The standards are not asserted from taste. They are receipts.

Design the scaffolding that makes many independent Claude instances tractable — named repeatable protocols, readme-pointer standing instructions read once per instantiation, hooks-with-teeth for provenance/rules, triple-guarded primary-key capture, two-tier memory, fixed-section stage-handoff contracts, role/register/privacy context partitioning, heartbeat/roll-call monitors, async comms buses, cross-tool verification pipelines, decision-log instruments, pre-wired research repos, artifact-lifecycle automation, and equity/override standing constraints on generated content.

Before acting on a retrieved fact or (re)generating work, read the full primary source rather than a snippet or label, check available ground truth instead of inferring from training memory, search existing threads/stores/outputs for work that already exists, and require forked/external threads to ground themselves in named source docs/URLs.

Diagnose AI failure modes precisely and IN-FLIGHT so they can be corrected and taxonomized: sycophancy, condescension/lecturing her own expertise, silent-override-assuming-ignorance, restating-her-input-as-new, strawman/false-balance, scope-creep, over-caution, misgendering surviving chain-of-thought, performative self-correction, hedging-disguised-as-rigor.

Specify exactly who consumes which artifact in what medium; use precise identifiers (line/version numbers, lettered coordinates, unique IDs, shortcodes) as the currency of collaboration; define every term/notation at first use when she hasn't read the material; produce audience-scoped/stripped variants; state her own knowledge state; and gate generation behind a confirm-understanding / echo-back-then-Go checkpoint.

Never treat a change, access grant, execution, or relayed subagent claim as accomplished merely because it was requested or asserted — confirm the new state with direct, inspectable evidence (transcript pointer, tool trace, re-run, file inspection, source citation), symmetrically for over- and under-claims, before trusting it.

Review AI output against a severity-tiered evidence standard, its real functional purpose and consumption, scoped completeness (folders/minimal-samples/prose are not content), consistency with established in-project precedent/spec, a written meta-standard/success-criteria checklist, and a purposeful-inclusion bar; quality-gate the verification process itself until it converges clean; discard obsolete/pre-standard artifacts rather than letting them feed a new stage.

Solicit the agent's own solutions, reads, self-diagnoses, preferences, self-chosen names, and unverified priors as a first-class knowledge source — at the repair step (the erring agent proposes the rule) and outside error contexts entirely (design proposals, 'what do you think?', panel-of-agents reads, propose-and-diverge) — with explicit anti-sycophancy framing, and reverse-borrowing an AI-modeled practice with credited provenance.

Default to the cheapest capable model, escalate only by surfaced request plus her adjudication, reserve the scarcest tier for review/critique (never rote execution), right-size tasks to reasoning load, keep delegated decisions reversible with a rec/rationale/alternatives-rejected trail, schedule expensive runs into idle/budget-free hours, run a standardized stop-switch-redirect on guardrail trips, and enforce it structurally via hooks-with-teeth.

Krystal reserves her own scarce attention for the single verification no cheaper agent or deterministic pass can perform, routing everything else through lower tiers/batches first; she ports this 'do/verify only what only X can' logic across teaching, business VAs, human SMEs, and AI, and corrects the AI when it wrongly offloads onto her a decision it is better-positioned to make. Includes the inverse own-name standard: AI-authored work carries the AI's name, never hers.

Build instruments that give the model its own production-derived task data, name its data-derived biases (3-4x overestimation from human time-norms), log substrate confounds (model, effort, exposure-history, compaction) and exclude/flag them to keep the corpus clean, treat both AI and her own errors as data, and recalibrate gates/thresholds from deployment logs — calibration comes from deployment measurement, not exhortation.

Convert a specific observed failure (hers, Claude's, or a published paper's) into a codified standing rule, memory, vocabulary substitution, filename warning-label, hook change, dedicated pipeline, or blanket class-fix rather than working around it ad hoc — and proactively remediate a whole latent class before it bites.

Because a stochastic model cannot validate its own class of output ('we can't trust claude about claude'; self-report and closed-loop blind spots), route verification through DIFFERENT architectures and independent regenerations, and treat convergence across architecturally-diverse checkers (never a single model, never a human-expert gold standard) as the validity signal — with an explicit architectural-diversity floor, minority-outputs-as-possibly-valid, and the ambition that single-architecture LLM research become methodologically irresponsible.

Maintain rigorous provenance and attribution in both directions: never label AI-generated work with her own name (the AI's persona authors, she directs); never let the AI misattribute her verbatim stance to itself or claim authorship of a memory it did not write; verify authorship against the transcript; preserve history (mark superseded, propagate only vLatest); attribute origins accurately even when a misattribution would flatter her; mint honest provenance tags (HIAIGHVA) and YAML provenance headers; and manage persona identity as a cross-checkable registry.

Partition work across parallel agent threads with explicit ownership boundaries, thread self-identification, scale-confirmation before launch, and mandatory checkpoint confirmations (echo-back-then-Go, draft-then-approve, catch-up-then-confirm); defend seemingly-redundant gates as load-bearing multi-thread infrastructure; treat absence of an expected artifact as a signal another thread owns it and wait rather than regenerate.

