Lead agents without acting like a developer.
Agent Propeller helps non-developers turn messy requests into clearer, safer agent work across tools like Codex, Claude Code, Cursor, and Antigravity.
It is a manual-first skill and document system for people who want better outcomes from agent tools without becoming prompt engineers or shell experts. The goal is simple: help the user stay in charge of the work, understand what is being approved, and build better operator judgment over time.
Most agent setups assume at least one of these is already true:
- you know how to phrase the task in execution-ready language
- you are comfortable approving bash commands you do not fully understand
- you can tell when a change is risky, reversible, or under-verified
Agent Propeller is built for the gap in between.
It helps the user:
- clarify the ask before risky execution
- understand proposed commands before approval
- isolate or checkpoint work before meaningful risk
- preview verification before work starts
- see what was and was not actually verified afterward
- carry operator preferences and lessons into the next session
Agent Propeller is especially useful for:
- founders using agents across planning, docs, and execution
- operators and PMs who want more control without adopting deep developer habits
- researchers, writers, and students using agents for structured thinking and drafting
- anyone who wants to lead agents rather than blindly approve them
- Read SKILL.md to understand the operating rules.
- Paste one of the prompts below into your agent tool.
- Open the matching workflow or safety file only when the task needs it.
First prompt to try:
Read this request and help me clarify it before doing anything. Ask only the highest-value reverse questions, explain any risky command before I approve it, and tell me how you plan to verify the result.
Best follow-up files:
- vague request: workflows/request-clarification-loop.md
- risky command or approval decision: prompts/bash-command-explainer.md
- risky edits or uncertain automation: safety/workspace-isolation-playbook.md
- development task that needs honest testing: workflows/test-preview-and-verification-loop.md
The agent restates the job, surfaces ambiguity, and asks only 1-3 reverse questions that change the direction of the work.
Before broad or uncertain changes, the harness pushes for inspection, isolation, and recoverable checkpoints.
The agent previews how it plans to test or verify the result, then reports honestly on what was actually checked and what remains unverified.
For development work that could introduce security risk, the harness expects a visible security-minded review instead of stopping at functional correctness.
The repo includes memory patterns for operator preferences and agent routing so the next session can fit the user better.
Agent Propeller is a strong fit when:
- the request is messy, high-context, or under-specified
- the agent wants to run commands you did not explicitly ask for
- the work feels broad, risky, or hard to undo
- you want a verification plan before you trust the result
- you use multiple agent tools and want clearer routing between them
- you want to improve your operating habits, not just finish one task
This repo does not give you a UI product or hidden automation layer.
It gives you a practical operating layer made of:
- a reusable skill in SKILL.md
- clarification, verification, and reflection workflows in workflows/
- plain-language prompt helpers in prompts/
- safety playbooks and checkpoint templates in safety/
- operator-profile and routing memory templates in memory/
Read this request and help me clarify it before doing anything. Ask only the highest-value reverse questions.Before you run any bash command, explain what it does, why you need it, what it may affect, and whether I should isolate or back it up first.Do the task, but tell me how you plan to test it before you start, and report what you actually verified afterward.
Help me use Codex and Claude Code like a leader, not a passenger. Clarify my ask, explain risky commands, and keep the work reversible.Treat this development task as high-accountability work: preview verification, run a security-minded review, and tell me exactly what is still unverified.I want a harness that remembers when Codex is better for structure, Claude Code is better for implementation, and Antigravity is better for exploration. Apply that here.
I have a messy product request and I want to use an agent without sleepwalking into risky edits. Clarify the request, ask only the reverse questions that matter, explain any command before I approve it, keep the work reversible, and give me an honest verification report at the end.
Most agent-oriented repos optimize for automation, evals, or developer-centric control surfaces.
Agent Propeller optimizes for operator legibility.
That means the user should be able to tell:
- what the agent thinks the task is
- why a command is being proposed
- what might be affected before approval
- what the rollback path is before risk
- what was actually verified before trust
This repo is not trying to be a benchmark suite, a dashboard-first product, or an opaque safety wrapper. It is trying to be a practical operating layer that makes agent work easier to understand and easier to trust.
Use workflows/request-clarification-loop.md and prompts/request-clarifier.md when the ask is vague or high-context.
Use prompts/bash-command-explainer.md, safety/command-risk-rubric.md, and safety/workspace-isolation-playbook.md when the agent proposes meaningful action.
Use workflows/test-preview-and-verification-loop.md to make testing visible before and after work.
Use workflows/post-action-reflection.md, memory/operator-profile-template.md, and memory/agent-routing-memory-template.md to carry forward what the user prefers and what each agent is best at.
If you want the short version:
- Codex: best direct fit as a skill because this repo already follows the
SKILL.mdplus support-files structure - Claude Code: best adapted into
CLAUDE.md, imported memory, and reusable commands - Cursor: best split into focused project rules such as clarification, command explanation, and verification reporting
- Antigravity: best used as an operating playbook for clearer task framing, safer approvals, and visible artifact review
- SKILL.md: the main operating rules
- references/01-product-direction.md: product thesis and target user
- references/04-session-handoff.md: current state and next-step guidance
- workflows/request-clarification-loop.md
- workflows/test-preview-and-verification-loop.md
- workflows/post-action-reflection.md
- workflows/harness-update-loop.md
- prompts/request-clarifier.md
- prompts/bash-command-explainer.md
- safety/workspace-isolation-playbook.md
- safety/backup-checkpoint-template.md
- safety/command-risk-rubric.md
- memory/operator-profile-template.md
- memory/operator-profile-example.md
- memory/agent-routing-memory-template.md
Agent Propeller is currently strongest as:
- a manual-first operating layer
- a skill for clearer and safer agent collaboration
- a repo for tightening operator-facing workflows through small review loops
It is not yet:
- a UI product
- an automated memory system
- a benchmarked platform with proof across many real sessions
What the repo already proves:
- there is a coherent operating model for clarification, command explanation, reversible execution, verification, reflection, routing, and security review
- the model is inspectable as real skill files rather than only product ideas
What would make the repo much easier to trust in public:
- one realistic terminal transcript
- one before/after example of a messy request becoming a safer execution flow
- one development example showing verification preview plus security review
If you add only one proof asset next, make it a realistic before/after walkthrough.