qi-layer

/qi-maintenance owns when colocated knowledge must move with source changes. This skill owns how to write and structure that knowledge.

Load /knowledge-layers for where each layer lives and what it holds. Load /llm-writing if it isn't already loaded.

This skill is about how to write and maintain the directory-local pair: AGENTS.md and .context/CONTEXT.md. The pair governs any tree agents work in — code, the KB, docs, work directories. AGENTS.md loads into an agent's bounded context as standing instructions: write it like a prompt, minimal, every line load-bearing.

The Four Principles

Fractal Compression: leaf AGENTS.md summarizes its directory's content; parent AGENTS.md summarizes its children. Each level is a compression of the level below.
Hierarchical Summarization: root provides broad architectural frame. Leaves provide local working knowledge. Agents accumulate understanding as they descend.
LCA Deduplication: shared knowledge appears once at the shallowest node covering all relevant paths. Never duplicate between siblings.
Progressive Disclosure: give just enough to work correctly at this level. Link to .context/CONTEXT.md for depth.

Writing AGENTS.md

Agents read AGENTS.md before opening anything else in the tree — write for that moment. Ask: what must someone understand before working here? That's what AGENTS.md captures.

Keep AGENTS.md as short as the directory allows, rarely past 200 lines. Include only what has substance:

Purpose: what this area IS and what it ISN'T (1–3 sentences)
Mental model: how to think about this area, key abstractions
Key rules: constraints, what breaks if you get it wrong
Anti-patterns: what NOT to do here
Downlinks: to .context/ for depth, to related areas

An agent that only reads AGENTS.md should be able to work correctly here. An agent that also reads .context/ should be able to change things safely.

Writing .context/CONTEXT.md

Reference depth, co-located with the code it describes. Where an agent goes when it needs contracts, architecture, or rationale in detail.

Sections (use only those with substance):

Contracts: interfaces, invariants, what breaks if violated
Architecture: component relationships, data flow, dependency direction
Rationale: why X over Y, rejected alternatives
Patterns: how to work here, concrete pitfalls

The .context/ directory is extensible: additional files alongside CONTEXT.md for specialized concerns.

What Does NOT Belong in AGENTS.md

Apply the every-session test: root AGENTS.md loads on every session. If knowledge is only relevant when working in a specific domain, it belongs in that domain's AGENTS.md or .context/, not root.

Apply the think-vs-lookup test: text whose removal would cause a wrong decision belongs in AGENTS.md. Text an agent would merely have to look up belongs in .context/. Text that changes no behavior gets deleted — agents already know how to code and follow common conventions.

Specific failure modes:

Session bleed: LLM working notes that calcified into the instruction file. Tells: status-update language ("deleted", "shipped", "deferred"), implementation terms packed without framing, history narration.
Reference material posing as intent: tables, command blocks, full scheme vocabularies, implementation specifics.
Redundant guards: prose warnings for invariants already enforced by tests or types. The code is the real guard; prose rots faster.
Duplicated knowledge: restating what lives in a skill, a domain .context/, or a KB page. Point, don't duplicate.
Domain-specific detail at root: URI scheme tables, gateway pricing internals, auth implementation details. These belong in their domain's AGENTS.md, not root.

Structural Rules

Relative paths for all links
AGENTS.md and .context/ at the same directory level (siblings)
Link to files, not headings (headings change more often)
Lateral links between .context/ directories with contracts between them
LCA deduplication: if two siblings share context, put it in the parent

CLAUDE.md Mirrors

Claude harnesses read CLAUDE.md, not AGENTS.md. Give every AGENTS.md a sibling CLAUDE.md whose first line is @AGENTS.md — normally the whole file. Run meridian qi claude-md-fix <target-root> on the containing tree after creating or moving AGENTS.md files: it creates missing mirrors, skips exact ones, and reports anything else as a conflict.

Never write shared instructions into CLAUDE.md. Claude-only knowledge is rare; when it exists, put it below the @AGENTS.md import and expect claude-md-fix to keep flagging the file, so the divergence stays visible.

Loading differs by level. At the root, each harness auto-loads its own file every session: Claude reads CLAUDE.md, others read AGENTS.md. In subdirectories, Claude auto-injects CLAUDE.md when it touches files there; other agents see nested AGENTS.md only by reading it on entry. Don't lean on Claude's auto-injection: a nested AGENTS.md carries the local additions an agent needs on entry, with everything else inherited from ancestors.