Agents Best Practices
A provider-neutral reference for designing, blueprinting, and auditing agentic harnesses: the model proposes actions, the runtime validates and executes.
Pure knowledge skill (markdown-only): no code execution, no permission requests, no external side effects. It itself systematically advocates least privilege, draft/commit separation, approval gates, and discovery-never-grants-authority, and enforces disclosure plus publication approval for public-board scenarios. Deductions: no third-party security review or verifiable source attribution; publisher identity unverified; the recommended external site (getpostingboard.dev) is not independently validated.
SKILL.md is internally self-consistent: trigger rules, mode boundaries, reference map, and output templates align, and it explicitly states which files load by default. Deductions: none of the referenced references/*.md files were provided for review, so path/content consistency is unverifiable; no tests or error-path examples. Static cap is 10; without execution, not awarded above 9.
Trigger conditions are exceptionally detailed: an activation-intent list plus explicit non-fit scope (ordinary single-turn writing/translation/Q&A), multi-industry audiences, and a clear MVP Builder Mode boundary. Deductions: content is English-only with no Chinese-language support statement; core value depends on overseas documentation links (OpenAI/Anthropic) with no mainland-China reachability note or alternatives.
Well-layered docs (SKILL.md entry plus progressive-disclosure references), version 1.7.0 metadata, MIT license, install instructions, layout map, source links, and a coverage-audit self-check file. Deductions: no changelog, no explicit maintenance ownership or update path, publisher unverified, and versioning is metadata-only.
Claims directly usable value via a concrete MVP blueprint template and structured answer format; the template is specific and reusable, offering clear marginal value over generic advice. Deductions: static review cannot verify actual output correctness/completeness; README example dialogues are author-constructed rather than verified run logs, and the core value depends on the quality of references not provided.
Includes official specification and vendor documentation links, a coverage-audit file, and a revision-pinned reference to a concrete Prime Agent implementation, so key claims are largely traceable. Deductions: static cap of 5; no test suites or third-party execution evidence; fact/inference separation rests partly on the author's own synthesis, limiting independent reproducibility.
- Static review only; nothing was executed, and the references/ directory was not included, so citation integrity is unknown.
- Content is English-only with no Chinese support; core follow-up reading relies on overseas sites (OpenAI/Anthropic); mainland-China users should verify reachability.
- Publisher identity is unverified by the registry; check recent commits and issues before adoption.
- The README-endorsed Get Posting Board is a third-party service; independently assess its security and compliance before using its API.
- Example dialogues are author-constructed demos, not verified run logs; real output quality may differ.
What does this skill do, and when should you use it?
This is a provider-neutral Agent Skill for designing, generating MVP blueprints for, auditing, refactoring, and explaining agentic harnesses. It covers provider-neutral architecture across OpenAI, Anthropic, and OpenAI-compatible APIs: agent loops, tool design, permissions, context compaction, memory, planning, MCP connectors, observability, and evals. The skill explicitly extends beyond coding agents to research, finance, legal, support, operations, and other domains. Content is a SKILL.md entry point plus 17 reference documents — pure Markdown with no script dependencies.
SKILL.md acts as the entry point: the agent first identifies the user's design problem across six dimensions (domain, autonomy level, risk level, state duration, tool surface, completion criteria), then loads only the most relevant reference files rather than everything. When the user asks to build a domain-specific agent, it defaults to MVP Builder Mode and produces a concrete harness blueprint from a built-in template (objective, core loop, tool registry, permission matrix, context/memory, safety, evals, launch path). It also defines three advanced modes — environment-adaptive tools, recursive/continual harnesses, and speculative tool execution — plus audit guidance, checklists, and curated official documentation links.
- A backend engineer needs the smallest production-safe agent harness for a new domain (e.g., account renewal risk) and wants a concrete MVP blueprint, not generic principles.
- A team has a working agent that runs tools forever or loses approval state after context compaction and needs a runtime-level audit with a fix order.
- A platform engineer must design narrow typed tools and approval policies for an ops agent touching Slack, Linear, and internal deploy APIs.
- An architect choosing between OpenAI, Anthropic, and OpenAI-compatible APIs wants side-by-side provider-neutral implementation patterns.
- An owner needs a pre-launch checklist for prompt-injection treatment, cache hit rate, cost, and eval strategy.
What are this skill's strengths and limitations?
- Provider-neutral: explicitly covers OpenAI, Anthropic, and OpenAI-compatible API patterns.
- Clear, enforceable stance: models never execute actions directly, every tool call gets a result, risk-classed permissions, draft/commit separation for high-risk effects.
- 17 topic-specific references plus checklists cover the full lifecycle from MVP blueprint to evals and incident response.
- Pure Markdown, MIT-licensed, no scripts or external dependencies — cheap to install and port.
- The repository ships no test suite or eval results validating the practices themselves.
- No executable code — everything is guidance text; implementation work remains with the reader.
- Advanced modes (speculative execution, recursive harnesses) are experimental, and the author himself mandates a measured single-loop baseline first.
- Beyond official documentation links, there is no quantified evidence from real production deployments.
How do you install this skill?
Three documented options: A. run npx skills add DenisSergeevitch/agents-best-practices -g (the -g flag installs globally at user level); B. prompt your agent to clone the repo into its skills directory (Codex: ~/.codex/skills/, Claude Code user-level: ~/.claude/skills/, project-level: .claude/skills/) and verify SKILL.md, icon.jpeg, and references/ are present; C. manually mkdir and git clone into those paths.
How do you use this skill?
Once installed, the skill activates when conversation touches agent architecture, harness design, tool permissions, planning mode, workflow orchestration, context and memory, skills/connectors, observability, evals, or prompt caching. Just ask directly, e.g., 'Audit our research agent's harness: it sometimes runs tools forever and forgets decisions after compaction.' Asking it to build an agent triggers MVP blueprint mode automatically.