How expertise forms

Teams want their dinner planner to get faster after a few good weeks. The usual approaches shorten the prompt, cache the last plan, or fine-tune a model on household chats. Each can produce speed. None produces expertise. Expertise is reusable, guarded judgment that survives new people, new models, and regression suites.

Expertise is how Thoughtware earns automaticity without abandoning judgment. Teams sometimes describe the result as "the product finally getting us." The architectural reality is an artifact with fields.

Helpful context: Deliberation and automaticity explains slow versus fast paths. Evaluation as engineering supplies the evidence promotion requires. Experience becomes expertise covers the pipeline from raw trajectories to promotion.

Expertise as an artifact, not a vibe

Expertise must be legible to the next engineer, the next evaluator, and the next model swap. Legibility requires fields. A prompt that shrank after week three is not expertise. A cached JSON plan is not expertise. A behavioural shift nobody versioned is not expertise.

The Meal Companion's Busy Week Pattern v0.2 is the reference shape on this page. Other notes link here for the definition. They do not redefine expertise in passing.

What expertise contains

Using Busy Week Pattern v0.2 as the reference:

Field	Purpose
Applicability conditions	When this strategy may run
Required information	What must be present in working state
Strategy skeleton	Ordered moves: anchor busy evenings, spinach deadline, leftover subplan
Checks	Exact validations and critique hooks
Failure modes	Named ways the pattern misfires
Fallback	Return to full deliberation when guards fail

Expertise is legible. Someone joining the team can read v0.2 and know what the fast path does and when it must not run. If reading the artifact requires interviewing the person who tuned the prompt, the artifact is incomplete.

Mature expertise does not demand perfect prediction. It improves calibration between evidence, confidence, and consequence.

Introduction to Thoughtware · Ch. 12

How it forms

Expertise forms when similar weeks produce similar successful trajectories, when evaluation separates lucky outcomes from sound process, when applicability conditions stabilize across runs, when promotion criteria pass with approval where required, and when a library entry replaces ad-hoc rediscovery. Each step adds a layer of confidence that compression is earned rather than assumed.

It does not form when a model is fine-tuned once on chat logs, when one good plan is cached forever, or when a temporary correction like "avoid pasta this week" silently hardens into policy. Those paths produce speed without the accountability that makes speed trustworthy.

Formation is slow relative to demo culture. That slowness is the price of speed teams can trust. Teams that skip formation still ship fast paths. They often discover false familiarity in production instead of in regression.

Promotion is a gate, not a mood

Promotion committees sometimes confuse repetition with maturity. Ten busy weeks suggest a pattern. They do not prove guards are complete. Evaluation must ask whether process was sound given available information. Regression must include near-miss weeks. Approvers must exist when authority requires them.

Busy Week Pattern v0.1 might promote after three stable weeks with documented corrections. v0.2 might follow a misfire when only one low-time evening existed. Version bumps are evidence of learning, not embarrassment. Unversioned fixes hide learning from the organization.

Expertise lives in libraries

Promoted expertise belongs in a library of cognition: searchable, sealed, substitutable. It does not belong buried in agent-private prompts. GenerateCandidates and AssessMealPracticality may live in a domain library. Busy Week Pattern lives where expertise artifacts live: versioned, inspectable, callable.

Libraries of cognition covers placement rules. This page owns the shape. The library note owns sharing and substitution.

Expertise trapped in agent-private prompts does not scale across teams or models. Private shortcuts are where false familiarity hides: unaudited, unversioned, unshared. When someone leaves the team, private expertise leaves with them.

Runtime behaviour when expertise runs

When guards pass, the agent loop changes shape. Fewer exploratory cognitive units run. The strategy skeleton drives ordering. Checks and critique hooks remain. Fallback stays armed. Automaticity compresses deliberation where evidence supports compression. It does not remove accountability.

