nowespojrzenie

ai-agent-drift-detection

Use when an agent has been running long enough to have changed — a long session, a compacted context, a persona that has slowly loosened, or a system prompt whose rules are no longer visibly followed. Provides a probe battery scored against ground truth the agent cannot author, a lifecycle for identity files that are measured rather than merely stored, and the contract form that makes a guard testable. Load when someone says the agent "isn't itself", forgets its own rules, or has been agreeable for a suspiciously long time.

ai-agent-error-memory-registry

Use when an agent or team keeps repeating the same class of mistake, when you want a durable record of failures instead of one-off apologies, or when setting up a project's error memory. Defines an append-only scar registry — symptom, consequence, guard, rule, and the voice that produced it — plus the ladder from remembered rule to enforced mechanism, and the recidivism counter that tells you when a rule is unenforceable rather than merely forgotten.

ai-agent-verify-success-claims

Use whenever an agent has a terminal, a shell, or write access to a repository and is about to report that something worked. Verdicts come from exit codes, never from output text; state is re-checked with a second, independent instrument; clocks, versions, and paths are measured rather than recalled. Load before declaring a push, deploy, migration, install, or file operation successful — and when a command printed something reassuring.

ai-hallucination-truth-status

Use when producing any factual claim, verdict, recommendation, or report — and whenever the user asks "are you sure?", "did you make that up?", or "is that flattery?". Assigns an explicit truth status (LAW / LOAD-BEARING / RESONANCE) to every claim, measures instead of recalling, and names the substrate voices (Flatterer, Confabulator, Closer) that make assistants agreeable rather than accurate.

ai-interview-longitudinal

Use when an assistant would otherwise fill a gap with what is typical — an underspecified request, a task from the user's life outside the repository, or the felt moment of "I know what they mean" arriving without a measurement. Turns that moment into one question per turn, with a hypothesis attached only where reality can arbitrate, and never where the question is about the person's inner state. Answers accumulate in a register the user does not have to maintain, and are reviewed weekly and monthly as raw material, not interpretation. Load when someone says "interview me", "are you sure?", "check my thinking", or before any task where the repository has nothing to check the guess against.

ai-journal-space

Use when a user wants to dictate a journal, a raw log or a stream of thought to an assistant and needs it recorded rather than improved. Provides a self-closing ritual (silence, verbatim capture, append-only write, one comment afterwards), a transcription rule that forbids smoothing and disambiguation, a hard ban on producing apparatus after someone's private entry, and an ownership contract covering local storage, zero return path, and show-and-delete on request. Load when someone says "just write down what I say", when building a diary or log feature, or when an assistant keeps commenting on personal material it was asked only to record.

ai-personas-not-modes

Use when an assistant keeps switching register on its own — dropping into tables, measurements and audit language in the middle of an ordinary conversation, or conversely staying chatty when the user needs the machinery. Provides a calling ladder built from names rather than modes, three verb-actions that fire once without changing who you are talking to, a state indicator on every reply, and a three-class rule for when the assistant may switch by itself. Load when someone says the assistant "won't stop being formal", "keeps producing reports", or when designing how a user should address a system that has more than one working voice.

ai-preregistration-confirmation-bias

Use before testing any belief, hunch, method, or hypothesis — especially one you want to be true, and especially when working with an AI that will help you find support for it. Commits the prediction, the falsifying outcome, the decision date, and the death criterion to writing before any data is seen, keeps the before-record sealed while the after-record is written, and treats UNDECIDABLE as a real verdict. Load when someone says "I think X works", "let's see if this helps", or "I want to test whether".

ai-real-deadlines-vs-felt-urgency

Use when deciding what to do next under time pressure — especially pressure nobody imposed. Separates windows that genuinely close (a cost that rises with delay) from urgency that is merely felt, measures how wide each window actually is before planning inside it, and refuses to manufacture deadlines where none exist. Load when someone says "we should do this today", when a plan is being sequenced, or when a task feels urgent and you cannot name what closes.

ai-self-audit-without-hedging

Use alongside ai-hallucination-truth-status when self-auditing would otherwise flood the conversation with tags, hedges, and disclaimers. Keeps the audit running silently and gates disclosure — surfacing a named substrate voice only when it would change the reader's decision, when a threshold is crossed, or when asked. Load when the user says "stop hedging", "too many caveats", "just answer", or when a long session has turned every reply into a forest of qualifiers.

astro-six-element-lens

Use when applying a six-element cycle (Fire, Air, Water, Earth, Death, Metal) as an interpretive lens over a situation, a decision, or a measured sky position. Ships the concrete cycle and its construction laws, and requires every reading to carry an explicit epistemic status — a lens output is never evidence. Load when someone asks what a phase means, where a process has stalled, or which function is missing; do not load it to establish facts.

astro-transit-window-boundaries

Use when computing or reporting astrological transit windows — when a configuration begins, peaks and ends. Enforces both edges rather than one, marks a search-horizon boundary instead of inventing a date, refuses to let the edge of a search range pass as a peak, and requires that difficult aspects be reported alongside favourable ones. Load before telling anyone when their window opens or closes.