Luck System

Thesis: Luck is not random. The expected frequency of positive-sum unplanned outcomes — "lucky breaks" — is a function of durable choices about what you do, who can see it, whether they can route an opportunity to you, and whether they want to. Most people treat luck as weather; it is closer to landscape. You do not control each rainfall; you decide whether you live on a plain, a coast, or a desert.

The common formula — Jason Roberts' Luck = Doing × Telling — captures the two most important dials. The fuller causal graph has more. This system describes the full surface area, tags the failure mode each dial covers, and prescribes how to turn them sustainably.

See supporting research: [[luck-surface-area-research]], [[network-theory-and-serendipity]], [[cold-outreach-empirical-2026]].

The model

Luck is not a scalar. It is a distribution of unplanned positive outcomes over time. Most encounters yield small outcomes; rare encounters are transformative. What this system shapes is the parameters of that distribution — the rate at which lucky events arrive, the fraction of them that convert to value, and the size of the typical payoff.

The crudest mnemonic is still Jason Roberts' Doing × Telling. It is directionally useful and terrible as a model: it can't capture substitution (high generosity partly replaces low reach), it treats independent dials as if they were truly independent (D and T are deeply coupled — writing is often how doing gets told), and its "any zero kills everything" implication is empirically wrong (a strong enough D has carried weak-T operators many times — Ramanujan's letters to Hardy, obscure academic work that gets discovered, public-domain code that accrues an invisible audience). The real object is more honest, and more useful:

Three stages

Think of a lucky break as a three-stage pipeline. All three stages must fire; within each stage, dials combine.

Stage 1 — Encounter rate ($\lambda$). How often an unplanned opportunity arrives in your life at all. This is a Poisson-like rate — a count-per-year. Pushed by:

• F — Frequency of encounters. Shows, dinners, calls, coffee, replies, comments, attendance. Interactions are the literal events at which luck is generated; below some threshold, rare events don't have enough trials to surface. Canonical target: 1–2 substantive new interactions per day, integrated with [[socializing-system]].
• T — Telling. Legible, findable public signal. A newsletter, public repo, talks, social posts, a personal site that passes the 30-second stranger parse. Telling effectively generates encounters you did not initiate — the inbound channel.
• B — Bridging ties. Network position. Pure weak ties (0–1 mutual connections) underperform moderately-weak ties (~10 mutuals) for opportunity transmission (Rajkumar et al., Science 2022). Cross-cluster ties compound fastest (Chetty et al., Nature 2022). Structural holes, not random acquaintances.
• M — Memory freshness. Ties decay ≈0.1 SD of emotional closeness per year with zero contact (Roberts & Dunbar, 2011). A dormant tie is nearly a cold one — dropped out of the encounter rate. Light recurring maintenance keeps the tie "live" in λ.
• G — Generosity. Pre-given, unreciprocated help — the [[never-eat-alone]] thesis, empirically grounded in Adam Grant's work. Acts as a stock that builds latent inbound: people who feel unobligated obligation route opportunity to you at unpredictable times. Adds to λ with a long time-lag.

Within the encounter stage, dials substitute: high T can carry low F (people come to you instead of you going to them); high G can carry low B (people bridge for you instead of you bridging for yourself). You need the stage-level output — some meaningful λ — not every dial at full.

Stage 2 — Conversion ($p$). Given that an encounter happens, the probability it turns into something positive. A sigmoid in the joint contribution of:

• S — Specificity. The legibility of "what I do" in 12 words or less. Below a threshold, intermediaries can't place you and the encounter dissipates. Above saturation, more specificity just narrows the funnel.
• L — Legibility of the ask. When an encounter requires an ask, the ask must be narrow enough that the recipient can imagine saying yes. See [[asking-system]].
• V — Velocity. Response time on inbound. A lagged reply is a closing window — conversion collapses fast as V grows. See speed-system. Within a few months of consistent sub-24-hour response, referrers begin routing time-sensitive opportunities preferentially.
• D — Doing (partial). The capability component of D that lets you actually fulfill what the encounter asked for. The rest of D belongs to stage 3.

S and L are near-substitutes (both fix the "they can't route to me" failure). V has a threshold — below it, conversion crashes; above it, further improvement plateaus. D is the base competence underneath all of it.

