Tracing all of them produces an exhaustive but overwhelming analysis. Instead, identify all candidates first, then select the 2-3 most critical using these filters:

1. Highest leverage: If pushed to the extreme, produces the most consequential or irreversible outcome?
2. Most hidden: Which variable's limit behavior would most surprise the decision-makers? (Variables where naive expectation ≈ actual convergence aren't worth tracing — they teach nothing.)
3. Most coupled: Which variable's limit behavior affects the other variables? (Trace the upstream variable first — its limit may redefine the others.)

List all candidates, mark the 2-3 you'll trace, and briefly note why the others are deprioritized. This keeps the analysis focused and the output actionable.

---

STEP 2 — Trace the Convergence

For each variable, don't jump to the extreme — approach it incrementally and watch what happens. This is where the insight lives.

CONVERGENCE TRACE: [Variable] → [Extreme]

Current state: [where things are now]
  ↓ Push slightly...
Incremental: [what improves or changes — usually positive]
  ↓ Push further...
Midpoint: [what secondary effects appear — often the first sign of non-linearity]
  ↓ Push toward extreme...
Near-limit: [what breaks, saturates, or reverses — the real finding]
  ↓ At the limit...
Convergence: [variable] → [extreme] means [outcome] → [what it converges to]

Does the outcome converge to what intuition predicted? YES / NO
If NO — what's the actual limit, and why is it counterintuitive?

Key patterns to watch for:

| Pattern | What it looks like | Example | |---|---|---| | Asymptotic ceiling | Returns diminish and flatten. You approach a maximum but never reach it. | Adding more engineers to a team: productivity asymptotes, then declines (Brooks's Law) | | Phase transition | The system changes state entirely at some threshold. No amount of incremental change prepares you for the discontinuity. | Water at 100°C: more heat doesn't make it "hotter water" — it becomes steam. Org maturity hitting L4: more tools don't help, you need structural change | | Reversal / collapse | The variable's effect flips sign. What was helping starts hurting. | Context in an agent: more context helps until it doesn't, then success rate drops (ETH Zurich finding). Automation of review: reduces burden until reviewers disengage entirely | | Convergence to zero | A human quality (attention, responsibility, skill) atrophies as the system takes over. | Pilots and autopilot: the more autopilot flies, the less skilled pilots become at manual flight — the limit of pilot skill as automation → 100% is dangerously low | | Divergence | No stable limit exists. The system oscillates or explodes. | Feedback loops without damping: over-correction → under-correction → chaos |

---

STEP 3 — Extract the Insight

The value of limit thinking is the delta between naive expectation and actual convergence. State it explicitly:

LIMIT INSIGHT

Naive expectation: "If we push [variable] toward [extreme], [outcome] will [improve/increase/get better]."

Actual convergence: As [variable] → [extreme], [outcome] → [surprising result].

The delta: [Why the actual limit differs from expectation.
            What force, feedback loop, or phase transition causes the divergence?]

Implication for the decision: [What should change about the plan, strategy, or design
                               given that the trajectory leads somewhere unexpected?]

---

STEP 4 — Find the Optimal Operating Point

Limit thinking doesn't just reveal where things break — it helps find where to stop pushing. If the convergence trace shows diminishing returns, reversal, or collapse past a certain point, that inflection point is the practical operating target.

OPERATING POINT ANALYSIS

Variable: [what we're tuning]
Benefit curve: [how benefit changes as variable increases]
Inflection point: [where marginal benefit starts declining significantly]
Recommended range: [where to operate, stated as a range not a precise number]
Signal to watch: [what early indicator tells you you've pushed past the optimal point]

---

Output Format

Variables and Extremes

List each variable being pushed and toward what extreme.

Convergence Traces

For each variable, the incremental trace from current state to limit:

• What the naive trajectory predicts
• Where secondary effects appear
• What the system actually converges to

Limit Insights

For each trace where actual ≠ expected:

• The delta between expectation and reality
• The mechanism causing the divergence
• Why this matters for the decision at hand

Optimal Operating Points

Where to operate if the limit reveals diminishing returns or reversal:

• The recommended range
• The early warning signal

Implications

What changes about the plan, strategy, or design given these convergence findings.

---

Thinking Triggers

Use these to sharpen the analysis:

• "If we push this variable all the way, what does the outcome actually converge to?"
• "Is this a linear improvement, or does it hit a ceiling / flip / collapse at some point?"
• "We're reasoning about a snapshot — what does the trajectory look like?"
• "What human quality atrophies as we automate this further?"
• "Is there a phase transition hiding between here and the extreme?"
• "Everyone is debating the current state — but where is this heading?"
• "What would a mathematician say the limit of this function is?"

---

Example Applications

• "Let's roll out AI code review to 100% of PRs" → Trace automation → 100%: reviewer attention converges to zero, critical thinking atrophies, the review becomes a checkbox
• "Let's add more context to the AI agent" → Trace context volume → ∞: performance improves then degrades, signal-to-noise ratio collapses past the optimum
• "Let's hire more engineers to go faster" → Trace team size → large: communication overhead grows quadratically, velocity per engineer converges toward zero (Brooks's Law)
• "Let's reduce meeting time to maximize focus" → Trace meetings → 0: alignment gaps emerge, decisions happen in silos, rework increases until net productivity is worse
• "Let's make the product free for adoption" → Trace price → 0: adoption explodes but perceived value converges to zero, support costs diverge, unit economics collapse
• "Let's keep investing monthly regardless of market" → Trace time → long: dollar-cost averaging smooths volatility, compounding dominates, the limit of a disciplined PAC is wealth accumulation despite short-term noise

---

Relationship to Other Skills

• Second-order thinking: Limit thinking is complementary — second-order asks "and then what?" for a single step; limit thinking asks "where does this end up?" across the full trajectory. Use limit thinking to set the destination, second-order thinking to trace the path.
• First principles: If limit thinking reveals a ceiling or collapse, first-principles thinking can question whether the variable being pushed is the right one at all.
• Scenario planning: Limit thinking can feed scenario planning by identifying the extremes that define the scenario space.
• Theory of constraints: If a variable hits an asymptotic ceiling, TOC can identify the constraint that creates the ceiling.