Chapter 05 · Vocabulary & Museum

The Codex

A shared vocabulary is navigation equipment — every term below is a labeled crystal the whole crew can pull off the shelf and mean the same thing by. The "first seen" column doubles as a timeline: read it top to bottom and you can watch the practice evolve.

Eggo the archivist holding up a glowing data crystal whose hologram links it to other cells in a wall of hexagonal vaults
every term // a labeled crystal
Eggo the announcer broadcasting glowing geometric shapes from a neon megaphone, three robots below catching matching shapes and lighting up identically

VVocabulary

Same word, same meaning, whole crew.

Teams that share precise terms move faster — not because jargon is clever, but because every shared term is a compressed agreement you never have to renegotiate. When someone says "clean handoff" or "no saved check, no claim", the whole crew lights up with the same shape.

Broadcast the terms until they land identically. Then stop explaining and start building.

TermDefinitionFirst seen
Context is KingEarly shorthand for making relevant evidence available to the agent.Mar 2025
Context engineeringDeliberate selection, compression, and presentation of agent evidence.Jun 2025
Scaffold → Outline → Instruct → ExecuteStaged workflow separating setup, planning, prompting, implementation.Mar 2025
Forced ContextManually curated search or file evidence supplied to the agent.Mar 2025
Three Pillars of Agent KnowledgePersonal knowledge, documentation, institutional/codebase knowledge.Jun 2025
Rule of ThreesGenerate ~three options or perspectives before choosing.Mid 2025
Print modeNon-interactive agent invocation that participates in shell composition.Aug 2025
Opus / Dopus / PopusHistorical wrappers for model choice, permission bypass, clipboard-seeded prompting.Aug 2025
Hypothesis ModeSmall, reversible try-fail-learn experimentation.Aug 2025
Compaction bailSave the useful artifact and restart rather than continuing degraded.Aug 2025
YOLO modeInformal name for bypassing permission checks; inherently risky.Aug 2025
Scripting 2.0Language-model judgment as a programmable runtime layer.Oct 2025
Diagram-Driven DevelopmentMermaid diagrams as persistent engineering memory.Nov 2025
Front-Stage / Back-StageDiagram split between user-visible behavior and supporting implementation.Nov 2025
SandcastlesCheap, disposable agent-built tools that must earn durability.Nov 2025
Progressive disclosureLoading specialized instructions only when relevant.Dec 2025
Council of FiveIndependent perspectives generated before a consequential decision.Dec 2025
TasteHuman judgment used to select among plausible alternatives.Dec 2025
Manager / workerContext-rich coordinator paired with narrow execution agents.May 2026
ImpA narrow, preloaded specialist agent or capability.Jun 2026
Browser BridgeBounded repo context sent through a browser model; local execution and verification.Jun 2026
OracleThe browser-routed external model used for broad planning, critique, synthesis.Jun 2026
No saved check, no claimPersistent external evidence required before asserting completion.Jun 2026

Vocabulary is a living surface: the newest terms (2026 column) are the most likely to shift as the practice settles. When a term and a tool disagree, trust the concept — names move faster than principles.

Names move faster than principles.— when a term and a tool disagree

Eggo the curator walking past stasis tubes holding retired gadgets, a compact modern tool glowing on his belt

MMuseum of Superseded Equipment

Sealed. Studied. Not reinstalled.

Preserved for study, not for use. Each exhibit is a hard-won directional change — something genuinely tried, that hit a wall, and got replaced. Knowing why the current gear won is what inoculates you against the next hype cycle.

RetiredCurrent gear
Large speculative rule libraries→Thin, reactive rules + progressive disclosure
Maximum-context dumps→Relevance density, preprocessing, fresh sessions
Unbounded agent fan-out→Bounded ownership; one–two supervised tracks
MCP everywhere→Mature CLIs first; MCP for real new capability
Trusting "done" or proprietary rewind→Git + saved tests, diffs, logs, screenshots
Model rankings as doctrine→Dated sidebars (see the --model stamp)
Permission bypass as convenience→Risky modes gated by isolation + verification
Agent prose as enforcement→Hooks and CI for non-negotiable invariants

The experiments that failed — container sandboxes, sprawling councils, unattended loops — stay valuable because their failures mapped the boundaries of autonomy. They are exhibits, not doctrine.

same meaning // whole crew