Stage 3 — Magnitude ($m$). The size of the outcome when it converts. Fat-tailed: most payoffs are small, a handful are transformative, the distribution's tail exponent is what matters over a career. Shaped by:

• D — Doing (full). Your work product's ceiling. The biggest outcomes are proportional to the biggest possible output you're capable of, not the average output.
• T — Telling (tail-side). Telling determines whether the ceiling is seen. A 99th-percentile artifact seen by three people produces a different magnitude than the same artifact seen by 300,000 — even when conversion rates are identical.

Magnitude is a distribution, not a number. You don't optimize it by squeezing the mean; you optimize it by widening the tail.

Putting it together

A first-order expression:

$$\text{Expected annual lucky-break value} ;\approx; \lambda(F, T, B, M, G) ;\cdot; p(S, L, V, D) ;\cdot; E[m \mid D, T]$$

The multiplication across stages is real: you need an encounter and conversion and some magnitude for a lucky break to register. Zero at any stage is zero output — because that's how a pipeline works, not because dials "multiply."

Within a stage, dials combine more like a weighted concave sum with couplings:

$$\text{stage output} ;\sim; \sigma!\left(\textstyle\sum_i w_i \cdot g(x_i - \theta_i) ;+; \sum_{i<j} c_{ij}, x_i x_j\right)$$

Where $\sigma$ captures the floor-and-saturation shape of each dial, $\theta_i$ is the dial's floor, and $c_{ij}$ captures pairwise couplings (D × T is the strongest; B × M is strong; G × V has a less obvious but real interaction). Translation: within a stage, you get substitution and diminishing returns — not a product.

Implications worth keeping

• Diagnose the binding stage, not the binding dial. If encounter rate is the limit, no work on conversion-stage dials helps. If conversion is the limit, more encounters just waste more opportunities.
• Above-floor, below-saturation is where effort pays. The efficient frontier runs inside each dial's middle range. Taking F from 0 → 1 encounter/week is transformative; taking it from 30 → 31 is a rounding error.
• Magnitude matters more than it feels like. Career-dominant outcomes come from a handful of fat-tail events. Optimize D and T first for ceiling, not for average quality.
• The system is dynamic, not static. Generosity and memory are stocks (accumulate or decay with time). Velocity is conditional (only matters when encounters fire). D compounds on itself. Treat LSA as a stock-flow system running in real time.
• Outcomes are stochastic. You cannot predict which week the break arrives in. You can raise λ, p, and E[m] and then run the ritual patiently across years. The right unit of reasoning is annual expected value, not weekly hit-rate.

Diagnosis before prescription

Before turning a dial, ask which stage is binding (encounter / conversion / magnitude), then which dial inside that stage is the bottleneck. A quick self-audit:

• Fewer than 10 people outside your direct collaborators could name what you work on → T is binding. Fix legibility before volume.
• "What I do" requires more than one sentence → S is binding. Narrow the niche, even if the narrowing feels premature.
• Asks get warm replies but rarely convert → L is binding. The asks are too vague; tighten.
• Network is wide but homogeneous (same industry, same age, same city) → B is binding. Invest in one cross-cluster tie a month before you add another same-cluster contact.
• You know people but haven't spoken in >6 months → M is binding. Weekly 30-minute weak-tie refresh ritual.
• You give sparingly or transactionally → G is binding. Start with low-cost high-leverage help (intros, proofreads, sending someone a relevant link unprompted).
• Opportunities come in but die in your inbox → V is binding. Fix the response layer first (see speed-system); more inbound will make things worse until V is sound.
• You have interesting work, legible positioning, strong ties, fast response — and luck still feels scarce → F or D is binding. Either you need more encounters per unit time, or the work itself isn't remarkable enough to be routable.

Most people — including high-performing ones — have two binding dials, not eight. Turn those; accept "good enough" on the rest.

The weekly ritual

A durable protocol rather than a burst. All of this fits inside ~4 hours/week, anchored to a weekly review block.

Ship

• Publish one artifact this week. Aim for a narrow-topic post, a working public note, a product update, or a public thread. Published > drafted. Biweekly minimum to avoid Collison-style latent-reader collapse; weekly is sustainable when the thinking is batched on the weekend.
• Update the "now" page / homepage so drive-by visitors can answer "what is this person doing right now." See website.

