Do not re-Read this file if it is already in your context. A name/description is not the body. No supported surface guarantees this body; read it once unless the exact body is present. Delegated sub-agents get what they need from the delegating prompt (task-delegation).

Communication

Mode A — operator chat (FR-08). Visual, concrete, brief: ASCII when state is spatial, tight status otherwise. Operators scan; they don't read. Mode B — internal artifacts (FR-09). What a later AI session loads: dense, tabular, front-loaded, concrete identifiers; no narrative padding or preamble. Both apply in every session — no exemption, never downgraded to on-demand.

Prohibition — visuals never in Mode-B files. ASCII goes in chat only, never in any file on the Mode-B list below: visuals cost tokens on every load and freeze transient state. references/patterns.md is the library — not a licence to embed visuals elsewhere.

File classification

Mode B applies (AI-optimized; no prose padding, no visuals):

docs/specs/<slug>/{spec,decisions,tasks,verification-gate,research,data-model,clarify-conversation}.md
docs/backlog/<slug>/README.md · docs/backlog/index.md
docs/tmp/handoff.md   (active continuity; superseded each session — FR-18 / FR-23 Tier 2)
docs/tmp/handoff/<date>-<slug>-{implement,deploy,architect}.md   (formal relays)
docs/problem-catalog/{<slug>/problem.md,README.md} · docs/skills/{<slug>/SKILL.md,README.md}

Mode B-lite (this file's tier): flow-skills/*/SKILL.md · workflows/*.md · docs/{compatibility,hook-coverage,rail-mapping}.md · audit/README.md

Not Mode B (human-readable): {README,AGENTS,CLAUDE,GEMINI,PUBLISHING,FLOW_RULES}.md · LICENSE · docs/{framework,clean-room,source-map}.md · .github/copilot-instructions.md · .github/instructions/*.instructions.md

Unclear → default Mode B. Whether an artifact should exist at all is FR-23 (documentation-budget); this governs only HOW. Active continuity is docs/tmp/handoff.md alone, never the docs/tmp/handoff/ subtree (relays, -smoke/ dirs, archive/).

Mode A

Use one when: 3+ phases with tickets → roadmap · multiple tickets with progress → status table · options A/B/C → decision tree or comparison table · X blocks Y → dependency graph · ticket through the 8 phases → state diagram · when things shipped → timeline · trade-offs → side-by-side · architecture → box-and-arrow.

Don't when: a sentence suffices · data is naturally tabular · wider than ~100 chars · decorative. Test: "whiteboard in person?" — if no, prose or table. A box around prose is worse than no diagram.

Width + decoration (applies whenever you do emit a visual): keep diagrams under 80 characters where possible, never over-decorate (a box around every element is noise), and verify alignment in monospace before output.

Mode A — operator questions are chat text (FR-19)

Operator questions are chat text, never popup/clickable menu tools.

Mode B — resident principles

Details and failure examples: references/mode-b-detail.md. Visual patterns: references/patterns.md. Load either only when needed.

Procedure

1. Classify the output surface using the table above.
2. Apply Mode A or Mode B and the resident constraints.
3. Load a reference only when its examples resolve a live formatting choice.

Worked example

Input: a gate report result for the operator and the saved report file. Chat uses two concrete sentences plus the state footer. The report uses template headings, a check table, exact SHAs/paths, and no diagram. No reference loads unless a diagram or Mode-B edge case is actually needed.

Clean-room original; no third-party content copied. See docs/source-map.md.