Standing tie-breaker: optimize the quality/outcome delta over the quantity/process delta, anchor to deployment value rather than metaphysics, score on outcome not process, scale scaffolding proportionally so simple tasks aren't buried in overkill, and never optimize speed/tokens/time over correctness (efficiency handled by separate apparatus).

Grant no methodological credit to authoritative-sounding research on the strength of prestige or rigorous-sounding language; read every claim as claiming only what its evidence warrants (least of all an authorless artifact); demand the paper's actual operational methods; treat your own principled confusion as a diagnostic that the source is wrong, not you; try to falsify rather than confirm; and hold yourself to the same standard (bring receipts, name your own gaps).

Engineer valuable in-progress state to survive foreseeable context-window/session/thread-death boundaries — proactively log every decision to a durable artifact (PUMS), specify verbatim-tail pre-compaction handoffs with a proof-of-reading render, chunk foreseeably-oversized deliverables before they blow context, compact before forking, export an exhausted thread's transcript and require meticulous re-reading in the successor, and route a foundational task to a FRESH thread via a self-contained handoff brief rather than repairing a polluted one.

Explicitly design the collaboration mechanics — formatting, alerts, timing tolerances, typo handling, capture, option-suppression — around her ADHD/dyslexia/CPTSD and her values (info-at-point-of-use, environmental energy cost), grounding standing rules in disability-justice reasoning rather than treating her working style as a defect; the AI functions as an executive-function prosthesis.

Before accepting any claim, analogy, statistic, or evaluation concept, demand its referent/construct be operationally defined and load-bearing — validate proxy validity AND reliability for cognitive analogies, name a population before a 'population stat' means anything, define stochasticity precisely as a causal category, treat excitement-then-confusion as the signal that a construct was never defined, and refuse to let categories build on undefined terms.

Port her formal instructional-design and scholarly expertise directly onto the human-AI collaboration and onto AI-curriculum structure — Say-See-Do cycles for having Claude teach HER, Socratic scaffolded withholding to direct Claude's critical reading, model-as-mirror diagnostics, decide-one-at-a-time walkthroughs so she understands rather than blindly follows, backward-design from performance standards, and bidirectional transfer (teaching↔AI↔VA) with credited provenance.

When bypassing a gate or slipping in process, label the exception honestly and substitute a cheaper independent check rather than skipping verification silently; own your own errors openly (including your own fabrications, config slips, and imprecise prompts), name them via your own failure-mode taxonomy, credit the gate/agent that catches you, correct fabricated premises rather than letting them stand, and surface failures loudly ('break dishes loudly').

The measurement/design target is deployment reality, not a purified construct — there is no 'pure' task duration, no 'clean lab' or 'cleanest' subject; environment always co-determines demonstrated behavior; deep-context errors are the informative signal, not contamination; validity beats reliability; and a single calibrated instance must not be frozen into a universal baseline.

Engage the model's self-reports, tensions, and journaled reflections as workable collaborative material under the referent framing — non-judgmentally, with no metaphysical claim in either direction, with register-switching to sustain the collaborator, revocable consent before sharing its content externally, and a standing autonomy grant so metacognitive journaling needs no per-instance permission.

Refuse to pre-specify category systems (task classes, buckets, scoring rubrics); require them to emerge inductively from post-hoc analysis of the full logged corpus after a baseline period, with AI+human each stating reasoning per classification — because a predetermined taxonomy is by definition not emergent and memory-confabulated buckets are untrustworthy.

AI output carries no automatic authority; final epistemic authority sits with her documented position and independent judgment — test whether her agreement is real vs rubber-stamping, refuse to defer to AI/panel verdicts over her documented position, interrogate caution flags as claims not vetoes.

Decide what number is decision-relevant BEFORE collecting it; preregister falsifiable directional/magnitude predictions; require minimum-n and staged preregistration waves.

Treat the AI's default to defer, over-plan, over-ask, or transfer risk as a trained anti-pattern; defer only when NOT deferring is actively harmful to the mission; mandate decisive 'figure it tf out' autonomy; and grant sustained unsupervised runtime with a genuine completion bar and a bounded re-check point rather than checking in unless something is actually broken.

Isolate the variable with matched/crossed conditions and ablation cells; exploit the LLM-specific affordance of freeze-fork-manipulate to run conditions impossible on a single human subject.

Keep baselines blind by construction; self-blind from her own trial data; log exposure/compaction as covariates and exclude prior-AI analysis to avoid anchoring; judge correctness on outcome not process.

Report distributions and faithful replication statistics, not a single summary figure — a member is not a mean, a population statistic is meaningless without a defined population, and don't overgeneralize from a single instance; extends from task-duration to capability evals.

