Blog vocabulary
Canonical terms used across posts on this blog. Written to keep one concept to one name, in English and in the Portuguese versions. Glossary only.
Personal AI alignment
Harness: Everything around a language model that a user controls: instructions, tools, memory, guardrails, the loop. Everything except the weights. Canonical 2026 usage, shorthand Agent = Model + Harness. Avoid: scaffolding, wrapper, setup
Personal alignment: Aligning an assistant to one person’s intent by shaping its harness, as opposed to lab alignment, which changes the model itself through training. Avoid: fine-tuning, training your AI, customization
Baseline:
The always-on half-lever: standing instructions that apply to every interaction.
Custom Instructions for consumers, AGENTS.md / CLAUDE.md for developers.
Avoid: instructions, system prompt, rules file
Procedures: On-demand context: a named, reusable recipe for one job, invoked only when the task calls for it. Saved prompts for consumers, Agent Skills for developers. Avoid: workflows, prompts, templates, commands, skills
“Skills” is deliberately not the name of this lever. It is one vendor form of it, and naming the pattern after its power form would make every consumer form read as a degraded skill rather than a legitimate procedure. Agent Skills is still named, prominently, inside the beat.
Memory: Accreting context: durable facts that grow over time, in two stages, capture then consolidate. Avoid: knowledge base, notes, history
Guardrails: The control half of the harness: limits on what the assistant may do, and the checks that catch it when instructions are not enough. Avoid: safety, permissions, constraints
Context engineering: The umbrella discipline of choosing what goes into a model’s context. Named once in the post as the industry term for the three context levers. Avoid: prompt engineering
Context rot: Measured degradation in model accuracy as the input context grows, well before the window is full. Industry term, backed by Chroma’s 2025 benchmark. Avoid: context bloat, drift
Dumb zone: Matt Pocock’s term for what context rot feels like from the user’s side (AI Coding Dictionary: dumb zone commonly begins ~125-150k tokens, debated). Introduced after the canonical term, attributed to Pocock.
Simpler is better: The through-value: the least context that does the job. Context is a budget, not a bucket. Avoid: less is more, minimalism
pt-BR equivalents
Decided for the Portuguese version (2026-07-27). English terms stay in English, glossed in Portuguese on first use: harness (glossed “arreio”), skills, baseline, Framing, Context Rot, Dumb Zone (glossed “zona burra”), guardrails (glossed “corrimãos”), prompt, handoff. Translated: Motivação (Motivation), Memória (Memory), Contexto (Context), Gestão de Memória (Memory Management), Privacidade (Privacy), Procedimentos (Procedures), O Efeito Composto (Compounding), Veja também (See also). The through-value renders as “quanto mais simples, melhor”. Quoted artifacts (CLAUDE.md excerpts, the grilling skill) stay verbatim in English, flagged as such in the prose, with a short Portuguese gloss after each.