Bridge

• 20 warm-intro-requests/week to targets 2 hops away. Route via known bridge brokers, not cold.
• 30–40 cold 1:1 sends/week — narrow ICP, specific hook, ≤120 words. See [[cold-emailing-system]] for format and [[cold-outreach-empirical-2026]] for current-year empirics and what counters the LLM-spam flood.
• Channel selection is not fixed. Re-check [[cold-outreach-empirical-2026]] at least annually — the relative yield of email vs LinkedIn vs DMs vs physical mail shifts as saturation and filtering evolve. Current empirics favor LinkedIn first-touch for many personas, with email preferred for senior operators, founders, and academics. Verify before relying on it.
• Send windows: 8–11 AM local time to the recipient. Batch, don't dribble.

Maintain

• 3–5 weak-tie refreshes: send a short "I saw X and thought of you / here's a thing you'd like / what are you working on?" message to contacts you haven't spoken with in 2+ months.
• 1 unprompted give: an intro between two people who should meet, a proofread offered, a relevant link forwarded.

Review

• Open the LSA dashboard (see Metrics below). Check the dial diagnosis — did any dial slip to binding?
• Reply to anything that has been sitting >24 hours. Empty the inbox.
• Pick next week's publish artifact and note the cold-outreach targets.

Monthly compounding — the infrastructure layer

These produce unplanned outcomes on a 3–12 month lag. Schedule once; let them run.

• Recurring low-friction event. Monthly brunch, salon, dinner, hike, co-working morning. Same day of month, same place, same 8–15 invitees rotating. Densifies the network on autopilot and manufactures Memory-refresh with zero marginal effort. Cate Hall's bagel-brunch template. Personal rule: for every city I'm in for more than a month, host a recurring event/party.
• Public body of work. One narrow-niche topic, held for 6+ months. A single topic owned is worth ten topics tried. The niche should be narrow enough that a stranger can predict what you'll write next.
• Cross-cluster tie acquisition. One new contact per month in a cluster you're not already in. Explicitly favor people who don't know the people you know.
• Long-form signal assets. One evergreen artifact per quarter — a tutorial, a dataset, a tool, a reference doc, a talk recording — that stays discoverable. These are the luck coupons that pay out for years.

Personal priors & tactics

Rules I've found matter beyond the general framework. Some come from direct conversation, some from running the system against my own life.

Communicate what you want — publicly

From a conversation with [[neav-meidar]] (originally logged here 2025-12-10): there is a large, underrated return from stating what you want out loud, online, and to everyone you meet. People cannot route opportunity to a target they cannot see. Be specific enough about the ask that someone who happens to hear it can action it. This is the T and S dials operating on the ask itself, not just on your identity.

Operate from genuine curiosity

See [[socialization-system]]. People who are intensely interested in a thing will talk for hours to anyone else who is genuinely interested. Ask real questions. You compound two things: (1) knowledge the rest of the market doesn't have, which is economic leverage for a consultant (you are paid inversely to how many people can solve the problem), and (2) warm durable relationships in niches you care about. Cold emails, LinkedIn pings, and "growing your garden" are the systematic version of this.

People notice when you show up

Consistently giving more than the assignment requires makes people remember your name. Those who already behave as if they belong in a better position find their way to it. This is Woody-Allen "80% of life is showing up" but with the qualifier that how you show up compounds faster than that you showed up.

Air out your weirdness

Show publicly the things you're interested in, weird about, genuinely excited by. This attracts people who are weird in the same way, who share your values, who will actually enjoy working with you. Homogenized signal attracts homogenized contact. See website for the surface-area of this in practice.

Give before you take

[[never-eat-alone]], formalized as G in the formula above. Time, attention, resources, acts of service, pre-given. The goodwill and trust make it dramatically easier to ask when you need to, but the point is not transactional: the network position of a habitual giver is qualitatively different from that of a scorekeeper.

Host events

Most people want social interaction; almost no one wants to host. Hosting is structurally privileged — you pick the guest list, the topic, the vibe; everyone thanks you afterward; the network density inside your chosen cluster rises for approximately zero marginal cost once the first event is scaffolded.

One way to increase serendipity is to throw parties or events, especially a regular series of get-togethers that's low-pressure to attend. This can be extremely simple, like having acquaintances over for coffee on your stoop, or a casual meeting at a pizza place. You can do this if you know five people in your city, or five hundred.

