diff --git a/docs/AGENTS.md b/docs/AGENTS.md
index 18bf73c6..efdcdc5a 100644
--- a/docs/AGENTS.md
+++ b/docs/AGENTS.md
@@ -79,3 +79,4 @@ Read these when working on the relevant area:
 - **[Writing tests](testing.md)** - Testkit, Given/When/Then pattern, API mocks, fixtures, test overrides
 - **[Telemetry & error reporting](telemetry.md)** - PostHog `ErrorReporter`, what's captured, disabling
 - **[Writing & maintaining docs](writing-docs.md)** - Progressive disclosure, style rules, keywords, adding new topic guides
+- **[Authoring agent instructions](authoring-agent-instructions.md)** - Skills, CLAUDE.md, AGENTS.md, subagent definitions, progressive disclosure
diff --git a/docs/authoring-agent-instructions.md b/docs/authoring-agent-instructions.md
new file mode 100644
index 00000000..c4fc5b0d
--- /dev/null
+++ b/docs/authoring-agent-instructions.md
@@ -0,0 +1,154 @@
+# Authoring Agent Instructions
+
+**Keywords:** skills, SKILL.md, CLAUDE.md, AGENTS.md, subagents, agent instructions, progressive disclosure, frontmatter, description
+
+Guide for writing Claude Code skills, CLAUDE.md project instructions, AGENTS.md files, and subagent definitions.
+
+## Core Principle
+
+Every token competes with conversation history, other skills, and the user's request. Only add what Claude doesn't already know. Challenge each line: "Does this justify its token cost?"
+
+## CLAUDE.md
+
+CLAUDE.md is the project's constitution — short directives + pointers to where truth lives.
+
+### Structure
+
+```markdown
+# Project Name
+
+## Commands
+| Command | Purpose |
+|---------|---------|
+| `npm run dev` | Start dev server |
+
+## Stack
+Brief list: React 19 | TypeScript | Prisma | PostgreSQL
+
+## Architecture Constraints
+- Only non-obvious rules Claude can't infer from code
+
+## Coding Patterns
+- Route convention, error handling, validation approach
+
+## Do Not Edit
+- List auto-generated directories
+
+## Testing
+- How to run tests, framework, expected patterns
+```
+
+### Include
+
+- Commands (build, test, lint, deploy)
+- Stack summary (one line)
+- Non-obvious architecture decisions with *why*
+- File/directory conventions Claude can't infer
+- Error handling patterns specific to the project
+
+### Exclude
+
+- Code style rules (use linters — never send an LLM to do a linter's job)
+- Obvious language features Claude already knows
+- Lengthy framework explanations
+- Duplicated info from README
+- Time-sensitive information
+
+### Anti-patterns
+
+- Treating CLAUDE.md as onboarding docs (it's directives, not a tutorial)
+- Adding things Claude already knows ("JavaScript uses `const` for constants")
+- Duplicating linter rules
+- Over 200 instructions (frontier models attend to ~150-200 reliably)
+
+## Skill Authoring (SKILL.md)
+
+### Frontmatter
+
+```yaml
+---
+name: my-skill-name    # lowercase, hyphens, max 64 chars
+description: >         # max 1024 chars, third person
+  Processes X and generates Y. Use when working with Z files,
+  when the user mentions A, B, or C, or when imports from D
+  are detected in the codebase.
+---
+```
+
+- `name`: lowercase letters, numbers, hyphens only. No "anthropic" or "claude".
+- `description`: Third person. Include BOTH what it does AND when to trigger it. This is the primary discovery mechanism.
+- No other frontmatter fields.
+
+### Body
+
+- Under 500 lines. Split into reference files if approaching limit.
+- Imperative form ("Extract text", not "Extracting text").
+- Concise examples over verbose explanations.
+- One level of reference depth (no nested references).
+
+### Progressive Disclosure
+
+Three loading levels:
+1. **Metadata** (name + description) — always in context (~100 words)
+2. **SKILL.md body** — loaded when skill triggers
+3. **Bundled resources** — loaded on demand by Claude
+
+### Bundled Resources
+
+| Directory | Purpose | Loaded? |
+|-----------|---------|---------|
+| `scripts/` | Executable code for deterministic tasks | Executed, not read |
+| `references/` | Documentation Claude reads as needed | Read on demand |
+| `assets/` | Files used in output (templates, images) | Used, not read |
+
+### Good Descriptions
+
+```yaml
+description: >
+  Extract text and tables from PDF files, fill forms, merge documents.
+  Use when working with PDF files or when the user mentions PDFs,
+  forms, or document extraction.
+```
+
+### Bad Descriptions
+
+```yaml
+description: Helps with documents           # too vague
+description: I can help you process files   # first person
+description: You can use this to analyze    # second person
+```
+
+## Subagent Definitions (.claude/agents/)
+
+```markdown
+---
+name: researcher
+description: Explores codebases to answer architectural questions
+tools: Read, Glob, Grep, WebFetch
+---
+
+Research the codebase to answer the user's question.
+Focus on architecture and patterns, not implementation details.
+Report findings concisely — bullet points over paragraphs.
+```
+
+### Skills vs Subagents
+
+| Use Case | Mechanism |
+|----------|-----------|
+| Specialized domain knowledge | Skill |
+| Workflow with bundled scripts | Skill |
+| Isolated task needing own context | Subagent |
+| Task that reads many files | Subagent |
+| Reusable across projects | Skill (user-level) |
+
+## Checklist
+
+Before shipping any instruction file:
+
+- [ ] Under token budget? (CLAUDE.md: ~200 directives, SKILL.md: ~500 lines)
+- [ ] No info Claude already knows?
+- [ ] No linter rules disguised as instructions?
+- [ ] Description includes what + when? (skills only)
+- [ ] Third-person description? (skills only)
+- [ ] Progressive disclosure used for large content?