agentic-loop-prompt-skill

An agentic skill for improving prompts, tool contracts, runtime controls, and evaluation plans for LLM applications that run agentic loops.

The skill is designed for agent runtimes that support loadable instruction packs or "skills" with a SKILL.md file and supporting references. It is not tied to one provider or framework. It covers Vercel AI SDK ToolLoopAgent-style loops, LangGraph/ReAct-style agents, MCP/tool-calling agents, browser/computer-use agents, code/shell agents, multi-agent systems, structured-output workflows, reasoning traces, verification/judge patterns, and agents with memory or RAG.

Purpose

Agentic loops fail in ways that ordinary prompt-writing advice does not catch: models call the wrong tool, repeat the same tool forever, produce valid JSON that is not usable, trust hostile web/email/file content, leak data through side effects, or keep working after the task is already complete.

This skill gives an agent a compact field guide for reviewing and improving that whole loop contract:

• what belongs in the system/developer prompt
• what belongs in tool descriptions and schemas
• what must be enforced by runtime code
• how to make completion observable
• how to prevent repeated no-progress loops
• how to use structured outputs and repair loops safely
• how to handle prompt injection and excessive agency
• how to expose reasoning traces without leaking private chain-of-thought
• how to add verifier models, consensus, postconditions, and evals

The core principle is:

Prompt for behavior; enforce invariants in code.

Repo layout

The skill itself lives in the improve-agentic-prompts/ directory of this repo. That directory is what you load into your agent runtime.

agentic-loop-prompt-skill/             <- this repo
|-- README.md                          <- you are here
|-- LICENSE
`-- improve-agentic-prompts/           <- the skill folder
    |-- SKILL.md
    `-- references/
        |-- prompt-contracts.md
        |-- tool-interfaces.md
        |-- structured-outputs.md
        |-- loop-controls.md
        |-- verification-and-consensus.md
        |-- reasoning-traces.md
        |-- security-boundaries.md
        |-- model-adaptation.md
        `-- evals.md

What it covers

A short top-level improve-agentic-prompts/SKILL.md routes the agent into focused reference pages:

Reference	Topic
prompt-contracts.md	Main system/developer prompts, authority order, trust boundaries, completion states, examples, and prompt anti-patterns
tool-interfaces.md	What belongs in prompts vs structured tool definitions, tool names/descriptions/schemas, results, provenance, prepare/commit, and tool anti-patterns
structured-outputs.md	Tool calls vs final output, JSON/schema reliability, constrained output, parser repair, and semantic validation
loop-controls.md	State-machine design, active tools, tool choice, explicit done states, stop conditions, no-progress detection, memory, approvals, and idempotency
verification-and-consensus.md	Usable-result checks, postconditions, verifier models, critics, self-consistency, debate, model juries, and escalation
reasoning-traces.md	Hidden reasoning vs visible summaries, provider thinking blocks, scratchpads, trace logging, context retention, redaction, and UI exposure
security-boundaries.md	Direct/indirect prompt injection, excessive agency, tool poisoning, memory poisoning, cross-agent risk, egress, secrets, MCP hardening, and approvals
model-adaptation.md	Adjustments for frontier, reasoning, small, text-only, open/local, RAG, multimodal, code/shell, multi-agent, realtime/voice, and provider-specific behavior
evals.md	Functional tests, adversarial injection matrices, tool-risk cases, loop cases, metrics, regression practice, and trace-based iteration

Installing

Install the nested skill folder, not the repository root.

For Codex-style skill installation, use the GitHub repo plus path:

install-skill-from-github.py --repo HamStudy/agentic-loop-prompt-skill --path improve-agentic-prompts

Or use a GitHub tree URL if your runtime supports it:

https://github.com/HamStudy/agentic-loop-prompt-skill/tree/main/improve-agentic-prompts

After installing, restart or reload your agent runtime so it picks up the new skill metadata.

Status and currency

This skill is framework-agnostic guidance, but provider and SDK behavior changes. For exact API names and current behavior, always check the live docs for the framework/provider you are using, especially around tool-calling, structured output, reasoning/thinking traces, and loop-control settings.

License

MIT. See LICENSE.