knowledge-layers

Knowledge Layers

Load /creative-writing-skills:qi-layer when writing or editing AGENTS.md or .context/ files: it owns the craft (four principles, contents guidelines, what doesn't belong). Load /creative-writing-skills:information-hierarchy if not already loaded for disclosure tiers. Load /creative-writing-skills:llm-writing if not already loaded.

The Five Layers

Layer	What it holds	When to use it
AGENTS.md	Intent and mental model for a directory	First thing agents read on entry; what to understand before working here
.context/	Contracts, architecture, rationale (CONTEXT.md); deferred work (TODO, FUTURE)	Reference depth and tracked deferrals, co-located with the material they describe
KB	Cross-cutting decisions, domain concepts, patterns	Spans directories; outlives sessions; no single directory owns it
docs/	User-facing documentation	Different audience, different update cadence
Work directory	Temporary design decisions and scratch	Not colocated with durable content; scoped to an active work item

Placement Rules

If it's intent/mental-model for a directory → AGENTS.md. If it's directory-scoped depth an agent looks up → .context/. If it's deferred work scoped to this directory → .context/TODO (must-do) or .context/FUTURE (nice-to-have). Flat markdown lists — each entry names the affected path and concrete follow-up. Cross-cutting items or items needing external visibility get filed in the project's issue tracker instead. If it's big picture — cross-cutting decisions, domain concepts, higher-level system architecture → KB. If it's for end users → docs/. If it's the human correcting an agent → the KB corrections register (resources/corrections/register.md; code.md beside it for code projects). /creative-writing-skills:correct fixes the classes that repeat. If it's a temporary work artifact → work directory.

When in doubt, colocate. Knowledge that depends on one subsystem belongs in that subsystem's AGENTS.md or .context/ even when it reads like a concept; the KB stays independent of any one implementation, holding what survives a change to it. Knowledge that lives far from what it describes rots faster: changes to the thing don't trigger awareness that a distant doc needs updating.

The KB links to the work's own docs. Before writing to the KB about some work, read the docs, AGENTS.md, .context/, and comments the work updated. Link to what they describe, and write only what they don't say: the settled intent, the reason, the rejected alternatives, how it relates to the rest of the project. When they say all of it, the KB gets a link or nothing. Cut a KB section that restates them down to the link and the parts they don't say; the copy drifts the next time they change.

Current Truth Over History

Every durable layer — AGENTS.md, .context/, KB, docs/ — holds the best current understanding; work directories may keep intermediate reasoning while their work item is active. When durable content is superseded, delete it or (for KB pages) move it to archive/ (.kgignore'd, excluded from the knowledge graph). Live content never references archived content. Pages read as current truth, never narrate their own evolution.

Deletion needs no replacement. A page whose subject is gone, or that describes behavior the system no longer has, gets deleted on sight — even with nothing new to write in its place. Stale knowledge is worse than a gap: agents load it and reason from it. Deletion is the content-truth owner's call; structural roles flag instead.

History is the archive. In a version-controlled tree, commit untracked files before deleting them so the removal lands in history, and say what was removed and why in the commit message. Elsewhere, preserve what still matters in the project's archive before deleting. Deletion is cheap because nothing preserved is truly lost.

Truth is anchored per layer. AGENTS.md and .context/ describe the tree that contains them: update them in the same change as what they describe. The KB records the current settled intent — what the project should be — and an invariant is KB truth the moment the human settles it, done or not. Unsettled intent isn't KB material; it stays in the work directory. docs/ describes what has shipped to its readers. In a code project, load resources/code.md for how this maps to branches, merges, and shipped code.

Decision records are the exception: preserve superseded decisions in place when they explain why the system changed. Mark the old decision as superseded and link to the replacement.

Invariants and Current Choices

Record each settled point as one of two kinds. An invariant must hold however the work is done. A current choice is what the project is doing or trying for now. The test: if an agent found a better way that broke it, should it stop and ask the human who owns the project first? If yes, it's an invariant; if no, the agent changes it and updates the doc.

Default to current choice. The human's hedges ("for now", "let's try", "i think", "whatever is less work") mark a trial, and so does the human agreeing to something an agent proposed. An agent's own call stays a current choice until the human makes it an invariant. Quote the human with their hedges kept, and never record a trial as their ruling.

Write invariants as rules, in decision records or AGENTS.md. Write current choices as how things stand now, beside the work they describe. Measurements and the story of what went wrong go beside what they measure or explain, not inside a rule. A status line says only whether the thing is done, partly done, not done, or superseded; how it got there belongs in version control or the archive.

KB Conventions

Structure

The KB is a directory tree like any other: it carries its own AGENTS.md and .context/ pair, written per /creative-writing-skills:qi-layer — intent in AGENTS.md, governance depth (writing conventions, structure, validation) in .context/. Read the KB's own AGENTS.md before writing. Treat this skill as the cross-project default; the local KB guide wins for that KB. See resources/bootstrap.md for a suggested starting layout.

Wiki Page Conventions

Follow /creative-writing-skills:information-hierarchy. One concept per document; name files by what they describe (token-validation.md), not when (auth-redesign-notes.md). Cross-reference instead of re-explaining. Load /creative-writing-skills:shared-dao for vocabulary methodology. Use mermaid for anything spatial; capture what the files themselves can't easily tell you.

Operations

Ingest: new information enters the KB. Read the source, extract key knowledge, write or update wiki pages, and update indexes/cross-links touched by the change.

Maintain: keep the wiki current per Current Truth Over History above.

Lint: health-check the wiki. Look for contradictions, stale claims, orphan pages, missing cross-references. Use /creative-writing-skills:md-validation for link checking and diagram validation.

Flag content needing human attention with > [!FLAG] **Needs human review**.