When guards fail, expertise steps aside immediately. Apologizing for returning to the slow path trains the wrong expectation. Frontier terrain deserves deliberation. Users recognize the switch as care, not regression.

Expertise versus other memory forms

Expertise sits beside endorsed knowledge, context, and experience. Knowledge states facts approved for use. Context states material true for this run. Experience logs trajectories for evaluation. Expertise packages reusable strategy when promotion criteria pass.

Confusing expertise with knowledge produces guards that are really preferences. Confusing expertise with context promotes throwaway remarks into global policy. Grounded verbs keeps the vocabulary aligned with forms.

What this looks like in the household planner

After three evaluated weeks with two-plus busy evenings, acceptable leftovers, and stable corrections, the team promotes Busy Week Pattern v0.1. The regression suite includes weeks with one low-time evening that should fail the guard, weeks with medical mentions that should defer, and weeks with conflicting spinach constraints that challenge the skeleton. v0.2 adds a failure mode discovered in regression: spinning composition when only one low-time evening exists. The guard tightened accordingly.

Next run with guards passing: compressed strategy executes. Next run with doctor mentioning low-FODMAP: guard fails, expertise does not run, full loop resumes. Speed without false familiarity.

Fallback rates in production deserve documentation too. Rising guard failures may mean conditions are stale, not that the model degraded. Expertise maintenance is part of ownership.

Common mistakes

Expertise as prompt length reduction is the most common: shorter prompts without guards or evaluation. Promotion without conditions is a close second, where one success story becomes global policy. No fallback means the fast path fails silently instead of returning leadership to deliberation. Expertise without checks means the strategy skeleton runs without exact validations at the end. Naming drift means three names for the same pattern in code, docs, and support macros, and nobody can search for the canonical version.

Relationship to economics and compression

Expertise lowers cost per run when guards hold. It does not eliminate evaluation cost forever. Regression suites grow with failure modes. Libraries require curation. The economic win is conditional speed with bounded downside. Unconditional speed is gambling.

Loop compression in the Building track picks up runtime effects. This page supplies the artifact that makes compression legitimate.

Review questions before promotion

Every expertise promotion review benefits from a standard set of questions. Can a new engineer execute the fallback path from documentation alone? Does regression include at least one near-miss week per guard condition? Does product copy describe when the fast path will refuse? Does telemetry record guard pass and fail separately from model latency? Does an approver exist for every knowledge item the strategy assumes?

When any answer is no, promotion is premature. The team may still have valuable experience to log. They do not yet have expertise to invoke.

Expertise decay and naming

Expertise can become stale when household life changes, when regulations shift, or when the catalog of meal candidates changes materially. Libraries need owners who watch guard failure telemetry and correction patterns. Decay is normal. Unowned decay becomes false familiarity. Scheduling periodic revalidation treats expertise artifacts like any other dependency. They compress judgment the way a cached configuration compresses setup work. Outdated configuration is a known failure mode. Outdated expertise deserves equal visibility.

Expertise artifacts also need names households never see but engineers can search. "Busy Week Pattern v0.2" is a good internal name: shape plus version. "Strategy7" is a poor name: untraceable in support tickets. Library searchability is part of expertise design. If evaluators cannot find the artifact, they cannot regression test it. When two artifacts overlap, documenting substitution preference in the library metadata prevents the runtime from guessing which pattern to try first. Expertise authors pair with an evaluator before promotion. Authoring without evaluation produces beautiful documents that fail on first contact with near-miss weeks.

What to do next

Drafting one expertise artifact for a repeated workflow in the domain, with all six fields above even as stubs, reveals how far the current fast paths are from actual expertise. Most teams discover that the fields expose gaps in guard design or fallback paths that informal shortcuts never made visible.

Applicability conditions covers guard design in depth. Experience becomes expertise traces the capture, evaluate, promote pipeline. Deliberation and automaticity describes the runtime effects when guards pass or fail.

Read next: Experience becomes expertise.