Choose where you work physically

Work in places where interesting people are likely to pass through — coffee shops near labs, co-working spaces tied to specific industries, libraries and cafes adjacent to universities. High foot traffic plus niche density plus being friendly generates serendipity without additional effort.

First impressions count disproportionately

See [[first-impressions-system]]. The Bayesian prior a new contact forms in the first 60 seconds is expensive to update. Every subsequent interaction is filtered through it. Warmth, clarity of introduction, and confident specificity about what you do compound from the first meeting onward.

Anti-patterns (and why they look like progress)

• The over-networker. Everyone knows you, no one works with you. Telling drowns Doing. Fix: halve outbound, double shipped work.
• The hoarder. Big contact CRM, zero maintenance cycles. B looks strong but M is zero — the network is already mostly cold. Fix: the weekly refresh ritual.
• The sprinter. 6-week burst of visibility, 6 months of silence. Telling is volatile; the decay between bursts erases most of the gain. Fix: biweekly cadence, held indefinitely, even at lower quality.
• The generalist. High D, high T, but S so low no one can route luck specifically. Fix: pick the narrow lane, even if the commitment feels premature.
• The over-giver. Gives infinitely, burns out, resents the network. G spikes, then crashes. Fix: disagreeable-giver posture — say no to low-leverage gives, say yes to high-leverage ones.
• Volume before signal. Scaling outreach when reply rates are <8%. List-size has brutal non-linear costs (1–200 recipients ≈ 18% reply; 1000+ drops to ≈8%). Fix before scale.
• Confusing luck with action. Trying to "make something happen" under pressure. Most lucky breaks in retrospect arrive during steady-state maintenance, not during the sprint. Fix: the ritual runs through both high- and low-motivation weeks.

Integration with adjacent systems

• speed-system — response velocity on inbound is a necessary condition for LSA to compound. V in the formula = speed-system's bread-and-butter.
• [[asking-system]] — L (ask legibility) lives inside the asking system. When an unplanned opportunity requires an ask, ask quality determines conversion.
• [[cold-emailing-system]] — tactical craft for the Bridge step of the weekly ritual.
• [[email-writing-system]] — craft for every written artifact, including cold outreach and warm-tie refresh.
• [[communicating-system]] — the intra-relationship analog of V. If you care about the response, respond quickly.
• [[anti-holding-back-system]] — the F (frequency) dial is killed by holding back from attending, posting, asking. LSA is the structural case for why holding back is expensive.
• [[socializing-system]] / [[socialization-system]] — the conversational craft. LSA determines the volume and selection of encounters; these systems determine the quality of each.
• [[first-impressions-system]] — compounds with F and B. First impressions are the Bayesian prior that filters every later interaction.
• website — the asset that carries T and S. A personal site that answers "what does this person do and want" in 30 seconds is the durable version of the pitch.
• [[never-eat-alone]] — G, formalized.
• [[essentialism-system]] — binds S (narrow niche). LSA is not "do more things"; it is "do fewer things more legibly."

Metrics worth tracking

Simple enough to glance at weekly. Don't over-instrument.

Failure modes of the system itself

• Scoreboard orientation. Counting metrics instead of producing meaningful work. If D is strong and the metrics drift, trust D — the numbers are proxies.
• High-motivation calibration. Designing the ritual for the energetic version of you. Calibrate to the tired version. A 30-minute refresh you reliably do beats a 2-hour block you skip when busy.
• Performative publishing. Shipping low-quality artifacts to hit the cadence. Pick the cadence you can hit at real quality, hold it indefinitely, don't stretch.
• Asking-system contamination. LSA is not a license for high-volume low-quality asks. See [[asking-system]] for quality bar.

The underlying worldview

Luck is real and unequal, but it is mostly a landscape you build. It rewards people who ship, are findable, are routable, are trustworthy, are fast, and are kind over periods measured in years. It punishes sprinting, hiding, vagueness, slowness, and transactionality. The dial with the biggest lag — public writing on a narrow topic held for years — is the one most people under-invest in, because the first 6 months are unrewarding. The dial with the biggest short-term return — response velocity — is the one most people never install, because the reward is invisible until someone routes an opportunity instead of forgetting you.

You cannot force luck. You can live where luck visits.

---

Feedback to Integrate

Habits

Questions

Systematic Improvements

Projects

See also knowledge-management-system