SOURCE ON GITHUB ↗

ABOUT PRODUCTAGENT.DEV

Claude Code harnesses for product managers, free to read and run on your own Mac.

productagent.dev is a free, public library of 10 Claude Code harnesses for product managers and operators: skills, agents, shell scripts on a schedule, a permissions fence and a Telegram bot. The flagship, Loops, lets Claude keep working overnight and hands the decisions back to you in the morning. You give each loop a goal, a budget, a cadence and a model. Loops runs on your Mac and the Claude or ChatGPT plan you already pay for, works inside a fence (permission rules with the OS sandbox underneath), and leaves four plain lines per loop in the morning.

curl -fsSL https://productagent.dev/install.sh | bash

Who made this

Harrison Johnson, a product manager in New York. He has spent a decade building marketplaces and B2B software, as employee number one, as a co-founder, and as a product lead. He built productagent.dev with Claude Code and runs Loops on his own projects every night. Three pages come out of that work:

LinkedIn ↗ · GitHub ↗

He also sets up productagent for product teams and runs the first two weeks with them. Past clients include deepgreen.co. Message him on LinkedIn.

Who productagent.dev is for

  • Product managers who already run part of the day in Claude Code or Codex.
  • Founders and operators with more projects than evenings.
  • Anyone who wants an agent to keep working overnight without handing it the keys.

What you get: 10 harnesses

Each harness is a folder you can read before you run it. Open one to see every file.

  • flow-design (global): A progressive design interview: one question per screen, ending in a clickable HTML prototype of a linear user journey (onboarding, checkout, wizard).
  • loops (system): Give an agent a goal, a budget, a cadence and a model. It works one bounded run at a time while you sleep, inside a fence, and hands you four plain sentences in the morning.
  • prototype-swarm (global): Crawls a multi-screen HTML prototype for undefined click destinations and spawns one generator agent per missing screen, with a handoff packet written by the source screen, until the click graph is covered.
  • self-maintenance (system): The one way an unattended run may touch its own loop registry, and the one place a "need from you" becomes a decision you settle with a single command instead of a paste.
  • rootcause (global): User research synthesis: groups interview notes, support quotes, board screenshots and metrics by root cause, separates symptoms from causes, and drafts the questions to ask next.
  • show-dont-tell (global): Turns text-heavy docs and abstract ideas into accurate, code-drawn visuals: diagrams, charts, timelines, flows and comic strips that trace back to the source.
  • pm-strategist (global): A product and strategy advisor agent that blends three lenses (build taste, business structure, PM execution) and pressure-tests ideas instead of cheerleading.
  • todo (system): A TODO list for PMs in one markdown file, reachable from Claude Code and from Telegram: capture, see what's due, triage, and hand items to an agent that asks before it acts.
  • rev-intel-harness ↗ (link): A CSV of companies in, the people out: point your existing AI subscriptions at target accounts. Open source, MIT.
  • karma ↗ (link): A personal knowledge graph over everything the loops write: seeds in, relationships out, so a month of night reports stays searchable. Open source, MIT.

What Loops leaves you in the morning

## L-03 · Ship one fix
- Trying to: ship the "sort permits by county" backlog item.
- Did: added the county filter; the page renders in a headless browser.
- Decided: left the mobile layout alone; separate backlog item.
- Need from you: merge branch night/2026-09-14 if the screenshot looks right.

Four words run the whole system. A Loop is work with a goal that renews itself. A Run is one bounded pass at a loop. An Order is a one-off instruction for tonight. A Decision is anything a run could not settle; the run hands the decision back to you instead of guessing. Read the full concept page at harnesses/loops/README.md, and the runner and fence at MACHINE.md.

INSTALL LOOPS · MACOS · RUNS ON THE PLAN YOU ALREADY PAY FOR

I use
curl -fsSL https://productagent.dev/install.sh | bash

Puts the machine in ~/loops, notices that Claude Code is there, asks what to work on, how often, how much brain and what it may spend, then writes your first loop. Nothing runs until tonight, and only if the laptop is plugged in. Stop it any time with --off. Nothing phones home.

Don’t have Claude Code yet? Do this first.

Claude Code is the terminal version of Claude. Loops runs it for you overnight on a Claude Pro or Max plan, the one you already have. Open Terminal (press ⌘ space, type Terminal, press return) and paste:

curl -fsSL https://claude.ai/install.sh | bash

Then sign in once. Type claude and press return: a browser tab opens; sign in with the same account you use in the Claude app, then type /exit.

Want one skill, not the whole machine? Copy a skill folder (for example flow-design) into ~/.claude/skills/, or an agent file (for example pm-strategist) into ~/.claude/agents/. Restart Claude Code and the skill or agent shows up.

How productagent.dev is made

Harrison Johnson, a product manager, runs these harnesses on real projects every day and night. Each harness gets copied off Harrison's machine, scrubbed of personal paths, budgets and client names, and published here as a folder. Claude Code wrote most of the code and much of the prose; Harrison makes the calls on what ships and what the rules are. Every example on the site is synthetic. The one exception is the receipts: the real run record, numbers only, frozen at a cutoff date. The essay walks through what those nights taught.

The site reads the GitHub repository at build time, so productagent.dev and the repo always match. Want the case for loops in the words of the people building Claude Code? Read why.

Questions people ask

What is a Claude Code harness?

A harness is anything written down that tells Claude Code how to do a repeated job: a skill (a SKILL.md folder), an agent definition, a slash command, a permissions fence, a set of shell scripts on a schedule, or a whole bot. productagent.dev publishes the harnesses one product manager runs every day, as the folders they are.

Is productagent.dev free?

Yes. Every harness is a public folder on GitHub. Loops runs on the Claude or ChatGPT plan you already pay for. productagent.dev sells nothing and has no paid tier.

How do I install a harness?

Loops installs with one line in the macOS Terminal: curl -fsSL https://productagent.dev/install.sh | bash. For a single skill or agent, copy its folder into ~/.claude/skills (skills) or its file into ~/.claude/agents (agents), then restart Claude Code.

What is Loops?

Loops lets an agent keep working on a goal while you sleep. You set four knobs per loop: a goal, a budget, a cadence and a model. Each night the runner does one bounded run per loop inside a fence, then writes four lines per loop: what it was trying to do, what it did, what it decided, and what it needs from you.

Is it safe to let an agent run unattended?

Loops is built around that question. Each run has a dollar budget and a timeout. A fence of Claude Code permissions allows a short list of actions and blocks git push, rm, shells and edits to the loops file, and the OS sandbox underneath stops what text rules miss, such as a one-line python3 -c. The runner keeps its own list of every blocked call and prints the list under the morning report, so the agent cannot hide what it tried. Nothing runs until you write a loop, and running the installer again with --off stops the schedule.

Does it work with Codex or ChatGPT?

Loops runs Claude Code or Codex on your own Mac, on your own subscription. The skills and agents target Claude Code.

Does anything leave my machine?

No. Loops runs locally, writes its reports to local files and sends no telemetry.

Follow along

ABOUT01 / 10

drag to scroll