At publish boundaries — about to ship code, send a public-facing artifact, commit a foundational decision, send an email crossing a sensitive boundary, deploy to production, or any moment where the act is hard to reverse and external eyes are about to read what you produced — layer THREE orthogonal review types that catch THREE different bug classes. (1) Adversarial review (different reviewer than author, catches what author misses). (2) Session-code-review (author misses self-review of their own session's output). (3) Multi-agent review (different model families catch different bug classes). Distinct from flow-cold-read (single cross-substrate evaluator) and flow-consultation-dispatch (parallel multi-agent on a specific question) — this skill layers all three review types at the same boundary because each catches what the others miss. Completion pressure peaks at publish AND session-end boundaries — exactly when review matters most, exactly when the trained default is to skip review and just ship. Trigger ph
Persistent agent state (continuity files, project memory, identity documents, accumulated decisions, "facts" sections) drifts from reality over time as the world changes and old assertions go stale. Use this skill to dispatch a fresh-context adversarial reviewer against accumulated state with explicit reality-check questions, surface stale facts and outdated claims, then archive or update what's drifted. Distinct from cold-read (which reviews fresh drafts pre-commit) — drift-detection reviews STORED state that's been accumulating. Run on a cadence (weekly / monthly), after significant external state changes, or when "facts that were true 3 months ago" feel suspect. Triggers include scheduled drift-sweep, "audit the continuity", "what's stale in my project memory", "are my facts still true", or noticing a decision from N months ago is being acted on without re-verification.
When compressing observations, decisions, state, and relationships into a memory file, continuity document, decision log, or any artifact intended for both human scanning and future-agent loading, use FlowScript notation — a semantic marker system with required fields that forces precision while compressing density. Markers cover state lifecycle (?, thought:, ✓), waiting/parking/decided states with required rationale fields, relationship indicators (->, <-, <->, ><[axis]), and confidence/urgency modifiers (!, *, ~, ++). The required-fields constraint is load-bearing — you cannot write `[decided]` without rationale, so the notation forces better thinking. Use when generating compressed memory content for any agent, encoding architectural decisions, structuring active threads, or compressing session output. Hybrid prose/FlowScript is the default; switch to FlowScript when structure carries the meaning.
When you've drafted something substantial in one AI substrate (an analysis, synthesis, decision rationale, code review, architectural plan, essay, or any artifact about to be committed/shipped/sent), dispatch a single cross-substrate evaluator — a different model family from the one that drafted — to read it cold and surface what same-substrate review cannot see. Different training distributions catch different blind spots: motivated reasoning, over-claiming, retrieval-mistaken-for-synthesis, template-overreach, observability-suppression-by-prompt patterns invisible from inside the substrate that produced the draft. Use as a phase-gate before high-stakes commits, not as per-line polish. Multi-instance convergent flags from cold-read signal structural pivot — not chapter-by-chapter editing. $0 marginal cost when both substrates are accessible. Trigger phrases: "cold read this", "second pair of eyes", "what am I not seeing", "review before I ship".
When facing a non-trivial decision, code review, architectural choice, or any artifact about to be committed/shipped/sent, dispatch multiple independent AI agents in parallel for cross-substrate evaluation, then synthesize at the fan-in step. Each agent surfaces hidden assumptions before analyzing — that forced enumeration is where cross-substrate value lives. Convergence across different training distributions is structural signal; divergence is what's worth investigating. Use when the user says "review this", "get a second opinion", "consult on this", "sanity check", "what does the team think", or before any decision/commit that's hard to reverse. The human (or orchestrator) curates synthesis at fan-in — automating that step collapses the value back to a single-substrate echo chamber.
When facing a foundational architectural, scoping, framing, or design decision with high conceptual surface area — the kind of decision where you could spend weeks "deriving the right answer" before shipping anything — commit the shape early and iterate via adversarial review instead. Committed shape activates real critique machinery (consultation, partnership challenge, cold-read) that surfaces holes a solo-derivation pass cannot reach. Iteration against real critique beats iteration against imagined critique by a wide margin. Sibling of structural-invariants-beat-discipline patterns; this is scoping-as-structural-mechanism rather than scoping-as-willpower. Trigger phrases include "scope this", "shape this decision", "what's the right approach to X", facing a non-trivial architectural fork, drafting a foundational doc/spec, or when stuck deriving the theorem before shipping.
At the end of a meaningful work session, compress what happened into a structured continuity update using FLOW methodology — extract episodes (factual record), update state and top-of-mind salience, rewrite recent narrative, evaluate which developing patterns graduate to proven via citation-validated 3x rule, demote patterns that no longer hold under new evidence, archive temporally stale material, and commit. Use when the user says "wrap", "update continuity", "close out the session", or when a meaningful work unit completes (feature shipped, architectural decision made, multi-step research concluded). Produces compression-with-intelligence rather than appending-without-thinking — memory gets smarter (not just longer) as work accumulates.