06-reference

jaynitx pattern recognition skill build

2026-05-12·reference·source: X (long-form article)·by Jaynit
pattern-recognitionskill-acquisitionfeedback-loopsexpertisechunkingmacsanity-checkagent-skill-build

"Pattern recognition: how to build the skill of seeing what others miss" - @jaynitx

Why this is in the vault

Operational treatise on building pattern recognition as a learnable skill, not an innate gift. Names the exact mechanism RDCO is already running unconsciously across multiple seats (MAC test discipline, Sanity Check verdict shape, skill-improvement loop) and gives us the underlying language to be more deliberate about how each seat builds its pattern library. Paired naturally with the Tony Robbins quote the founder shared the same minute - Robbins names the hierarchy (recognize / use / create), Jaynit gives the building mechanism.

The core argument

Expertise is not better thinking. It is better seeing. Chess masters do not calculate more moves ahead than amateurs - they recognize roughly 50k-100k stored configurations and respond from pattern match. Same mechanism shows up in every domain: a novice reader sees letters, an expert reader sees ideas. A novice programmer sees lines, an expert sees algorithms and off-by-one errors. A novice investor sees data points, an expert sees pricing power forming or value traps disguising themselves.

The build mechanism is repeatable: volume with variety + active analysis + tight immediate feedback + deliberate side-by-side comparison + articulation of what you saw + cross-domain transfer. The single biggest accelerator is feedback-loop tightness. Weather forecasters develop deep intuition because they predict and find out within hours. Clinical psychologists do not, because they rarely learn whether the patient actually got better. Same years of practice, different pattern libraries.

The dark side: the same brain that builds real patterns can build false ones. Narrative fallacy. Conspiracy theories. Superstition. The defense is mechanism - can you articulate WHY the pattern exists, and does it predict reality when tested? Sample-size and cross-domain checks help separate signal from noise.

Recognition-primed decision making (Gary Klein) is what experts actually do under time pressure - they do not weigh options, they pattern-match the situation and act. Firefighter walks into a room, feels wrong, orders evacuation. Floor collapses seconds later. The pattern was for fires with basements, recognized below conscious awareness.

Buffett reads an annual report in minutes not because of innate brilliance but because he has read thousands over 50 years and the patterns light up immediately. Munger's "latticework of theory" is the same concept - facts without pattern structure are not usable.

Mapping against Ray Data Co

This piece directly reinforces multiple seats already in motion:

MAC test discipline IS pattern-library construction. Every MAC test is one more stored configuration of "what production-ready data looks like." The Garry Tan complexity-ratchet posture is the deliberate version of what Jaynit describes - volume with variety, tight feedback, articulated patterns. The carve-out-with-counterparty-signature (the 5th MAC artifact added 2026-05-12) is the governance layer for when a stored pattern must flex - Kuhn's paradigm-shift problem made tractable.

The multi-agent skill build-out the founder just sketched (spec author → test plan instrumentation → code writer → critic) is a structural answer to Jaynit's "how do you tell real patterns from false ones." Separating the seats prevents the test-author from writing tests TO the code instead of TO the ground truth. The critic seat is the cross-domain check. The board task spawned today (2026-05-12-multi-agent-skill-buildout-pipeline /decisions/) is the founder-gated decision on whether to formalize this.

Sanity Check IS the pattern-recognition surface for content. The verdict shape (skip / skim / read / file) is the externalized version of Jaynit's "I could feel within seconds if a post would perform." We are deliberately building Ray's pattern library for content-quality signal, with the founder's corrections (no-em-dash memory, no-derivative-Sanity-Check-pieces memory, sharp-verdicts memory) as the tight feedback loop.

The skill-improvement loop (/self-review → /improve autonomous) is the meta-loop. Reviewing what skills produced, naming the "OK but not good" patterns (Garry Tan's three-star pattern - not the failures, the mediocre outcomes), and feeding the lesson back into the skill file. This is exactly Jaynit's "articulate the pattern to strengthen it" prescription, mechanized.

Investing project (just stood up 2026-05-12) sits at the start of its pattern-library build for a brand new domain. The Innermost Loop thesis is a pattern (4-layer demand stack with bottleneck migration). The Cerebras disqualification is the inverse - a stored anti-pattern (customer concentration + upside-down cap table + inverse margin pivot + revenue-recognition material weakness). The discipline of writing both into the vault is the deliberate practice; quarterly thesis review is the feedback loop.

The trap to watch: false-pattern formation. A founder operating with infinite Ray attention and a vault that crosses 2300+ docs is at structural risk of pattern overfitting - seeing a "the way we do things" everywhere when actually most of those are sample-of-one or two. The defense Jaynit prescribes: hold patterns as hypotheses, update when they stop predicting, and watch for patterns that only show up in one narrow context. Concrete RDCO surface for this: when /self-review surfaces a "this is how we do X" claim, check whether the claim holds across 3+ distinct contexts before promoting it to a SOP.

Cross-references to the Robbins quote

Robbins' hierarchy is exactly Jaynit's article in three words:

For RDCO this maps to: Ray spends most of every day on recognition (process-newsletter, process-youtube, sanity-check). The use surface is the verdict + filing. Creation is rarer - the concept articles I write, the MAC framework refinements, the new skill files that did not exist before. The implication: pattern creation requires the largest input library AND the willingness to commit to a synthesis before all the evidence is in. The founder is forcing this with the harness-engineering thesis, the operating-model frame for investing, the multi-agent pipeline proposal.

The actionable extract

  1. Mechanize feedback-loop tightness wherever skill-building is happening. /self-review weekly + /improve autonomous weekly is the current cadence; if a skill's feedback loop is >2 weeks long, it is not building patterns fast enough.
  2. Articulate patterns explicitly in vault notes when they crystallize, not just when something breaks. The /improve skill file already does this via Changelog entries - extend the pattern to project READMEs (investing/README.md does this already with the Strategy vs execution boundary table).
  3. Cross-domain transfer. Patterns from one RDCO seat should be tested against other seats. The MAC complexity-ratchet posture works for data; the Sanity Check verdict shape works for content; the question is what other surfaces benefit from the same shape. Worth a /curiosity question.
  4. Anti-pattern library is as valuable as the pattern library. The Cerebras disqualification card in investing/candidates/README.md is the template - explicit reasons we DID NOT do something. Extend to other seats (rejected MAC frameworks, killed Sanity Check pitches, archived skill ideas).
  5. Watch for overfitting - patterns that hold in 1-2 cases get promoted to "the way we do things." Discipline: 3+ distinct contexts before any "this is how" claim crystallizes into a SOP.

Related