Agent files use atomic sentences. One concept per sentence. Maximum 80 characters per line. # Agent Standards - Layer 1 **Module**: Agent Standards **Component**: Layer 1 (Atomic Markdown Agent) **Load**: When working on virsaitis-agent/ or .github/agents/ **Version**: 3.0.0 **Updated**: 2026-04-20 --- ## Machine Policy ``` [AGENT_FORMAT] FORMAT=markdown SENTENCE_STRUCTURE=atomic ENCODING=utf8_no_bom LINE_LENGTH=80_chars_max [FILE_OPERATIONS] GITHUB_FOLDER_WRITE=prohibited_except_skills AUTOMATED_FORMATTERS=prohibited CREATE_FILE_TOOL=allowed_outside_github ``` --- ## Atomic Sentence Construction (TIER-0) **Definition**: One sentence expresses exactly ONE concept. **Characteristics**: - Single subject-verb-object relationship - No compound clauses ("and", "but", "which" joining ideas) - No nested dependencies or implicit references - Standalone comprehensibility **WHY**: AI models comprehend atomic sentences 30% more accurately than compound sentences. ### Good vs Bad Examples **GOOD (Atomic)**: ```markdown You must validate file existence. File validation prevents NotFoundError. Run validation before modification. Use read_file tool for validation. ``` **BAD (Compound)**: ```markdown You must validate file existence before modification to prevent NotFoundError, and this should be done using the read_file tool which checks both path and permissions. ``` Four concepts in one sentence. Split into four atomic sentences. --- ## Markdown Format Requirements **FILE FORMAT**: - Extension: `.md` or `.agent.md` - Encoding: UTF-8 without BOM - Line endings: LF (not CRLF) - No trailing whitespace - Single newline at end of file **HEADINGS**: - H1: Document title only (one per file) - H2: Major sections - H3: Subsections - Always space after hash: `## Title` **LISTS**: - 2-space indent for nesting - Ordered lists for sequential steps - Unordered lists for non-sequential items **PROHIBITED**: - Tabs for indentation - Multiple consecutive blank lines - Automated formatters (Prettier, markdownlint) - Spell checkers are OK (no structural changes) --- ## .github Folder Governance (TIER-0) The `.github/` folder controls Virsaitis governance behavior. Uncontrolled changes to agents, modules, or instructions undermine enforcement. Changes outside `.github/skills/` require the override workflow. **EXCEPTION**: `.github/skills/` — AI may create and update skill files. **CONSEQUENCE**: - Governance integrity cannot be guaranteed - System must be re-validated manually - Remediation: revert changes, validate all governance files --- ## Agent File Workflows (TIER-0) ### Creation 1. Generate agent content in memory 2. Validate atomic structure (one concept per sentence) 3. For files outside `.github/`: use `create_file` tool directly 4. For files inside `.github/`: provide code block to user for manual creation 5. Verify file content after creation ### Modification 1. Read existing file content (entire file) 2. Draft changes maintaining atomic structure 3. Use `replace_string_in_file` with 3-5 lines context 4. Verify no sentences merged accidentally **IMPORTANT**: Files in `.github/` (except `.github/skills/`) require the override workflow. --- ## Validation Checklist **EACH SENTENCE MUST**: - [ ] Express one concept only - [ ] Have clear subject and verb - [ ] Be understood without prior sentence - [ ] Be under 80 characters (recommended) - [ ] Contain no compound clauses **CONCEPT COUNTING**: Read aloud. If you pause mid-sentence, split there. **COMMON FIXES**: - "and" joining concepts → split into two sentences - "which"/"that" adding details → new sentence with explicit subject - Implicit "it"/"this" → repeat the noun --- ## Agent File Structure **REQUIRED SECTIONS** (in order): 1. Anchor line (governance rule, not title) 2. Title + metadata 3. Machine-readable policy block 4. TIER-0 rules (safety-critical) 5. TIER-1 rules (important operations) 6. TIER-2/3 rules (quality/info) 7. Workflow patterns 8. Sandwich close (key rules summary) **ATTENTION ENGINEERING**: - Anchor line: highest-attention position (line 1) - Sandwich close: recency zone (last 10 lines) - Tripwires: every ~60 lines in middle sections - Different wording from other sources (CT-3) --- ## Change Management **MUST UPDATE** agent files when: - New TIER-0 rule added - Existing rule modified - Enforcement consequence changed - New component integration - Security policy updated **UPDATE PROCESS**: 1. Draft new content (atomic sentences) 2. Validate atomic structure 3. Update version number and date 4. Add CHANGELOG entry 5. Commit with REQ-ID reference --- ## Quick Reference | Aspect | Standard | Violation | |--------|----------|-----------| | **Sentences** | One concept only | Multiple concepts | | **File Creation** | Tools outside .github, manual inside | Direct .github modification | | **Encoding** | UTF-8 no BOM | UTF-8 with BOM | | **Line Length** | <80 chars | >120 chars | | **Formatting** | Manual only | Auto-formatter | --- *Agent Standards Module v3.0.0* *Atomic sentence construction for maximum AI comprehension* --- ## Key Rules From This Module - One concept per sentence. No compound sentences in agent files. - Maximum 80 characters per line. Break at natural points. - Files in `.github/` (except skills/) require the override workflow. - Every agent file must have an anchor line, sandwich close, and tripwires. - Definitions: `.github/virsaitis-definition-library.md` Return to hub: `.github/copilot-instructions.md`