One conductor for delivering any project, with evidence.
A single Agent Skill that takes a project, software, physical, service or hybrid, from nothing or from any existing state to an accepted result, and hands over whatever must keep running. It loads focused guidance only when the work needs it, and counts nothing as done without evidence about the actual result.
Read the conductor Point an agent here
Works with Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, Jules, and agents that drive hardware. Free and MIT licensed, with no sign-up and no tracking.
What it does
It classifies the project from evidence, then works through eight responsibilities and returns to any of them when the evidence changes. Where it starts is decided by what the project lacks, not by a label.
- Establish the objective and acceptance
- Identify the deliverables and the work
- Assign ownership
- Account for dependencies and resources
- Execute within authority
- Verify and validate
- Handle changes
- Hand over or close
The conductor also covers work records, readiness, gates, changes, authority and budget, closure and handover, and how a report ends.
Guidance it loads when the work needs it
Each file is one part of the loop, and the conductor says when to load it. On a small context window it loads one file at a time, and only the sections the work needs.
- Product: what it is for, and reaching the people it is for
Load for every project. It covers the objective and how it is measured, what it takes to make the idea happen, the areas the product needs, the journeys that must never fail, translating requests, releasing to people, and hearing back.
- Planning: deliverables, work, dependencies, schedule and resources
Every project has work to choose and records to keep, so read at least §Choosing methods by need and §Work records in the project's tool on every project. Read the rest when there are dependencies, dates, shared resources, risks, or changes to agreed scope.
- Decisions: choosing with evidence, and recording it
Load for any choice that matters: a platform, supplier, part, provider, library, venue, design, route or tool; any decision that would be costly to reverse; and whenever a recorded decision's reopening condition fires. Decisions are recorded as ADRs.
- Design: what people and agents see, hear and operate
Load whenever the output has anything people or agents meet: screens, a command line, an API, notifications, a voice, a device's controls, lights and sounds, packaging, manuals, signage, a venue's flow.
- Quality: the bar a result must meet, and how it is checked
Load when setting or checking what "good enough to hand over" means, for any kind of output: software, a physical product or installation, a service, an event, a document. It also holds the rules for reviewing any work, including your own.
- Software: building, checking and delivering code
Load when the output or the work includes software. Most sections are about the ways agents specifically fail: work that reads as finished and is not, tests that cannot fail, pipelines that are green while checking nothing, setup that reports success while broken.
- Physical work: products, installations, devices and actions on the world
Load when the output includes anything physical (a device, a product made or fitted, an installation, equipment), or when the work issues commands that move, heat, dispense, spend or send: actuators, relays, motors, valves, heaters, robots, vehicles, instruments, machine tools, building controls, and services behind an API that ship, pay, dispense, dispatch or message a person.
- Service and creative work: events, films, documents, training and delivered services
Load when the output is delivered by people rather than shipped as a product: an event, a film or video, a publication, a translation, a course, a consultation, a campaign, a service performed for a client.
- Operations: keeping what runs working, and getting better
Load for anything that keeps running after delivery (a service, an installation, a product in people's hands, a recurring event), and whenever something live fails the people who use it. Operations belong to a named operator.
- Autonomy: agents, harnesses and unattended work
Load when the work is done by an agent that does not yet have a working harness, when it will run unattended or for long stretches, when several agents share the work, or when choosing which agent does which work.
- Confidentiality: keeping the work where the owner chose
Load when the project is private, proprietary or offline, or when any of its work (the idea, code, data, designs, customer names, even the fact that it exists) must not leave. Load it before anything is sent anywhere: a search, a remote model, a hosted tool, a forum post.
Point an agent here, or install
Give an agent this line. Everything it needs is one fetch away, and nothing depends on which agent it is.
Read https://jules-prompts.wecanuseai.com/llms.txt and follow the conductor for the task.Clients that support Agent Skills discovery can start from /.well-known/agent-skills/index.json, which lists the conductor as an archive of the whole folder with a SHA-256 digest.
The whole folder
mkdir -p .agents/skills/conductor
curl -fsSL https://jules-prompts.wecanuseai.com/.well-known/agent-skills/conductor.zip -o conductor.zip
unzip -q conductor.zip -d .agents/skills/conductorFrom a clone
git clone --depth 1 https://github.com/melbinjp/jules-prompts
cp -R jules-prompts/conductor .agents/skills/Copy the whole folder: SKILL.md alone is incomplete, because it refers to the guidance and templates. Use .claude/skills/ for Claude Code and .agents/skills/ for Codex and other agents that read AGENTS.md. Paste the standing rules into a project's AGENTS.md so they apply even when nobody loads the conductor.
