human-in-the-loop-design

human-in-the-loop-design is a cursor rule for Cursor from mohitagw15856/pm-claude-skills. It costs 114 tokens per session (1,337 once invoked), scanned A, original, MIT.

A framework for deciding which actions an AI agent may perform automatically and which require human approval. It also defines the approval screen, escalation rules, and records needed to reconstruct decisions later.

In plain words
What is it for?
Use it to design agent approval workflows, classify actions as automatic or restricted, handle missing or delayed approvers, and specify audit trails.
Why use it?
It helps human review stay meaningful instead of becoming a routine stream of approval clicks, especially when actions affect money, data, permissions, or external users.

Cursor rule for Cursor

Written for Cursor: a Cursor rule (.mdc).

Good fit Use it to design agent approval workflows, classify actions as automatic or restricted, handle missing or delayed approvers, and specify audit trails.

Compare 6 cursor rules from other repositories ↓
Install with agentmods
npx agentmods add rules/mohitagw15856/pm-claude-skills/human-in-the-loop-design
About the project

PM Skills is a collection of plain-Markdown instructions that teach AI assistants structured methods for handling professional, personal, and life-admin tasks. People use it with Claude, ChatGPT, Gemini, Cursor, Codex, and other supported agents for work such as writing product requirements, reviewing documents, or planning difficult situations.

mohitagw15856/pm-claude-skills · 1,352 stars · on GitHub · mohitagw15856.github.io

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

Clone the repo
git clone --depth 1 https://github.com/mohitagw15856/pm-claude-skills

Made for: Cursor.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for human-in-the-loop-design

README.md
[![agentmods](https://agentmods.dev/badge/rules/mohitagw15856/pm-claude-skills/human-in-the-loop-design/github.svg)](https://agentmods.dev/rules/mohitagw15856/pm-claude-skills/human-in-the-loop-design)
Your own site
<a href="https://agentmods.dev/rules/mohitagw15856/pm-claude-skills/human-in-the-loop-design"><img src="https://agentmods.dev/badge/rules/mohitagw15856/pm-claude-skills/human-in-the-loop-design/github.svg" alt="Measured on agentmods" height="20"></a>

Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.

agentmods 80×15 button for human-in-the-loop-design

Your own site · 80×15
<a href="https://agentmods.dev/rules/mohitagw15856/pm-claude-skills/human-in-the-loop-design"><img src="https://agentmods.dev/badge/rules/mohitagw15856/pm-claude-skills/human-in-the-loop-design.svg" alt="Reviewed on agentmods" width="80" height="20"></a>
Per session 114 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 1,337 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. A grade says what 26 rules found in the file — not that it is safe.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5.1 $0.00114 $0.01337
Opus 5 $0.00057 $0.00668
Sonnet 5 $0.00023 $0.00267
Haiku 4.5 $0.00011 $0.00134

Measured 8d ago against content hash 83a144f1f515, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-11, from the pricing page.

Security

Grade A, and why

human-in-the-loop-design scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 8d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

exports/cursor/pm-agentnative/human-in-the-loop-design/human-in-the-loop-design.mdc · 78 lines

How it starts

The opening of the file, as written. The whole thing — 78 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Human-in-the-Loop Design Skill

The failure mode of agent oversight isn't too little review — it's review that decays into a rubber stamp. Forty approval prompts a day trains the human to click yes; then the one that mattered goes through with the rest. This skill designs the loop so human attention lands exactly where it changes the outcome, and nowhere else.

What This Skill Produces

  • An action-tier policy: every agent action classified auto / notify / approve / forbidden
  • An approval UX spec: what the human sees, batching rules, and the anti-rubber-stamp mechanics
  • Escalation & fallback rules: timeouts, absent approvers, disagreement
  • Audit-trail requirements: what gets recorded so any decision is reconstructable

Required Inputs

Ask for (if not already provided):

  • The agent and its action inventory — everything it can do (from its tool list, not its marketing)
  • Blast radius per action: reversible? outward-facing? money/data/permissions involved?
  • Volume estimates: how many times per day each action fires (approval load is a design constraint, not an afterthought)
  • Who approves — role, how many people, what else competes for their attention

Design Method

  1. Tier every action by consequence, not by feel. Two axes decide the tier: reversibility (undo in one step ↔ irreversible) and reach (internal draft ↔ external/financial/permanent). Then:
    • Auto — reversible + internal (drafts, reads, internal scratch writes). Log only.
    • Notify — reversible + modest reach (filed a ticket, updated a record). Do it, tell the human, easy undo.
    • Approve — hard to reverse OR outward-facing (send, publish, pay, delete, grant). Blocks until a human decides.
    • Forbidden — irreversible + high reach where the org has decided no automation belongs (auth changes, legal commitments). Not gated — absent from the toolset.
  2. Budget the approvals. Multiply approve-tier actions by daily volume. If the number exceeds ~10-15 meaningful decisions per approver per day, the design is broken before launch: move volume down-tier by adding reversibility (drafts, holds, delayed sends) rather than by lowering the bar.
  3. Design the approval moment against rubber-stamping.
    • Show the decision, not the transcript: what will happen, to whom, why the agent believes it's right, and what's unusual about this one.
    • Surface anomaly, hide routine: same-as-last-50 approvals batch into one digest; the outlier renders differently and alone.
    • Require typed engagement for the highest stakes (type the amount, name the recipient) — friction proportional to consequence.
    • Track approval latency and edit rate per approver: 100% instant approvals is a broken loop, not a good agent — say so in the metrics section.
  4. Write the escalation rules. Approver silent for [X]: action expires safely (never auto-proceeds). Approver rejects: agent gets the reason as context, may revise once, then stops. Two approvers disagree: named tiebreaker. After-hours urgent: the on-call path, or an honest "waits until morning".
  5. Spec the audit trail. Per gated action: what the agent proposed (verbatim), the evidence it showed, who decided, what shipped (diff vs proposal), timestamps. The reconstruction test: six months later, "why did this go out?" is answerable from the trail alone.
  6. Plan the tier reviews. Tiers loosen with evidence, not with comfort: an action moves down a tier after [N] consecutive approvals with zero edits and a human review of a sample. Tightening is immediate on any incident.

Read the full file on GitHub · 78 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 8d ago First seen · 78 lines · 114 tokens per session scan A 83a144f1f515

Subscribe to this mod's changes

human-in-the-loop-design is a cursor rule published in the GitHub repository mohitagw15856/pm-claude-skills (1,352 stars, last pushed 4d ago), licensed MIT. It adds 114 tokens to every session and 1,337 once invoked, about $0.0006 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-09-03.