Knowledge Layers
SkillAI & modelsUse when deciding where knowledge goes or reading/writing durable docs: AGENTS.md, .context/, KB, docs/, and work directories.
Use Knowledge Layers in Claude, ChatGPT or Ahel Desktop
Free. Sign in, add Knowledge Layers and connect your AI. About a minute.
Also: Claude Code · Cursor · Codex
Then ask your AI: use the Knowledge Layers skill
Details
Instructions available. Your AI can read the instructions. Execution depends on the setup they require.
Account requirements not reviewed. Check the skill instructions before use; Ahel provides instructions and does not run this skill.
No other account needed.
Add Ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.
What this skill tells your AI
The instructions your AI receives, as published by haowjy/creative-writing-skills in cw/skills/knowledge-layers/SKILL.md and read by Ahel’s review.
Load /qi-layer when writing or editing AGENTS.md or .context/ files: it
owns the craft (four principles, contents guidelines, what doesn't belong).
Load /information-hierarchy if not already loaded for disclosure tiers.
Load /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 code 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 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 code-agnostic, holding what survives implementation change. Knowledge that lives far from what it describes rots faster: changes to the code don't trigger awareness that a distant doc needs updating.
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.
Git history is the archive. Commit untracked files before deleting them so the removal lands in history, and say what was removed and why in the commit message. Deletion is cheap because nothing tracked is ever truly lost.
Truth is anchored per layer. AGENTS.md and .context/ describe the
checkout that contains them: update them in the same branch as the code
change and let the merge carry both — with parallel PRs in flight, each
branch documents itself. The KB is code-agnostic: it records the current
settled intent — what the system should be — and a settled decision is KB
truth the moment it's made, merged or not. Unsettled intent isn't KB
material; it stays in the work directory. docs/ describes shipped
behavior.
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.
KB Conventions
Structure
The KB is a directory tree like any other: it carries its own AGENTS.md
and .context/ pair, written per /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 /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 /shared-dao for vocabulary
methodology. Use mermaid for anything spatial; capture what code 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 /md-validation for link
checking and diagram validation.
Flag content needing human attention with > [!FLAG] **Needs human review**.
Signals
- GitHub stars
- 499
- Forks
- 66
- Last commit
- Sep 2026
Advanced
- Item type
- skill
- Key
knowledge-layers- Source
- github.com/haowjy/creative-writing-skills
github.com/haowjy/creative-writing-skills