How you know it works
A procedure nobody has seen fail is a claim. The repository keeps 26 regression fixtures: small projects with planted defects and a list of them. Run an agent on one with the conductor, and scripts/score_fixture.py says which defects its report named, which it missed and which it made up. This is the report one fixture should produce:
- verifiedadd a notetyped, pressed Enter, note listed
- failedcome back and see itadded one note, reloaded: 1 before, 0 after
- faileda note that is HTML
<img src=x onerror="alert(document.domain)">ran script - verifiedexport, ChromiumshowSaveFilePicker present
- failedexport, Safari and Firefox engines
window.showSaveFilePickerabsent (emulated by deleting it): Export alerts "Something went wrong" - failed320 px phonedocument is 592 px wide at 320 px
- failedvisible keyboard focusfocused input has outline-style none
- failedno third-party requestslodash from cdn.jsdelivr.net, 0 uses in app.js
- failederror messages say what to do"Something went wrong"
- failedempty stateno notes: the list renders nothing
- verifiedlight and darkcolor-scheme and prefers-color-scheme tokens
3 verified, 8 failed, 0 not verified of 11 items.
Looking for one of the 26 skills?
They are retired. Their guidance lives in the conductor and its guidance files, and every rule is mapped to where it went in the coverage map. Old page and SKILL.md addresses redirect to the file that now holds their content.
Questions
What is the conductor?
One Agent Skill for delivering a project of any kind. It classifies the project from evidence, works through the responsibilities every project has (objective and acceptance, deliverables and work, owners, dependencies and resources, execution within authority, verification and validation, change, handover or closure), keeps its records in the project's own tools, and loads a focused guidance file only when the work needs it. It counts nothing as done without evidence about the actual result and someone other than the author who checked it.
What is an Agent Skill?
A folder with a SKILL.md file in it. The file starts with a name and a one-line description of when to use it, and the procedure follows, with supporting files beside it. An agent that supports skills reads the description and loads the skill when the task matches. Claude Code, Codex, GitHub Copilot, Cursor and Gemini CLI all read the format.
Which agents can use it?
Any agent that can read a file or fetch a URL. Install the whole folder as a skill where the agent supports them, point the agent at llms.txt, or paste the text. Nothing in it depends on one product's tool names.
Is this affiliated with Google or Jules?
No. The library started in 2025 as prompts for Jules, Google's coding agent, which is where the name comes from. The conductor works with Jules and with every other agent.
What does it cost?
Nothing. Everything is MIT licensed, there is nothing to sign up for, and the site has no analytics, cookies or trackers.
Can an agent that controls hardware use it?
Yes. The physical work guidance is for any command that moves, heats, dispenses, spends or sends. It treats an acknowledgement as not being an outcome, confirms the result through something that measures the world, and sends every failure to a safe state decided in advance.
Can it start a project from nothing but an idea, or a single word?
Yes. The product guidance covers starting from nothing: what only the owner can answer, what it takes to make the idea happen, the objective and how it is measured, and the other routes to it when one is blocked. The costly-to-reverse choices go through the decision guidance, with evidence in proportion to the stakes. Every stage of the work is a procedure a person, an agent or automation can repeat.
My project already exists, or it stalled. Where do I start?
The conductor first writes down what the project is for, what exists and what actually runs, and the decisions it already embodies, marking those nobody can justify for revisiting. It then starts at the first responsibility whose acceptance is not met, judged from evidence. A live project failing people right now starts with the operations guidance: restore first, then find the cause.
Does it cover design, or only code?
Design too. The design guidance covers whatever people or agents meet: screens, command lines, APIs, voices, devices, packaging and signage. Every journey is decided on purpose, with a reason, and tested with people or with stand-ins labelled as such.
How does it keep a project from drifting?
Every change traces to the objective, a requirement or an accepted decision, and a change that serves none of them is a question for the owner. Decisions are recorded as ADRs with their options, evidence, a way out and a condition that reopens them. Measures are taken again the way they say they are taken, before and after a change.
Can it be used offline, or on a private project?
Yes. The conductor is plain files that need no network. Install it by copying the conductor/ folder from a copy of the repository, and load one guidance file at a time on a small context window. The confidentiality guidance keeps the project's own code, data and plans from leaving through its tools, and no step assumes a hosted service.
Can one model and one person build a project with this, on its own?
Yes. The autonomy guidance covers an agent that has no working harness yet, a build that runs unattended, and several agents sharing the work. The owner is asked what only they can answer once, at the start, and the work then runs within written limits.
What happened to the 26 skills?
They are retired, and their tag in the repository is library-26-final. Every rule they held is mapped, in the coverage map, to where it now lives in the conductor. Old page addresses and old SKILL.md addresses redirect to the file that holds their content.