productize-all

Productize

Productize is an engineered distribution and runtime system for building, shipping, operating, and growing AI products inside your coding harness.

Three playbooks: /productize-0-1 (raw idea → shipped product), /productize-operate (production, incidents, error budgets), /productize-grow (aha-moment → retention → PLG). Nine gates that gate every move. One meta-runner. 238 routed skills underneath. Five host adapters. One workflow runtime. MIT.

Your AI coding agent — Codex, Claude Code, Cursor, OpenCode, Factory, or the next harness you adopt — is the harness. It writes the code, runs the tools, edits the files. Productize is the product team around the harness: thesis-review sharpens the bet, product-review locks the spec, design-review kills the slop, eng-review challenges the architecture, qa runs the evals, release owns the ship gate, docs catches drift, dx-review protects developer experience, and comms-review writes the narrative.

Without a harness, Productize is inert. Without Productize, the harness is undirected. It will write any code you ask, beautifully, for the wrong product.

Quick start

Install Productize into the harness you use:

npm install -g @productize/agent-skills
productize --host auto

Or install from source:

git clone --single-branch --depth 1 https://github.com/itseffi/productize.git productize
cd productize
./setup --host auto

Then ask your harness to build a product:

/productize-0-1
Build a Twitter clone. Make it secure. Have real AI agents use it.

Productize routes the request, fires the right gate at every move, produces the artifacts, edits the code through the harness, runs the evals, applies the fixes, and tells you what to decide next.

See it work

You:    /productize-0-1
        Build a Twitter clone. Make it secure. Have real AI agents use it.

Agent:  Before I write anything, I'm tightening the ask.
        A Twitter clone is the implementation shape, not the product.
        The product is an agent-native social arena: humans and AI
        agents can post, reply, follow, report, coordinate, spam,
        hit rate limits, and expose the trust and moderation
        problems that only appear when autonomous users are active.

        [Move 1 — Scope capability + curate data]
        [thesis-review + product-review fire — rate the premise, find
        three risky assumptions, ask one blocking question about the
        wedge]

        Beachhead: builders testing real agent behavior in social
        products. Wedge: a small social graph where agent accounts are
        first-class users and adversarial QA from day one. First
        capability: human accounts, agent accounts, posts, replies,
        follows, reports, rate limits, audit logs, seeded agents, and
        evals that score whether agents can use the product without
        breaking trust or safety.

You:    Approve. Continue.

Agent:  [Move 2a — Design]   design-review fires on the feed, composer,
                                profile, notifications, reports, and
                                moderation queue.
        [Move 2b — Architect] eng-review fires on auth, permissions,
                                agent identity, event logs, prompt
                                injection boundaries, content provenance,
                                rate limits, and abuse controls.
        [Move 2c — Spec]     product-review fires on acceptance criteria.
        [Move 2d — Build]    writes the first slice: accounts, agent
                                accounts, feed, posting, replies, follows,
                                reports, seeded agent scripts, and
                                browser-driven evals.

You:    /productize-autoplan

Agent:  [Runs thesis-review, product-review, design-review, eng-review,
        qa, release, docs, dx-review, and comms-review. Auto-applies
        safe decisions. Surfaces two product/security calls. You decide.
        It proceeds.]

You:    Move 3 — design the evals.

Agent:  [Move 3 — qa + product-review fire. Eval users defined:
        normal poster, reply farmer, spammer, brigader, reporter,
        lurker, and moderator. Ship gates cover signup, posting,
        following, reporting, moderation, rate limits, and abuse paths.]

        [Deploy gate — release fires. One block: no kill switch for
        seeded agents. Add it before running live evals.]

        [Move 4 — qa runs the evals in a browser. One failure:
        agent-generated text leaks unsafe instructions into notifications.]

        [Move 5 — eng-review analyzes the trace and finds the boundary:
        feed content is escaped, notification previews are not.]

        [Move 6 — design-review + qa apply and verify the fix: sanitize
        previews, add regression coverage, rerun the abuse-path evals.]

You:    Ship.