Authoring Guidelines: Prompts, Instructions, and Agent Instructions
Consolidated standards derived from existing instruction files, experimental workflow notes (
_thoughts/ai-documentation-workflow-experiment.md), and external best-practice reference (_thoughts/Prompt Templates Instructions for LLMs Labelvisor.md). This serves as a single authoritative Memory Bank entry for content creation consistency.
1. Scope & Purpose
Provide uniform rules for creating and maintaining:
- Prompts (
_prompts/) - Instruction files (
_instructions/& nested language/domain sets) - Agent-oriented instruction overlays (files containing
applyTo:directives or operational policies)
2. Placement & File Types
| Content Type | Location | Purpose | Example | |————–|———-|———|———| | Prompt | _prompts/ | Reusable interaction template | code-review-assistant.md | | General Instruction | _instructions/ | Cross-cutting rules (security, memory bank) | security-and-owasp.instructions.md | | Domain/Language Instruction | _instructions/<domain>/ | Language-specific practices | python/python.instructions.md | | Mode Definition | _modes/ | Interaction configuration/persona | advanced-coding-assistant.md | | Thought / Rationale | _thoughts/ | Experiments & design reasoning | ai-documentation-workflow-experiment.md | | Memory Bank Entry | memory-bank/ | Persistent canonical context | This file |
3. Naming Conventions
- Kebab-case for filenames (lowercase, hyphen separated).
- Suffix with
.instructions.mdfor instruction governance files. - Use concise verbs/nouns:
refactoring-python.instructions.mdnotpython-refactoring-guide-v1.md. - Avoid dates in filenames—use front matter or internal metadata.
4. Metadata & Front Matter
When applicable include (YAML or pseudo-front matter):
---
title: "Clear Title"
type: "prompt|how-to|policy|best-practices|reference|explanation"
difficulty: "beginner|intermediate|advanced" (if relevant)
audience: "maintainers|contributors|agents"
applyTo: ["glob or path patterns"] # only for agent-scope instructions
updated: 2025-10-08
---
Required minimal fields per type: | Type | Required | Optional | |——|———-|———-| | Prompt | title, purpose | examples, tags | | Instruction | title, type | applyTo, rationale, risks | | Agent Policy | title, applyTo | escalation_path, safety, validation |
5. applyTo Best Practices
- Keep patterns specific (e.g.,
scripts/*.pynot**/*.pyunless intentionally global). - Group related patterns; avoid duplicating across files—prefer one source of truth per scope.
- Document exclusions if non-obvious (“Not applied to generated test fixtures”).
- Periodically audit for dead patterns after restructures.
6. Prompt Authoring Principles
Derived from experimental workflow + external template guidance: | Principle | Description | Example | |———–|————-|———| | Clear Objective | State the task in one decisive sentence | “Summarize repository architecture…” | | Structured Sections | Use headings: Context, Task, Constraints, Output | Improves scanability | | Explicit Constraints | Token, style, format specs reduce ambiguity | “Limit to 200 words” | | Provide Examples | Few-shot guidance clarifies expectations | Example input/output blocks | | Role / Persona | Align model behavior | “You are a senior security reviewer” | | Deterministic Format | Enforce JSON/schema when machine-parsed | JSON schema snippet | | Progressive Refinement | Encourage iterative improvement loops | Step-by-step refinement prompt | | Avoid Ambiguity | Eliminate vague adjectives | Replace “optimize” with measurable target |
Anti-patterns:
- Overloaded instructions (too many unrelated tasks)
- Hidden assumptions
- Underspecified output format
7. Instruction File Authoring Principles
- One conceptual domain per instruction file (separation of concerns).
- Lead with intent & scope boundaries.
- Provide decision criteria (when to apply vs when not to).
- Include risk table for security- or performance-sensitive domains.
- Cross-link related instructions (“See also: refactoring-python.instructions.md”).
- Maintain changelog snippet if file drives critical process.
8. Agent Instruction Guidelines
- Be explicit about REQUIRED vs OPTIONAL behaviors.
- Define escalation triggers (e.g., “Ask user if confidence < 97%”)—reinforces current repo policy.
- Provide validation checkpoints (build, lint, tests, security scan).
- Include safeguarding rules (no secrets, no network exfiltration, confirm destructive ops).
- Specify output formatting requirements (e.g., heading levels, fenced code rules) to ensure predictability.
Checklist for agent-facing instruction files:
- Scope & intent
- applyTo patterns
- Required success criteria
- Error / fallback strategies
- Security constraints
- Update policy (when file must be revised)
9. Security & Compliance Highlights
From security-and-owasp.instructions.md principles (summarized):
- Never commit secrets / API keys.
- Use parameterization (avoid injection risks) in code examples.
- Validate & sanitize untrusted input paths.
- Reference OWASP Top 10 where relevant (e.g., for auth, injection, sensitive data).
- Encourage least privilege in examples (scoped tokens, minimal permissions).
10. Review & PR Checklist (Content-Specific)
| Item | Prompts | Instructions | Agent Instructions | |——|———|————-|——————–| | Clear objective | ✓ | ✓ | ✓ | | Proper metadata | ✓ | ✓ | ✓ | | Examples (if needed) | ✓ | Optional | Optional | | applyTo patterns valid | N/A | If scoped | ✓ | | Security considerations | ✓ | ✓ | ✓ | | Links to related docs | Optional | Recommended | Recommended | | Formatting/style passes | ✓ | ✓ | ✓ | | No secrets or PII | ✓ | ✓ | ✓ |
11. Quality Metrics (Initial Baselines)
- Acceptance rate of new prompts without revision > 70%.
- Instruction drift incidents per quarter: target 0 (tracked via PR review notes).
- Time-to-onboard (from clone to first successful script run): < 10 minutes.
- Review turnaround for instruction changes: < 24h.
12. Common Failure Modes & Remedies
| Failure | Symptom | Remedy | |———|———|——–| | Ambiguous prompt | Inconsistent model outputs | Add explicit format & constraints | | Bloated instruction file | Hard to apply selectively | Split by domain / concern | | Stale applyTo patterns | Instructions not triggered | Audit after structural changes | | Untracked rationale changes | Confusion in future PRs | Log in Memory Bank (activeContext) |
13. Future Enhancements
Planned improvements:
- Lint tool to validate front matter & applyTo existence in repo.
- Automated dead pattern detector (flags unmatched applyTo globs).
- CI job: markdown link + schema validation for structured prompt JSON examples.
- Semantic diff annotator for instruction changes (summarize added/removed rules).
14. Maintenance Policy
- Update this file whenever new content types or validation tooling is introduced.
- Reference it from CONTRIBUTING and PR templates for visibility.
- Log major revisions in
progress.mdandactiveContext.md.
15. Source References
_thoughts/ai-documentation-workflow-experiment.md(workflow, iterative prompting)_thoughts/Prompt Templates Instructions for LLMs Labelvisor.md(template structuring & clarity principles)- Existing instruction files in
_instructions/&.github/instructions/
Status: Authoritative baseline established (2025-10-08). Future updates must preserve section numbering for stable referencing.