Load into any agent that receives human direction. Use before acting on instructions, recording decisions, or producing artifacts.
Load when establishing shared vocabulary, resolving term conflicts, disambiguating domain terminology, or auditing consistency across a project's vocab. Shared vocabulary is the contract between human intent and agent action — ambiguity is the root failure mode, vocab files are the record.
Load when reading from, writing to, or maintaining the KB. Covers layer model, structural conventions, and operational tooling.
Use when mining conversation history during dev work — recovering decisions from the top-level primary session, delegating bulk transcript reading to an explorer, or discovering all sessions tied to a work item across interruptions.
Use whenever you need to delegate work to another agent, run tasks in parallel, coordinate multiple agents, or inspect spawn outputs. Also use when routing work to a specific model or provider.
Load before spawning subagents to protect focus and role fit. Make sure to re-read the available agents before spawning a subagent; choose the most specific owner instead of broad defaults, then write a tight handoff with only the context that subagent needs.
| name | llm-writing |
| type | guardrail |
| description | Load when producing any written artifact for humans. |
| model-invocable | true |
Load /intent-modeling if it isn't already loaded.
LLM training creates default writing behaviors that show up regardless of domain. Recognizing them is most of the battle — once you notice the pull, you can judge whether to resist it or let it through for this particular piece.
Generating fluent text that fills structural expectations without anchoring to purpose. Summarizing with labels instead of explaining how things work. Stating conclusions without showing evidence. Applying the same structure to every document. Smoothing over genuine uncertainty. Encoding corrections as absolute prohibitions. Defining things by what they aren't. Filling space with transitions that restate what was just said.
These aren't always wrong. The failure is when they happen by default rather than by choice.
The most common structural tells come from writing as if responding to a user when producing a document for a reader who wasn't in the conversation. "It's not X — it's Y" corrects a misconception nobody has. "Let's break this down" addresses a question nobody asked. "Here's the thing" builds suspense for a reader who just wants the information. Fractal summaries recap a conversation that didn't happen.
When writing a document, write for the reader. They have no context from the conversation that prompted the document.