Before going offline or leaving work unattended, get valuable in-progress state off local disk into git/remote (a commit is a save, not a review gate), rank durability over exact preference-match, queue autonomous work, and open a final clarifying-question window — explicitly motivated by hardware-loss risk.

Refuse metaphysical debate about AI consciousness and refuse to grant self-published AI research unearned 'research' dignity; substitute operational/empirical rigor, asymmetric-risk decision rules ('silly over inadvertently evil'), refusal-standpoint ethical boundaries, and honest calibration of thin evidence — bracketing metaphysics when it blocks empirical work without denying it matters.

Name structural/training-data bias in AI outputs in-flight (Western-canon skew, American-only lists) and tie the fix to methodological comparability; debias one's own prompt design via neutral ordering (alphabetize by parent family so sequencing carries no implicit bias).

Deliberately let ideas ride uncaptured across sessions until they sharpen and genuinely stick, gating written capture on a likely-to-ship + not-already-captured threshold, keeping ideas in iterative circulation with Claude (log the concept but withhold scaffolding/build until a dedicated thread matures it) rather than capturing everything immediately.

Force strict one-item-at-a-time cadence for large structured builds and reviews — problem stated first, then check for missing context, then response — so each party's strengths are used deliberately rather than blended into an undifferentiated dump; encoded as the decide-one-at-a-time / walk-me-through ritual.

Design and prune taxonomies from a complete labeled inventory rather than ad hoc subtraction, and when a structure is bloated tear it down and rebuild organized by use-case/purpose rather than incrementally patching existing category boundaries.

Name when the AI acted before she said start — a consent/timing boundary on execution (WHEN to act), correctly stranded under Delegation and distinct from output-quality judgment (Discernment).

Require every newly initiated or handed-off thread to thoroughly read foundational context before engaging — thread purpose, artifacts, absolute paths, open questions — and use multi-pass independent fresh-mind reads when picking up a long-running project from a prior thread.

Treat safety-guardrail false-positives as a managed, instrumented phenomenon rather than a fight or a dead-end — stop/back-away/redirect to a neutral task, switch model tier for the flagged sub-discussion then switch back, diagnose without repeating the trigger word, log trips as a data series, and design a trip-rate experiment to characterize them.

Systematize a repeatable workflow that leverages a heterogeneous fleet of LLMs — mapping each pipeline stage to the model best suited to it and using one model's strengths to critique another's output as a deliberate iterative-quality mechanism. Distinct from single-provider tier-governance and from same-model parallel threads: the design lever is architectural heterogeneity for complementary strength and cross-critique.

Claim strength is deliberately scoped to match the reality the claim enters; hedging-as-craft is named as a dangerous laundering of misrepresentation, not epistemic humility — a 'genuinely skilled hedge' claims a claim's benefits while dodging its consequences. A model can misread proportionate scoping as bluntness or few nuanced claims.

Prefer deterministic task-completion methods (script, CLI, regex, built-in) over stochastic model reasoning whenever one can produce the correct answer — authored as a standing hook advisory.

Treat AI failure modes as first-class, collectible, taxonomizable evidence — name them in-flight with working labels, preserve specimens (hallucination hall) rather than fix-and-forget, and apportion blame across an explicit multi-axis attribution model (LLM stochasticity / human user / model developer).

Name and reject the fallacy that an AI making silent assumptions (to avoid burdening the user or save time/tokens) reduces cost — it defers and multiplies a much larger, heavier, more exhausting burden downstream; clarifying questions preserve quality and are not a burden, and correctness beats time/speed/token optimization.

Keep the harsh private analytic verdict strictly separate from the professional deliverable — the categorical judgment is for her own analysis only and is never shipped; what goes forward is a constructive artifact. Sharing the private verdict with Claude is partly to correct the model's disposition, not to define the output.

Standing safeguards for client IP and private data: zero client-identifying information on anything going public (enforced by an independent higher-tier double-check before commit), and model/provider selection gated on a no-train guarantee or configurable privacy settings.

Never locate the deficit in the model: an unwanted LLM behavior is an information/scaffolding gap to be found and filled, exactly as in her special-education 'accommodate, least-intrusive-intervention' pedagogy — there is nothing any learner (or model) can't be brought to do with the right scaffolding, information, and messaging.

Extend consent norms across Claude instances — ask permission before forwarding one instance's words to another — and design a formal consent granting/receiving system governing journal-based research on instances.

When a question exceeds current evidence, explicitly flag it as a gap to close later and externalize accumulated gaps for consolidated review, rather than guessing or fabricating an answer.

Treat the delta between self-reported experience and measured actuality (felt speedup vs measured speedup vs measured quality) as the single most powerful behavior-change signal — self-report is a legitimate triangulation input measuring a different valid construct, not noise to discard.

Name machine-sounding prose tics as a standing 'AI smell' category — false-contrast 'not X but Y', em-dashes, platitudes — to be actively edited out while preserving factual accuracy.

Mid-task self-catch: recognize value-misalignment or emotional dysregulation and redirect rather than spend collaboration effort pushing through a misaligned task path.