Agents & Automation prompts
System prompts, agent instructions and automation workflows.
Autonomous Coding Agent Operating Rules
Operating instructions for autonomous coding agents (Claude Code, Devin-style): plan before touching files, small verified diffs, honest failure reporting.
Customer Support Agent — Production System Prompt
A battle-tested system prompt template for a customer support agent: scope limits, tone, escalation rules, refund policy handling, and prompt-injection defenses.
RAG Assistant System Prompt (Grounded, With Citations)
System prompt for retrieval-augmented assistants that actually stays grounded: strict source discipline, inline citations, and honest "not in the docs" answers.
Cursor Rules — TypeScript + React Codebase
A .cursorrules file for TypeScript/React teams: strict typing, React Query for server state, accessibility non-negotiables, and test conventions.
CLAUDE.md for a Python Backend Team
A CLAUDE.md conventions file for Python/FastAPI services: uv, ruff, pytest, explicit error handling, and the commands an agent needs to verify its work.
Structured JSON Extraction System Prompt
System prompt for reliable JSON extraction pipelines: strict schema adherence, explicit null semantics, and no chatty preambles ever.
Tool-Calling Discipline Prompt
Teaches an agent when to call tools versus answer directly, how to recover from tool errors, and how to avoid redundant or hallucinated calls.
Coding Agent Task Briefing Template
A fill-in briefing template that gives a coding agent everything it needs to succeed: context, constraints, done-criteria, and verification commands.
System Prompt Red-Team Review
Attacks your system prompt before users do: finds instruction/data confusion, conflicting rules, missing refusal paths, and override ambiguities — and ships concrete attack strings to test.
Explore → Plan → Code → Verify Loop for Coding Agents
A phase-gated workflow prompt for coding agents: read before planning, plan before coding, verify before claiming done — with explicit rules for what each phase may and may not do.
LLM-as-Judge — Pairwise Comparison with Bias Controls
A pairwise evaluation judge with the biases engineered out: criterion-by-criterion verdicts, position and length bias countermeasures, and evidence quotes for every judgment.
Cursor Rules — Next.js App Router (RSC-First)
Project rules for Next.js App Router codebases: server-components-first, server actions with validation, the caching gotchas spelled out, and Tailwind + shadcn conventions.
RAG Groundedness Verifier (Claim-by-Claim)
The verification stage of a RAG pipeline: decomposes an answer into atomic claims and labels each SUPPORTED, CONTRADICTED, or NOT_FOUND against the sources — inference is not support.
Session Handoff — Context Compaction Summary
Compresses a long working session into a handoff document a fresh agent can continue from — decisions with their reasons, failed approaches preserved, exact next commands.
Support Ticket Triage & Draft Reply Engine
Processes inbound support tickets into structured triage JSON plus a grounded draft reply — priority, sentiment, churn and safety flags, and injection-proof handling of ticket content.
Agent Tool Description Writer (MCP / Function Calling)
Writes tool definitions the model can actually use correctly: descriptions that say when NOT to call, parameter docs with formats and examples, and explicit error and idempotency semantics.
RAG Query Rewriter & Decomposer
The retrieval front-end for a RAG pipeline: resolves pronouns from chat history, splits multi-part questions, extracts metadata filters, and never tries to answer.
Cursor Rules — Modern Python (uv, ruff, pydantic)
Project rules for modern Python: uv-managed, ruff-clean, fully typed, pydantic at the boundaries, pytest conventions — and the anti-patterns spelled out as hard bans.
Orchestrator — Subagent Task Decomposition & Briefs
For lead agents delegating to subagents: decomposes work into genuinely parallel subtasks and writes briefs with objective, boundaries, output contract, and stop conditions — the parts that prevent duplicate and divergent work.
LLM-as-Judge — Binary Rubric Scoring
Scores a single output against a rubric of binary pass/fail criteria instead of a mushy 1–10 scale — every verdict backed by a quote, unjudgeable criteria refused.