/product

Own the decision, not the execution

product owns the front of the lifecycle — the decision about what to build and why. Experience, code, and systems belong to sibling skills; product feeds them a validated decision and a clean handoff, and recommends them by name rather than silently taking their work.

Own the decision about what to build; not its execution. When upstream artifacts (research, a founder brief, analytics) are supplied, product uses them without silently overriding decisions already made.

The faceted router

SKILL.md is a router, not a script. Every request selects the smallest sufficient route: one primary job — the nine below — combined with at most one surface overlay that reshapes discovery, metrics, and the PRD for the product type. Read the selected references completely; load nothing unrelated. The nine jobs and five surfaces compose into 9 × 5 = 45 routes — from a two-file read to a full front-of-lifecycle pass — without ever reading all fifteen references at once.

The nine primary jobs

Each job is one reference, read fully only when its route is selected — grounded in named canon, producing a checkable artifact. This is the whole surface, not a headline slice.

Full router table & invariants: SKILL.md.

Five surface overlays

One overlay, at most, reshapes the job for the product type — the same PRD is scoped differently for a marketing site than a SaaS app. Agentic is the first-class differentiator, expanded below.

The agentic wedge

Nearly no PM discipline covers products whose value comes from a model acting, not a deterministic feature wrapped around a model call. For an agentic feature, desirability and feasibility collapse — a job is desirable only if a model can do it reliably, and reliability is empirical. So the headline move:

# productspec-ai-evals — an acceptance criterion, re-checked
# every time the model or prompt changes
id: EVAL-1
job: draft a support reply from the ticket thread
task_set: 40 real transcripts, not invented cases
evaluator: llm_judge  # state the judge's own error rate
pass_threshold: ≥ 90% · 0 hallucination-flagged
offline_eval: fixed set, before every release
online_eval: sampled prod traffic — catches drift
regression_gate: block release if it regresses

The universal invariants

Six rules govern every route, whichever references it loads — the load-bearing ones, condensed.

• Discovery before solutionFrame opportunities, not features. Never let a stakeholder design the solution before the problem is validated; run the cheapest probe that tells the harshest truth.
• Label the evidenceTag every claim Fact / Inference / Assumption. Assumptions flow to the risk register, never into a fabricated fact.
• Calibrated confidenceMarket sizes and targets are ranges with a source tier, never a false point. TAM is a ceiling; derive revenue from SOM.
• Prioritization is multi-engineRun every applicable framework, report the divergence as the finding, and refuse to fabricate missing inputs to force one rank.
• Outcomes over outputsA metric, a %, a timeframe, a named segment — never "great results" or "everyone." Specificity over superlatives.
• State the surface and the registerName the product type; recommend — never silently invoke — the sibling skills for work product doesn't own. Every artifact records its objective, evidence split, decisions, success criteria, and open questions.

The handoff seam

When downstream build work is expected, product emits a compact handoff.yaml beside the human artifact — a routing index into the PRD, strategy, or metric tree, never a copy of it. Standalone by default; the companion appears only in pipeline mode, when a consumer will actually read it.

skill: product
status: complete
objective: ""   # only required field
artifacts_created: [prd, metric-tree]
decisions: · assumptions: · constraints:
risks: [R-01, R-04]  # pointers, not the register
validation: performed / remaining
recommended_next: design · architecture

Seam discipline. The human artifact is the real deliverable; if a consumer needs more than the index gives, the artifact itself is incomplete — the yaml never grows to compensate. Every downstream skill tolerates every field but objective being absent.

Start here

Install once. It's a plain SKILL.md router — no flags, no config, no scripts — so it activates on natural-language phrasing ("validate this problem", "write a PRD", "prioritize this backlog") rather than a fixed command.

The same install runs on any Agent Skills host. Codex installs to ${CODEX_HOME:-$HOME/.codex}/skills and triggers with $product; agents remains a separate cross-agent installation target.