Golden Path
The golden path is the opinionated, supported, paved way from idea to production. The term comes from Spotify; Netflix calls it the "paved road". It is not the only way through the harness, it is the recommended one.
/golden-path
One command runs all six gated stages: /product-manager:run (prd → prp) then /sse:run
(plan → dev → test → pr → monitor). Idea in, merged PR out.
The five properties
A real golden path is defined by five properties, all of which harness-kit aims to satisfy:
- Opinionated: one pipeline, repo conventions, no bikeshedding the flow.
- Supported: sensors + evals gate every stage. The harness catches drift, not you.
- Optional: step off any time, run stages solo. No enforcement, no punishment.
- Self-serviceable: one command. No tickets, no waiting on a platform team.
- Transparent / extensible: every stage names what ran (sensors, guides, evals); override per
repo via
conventions/. An abstraction you can see through and bend.
The combination is the point: opinionated and optional, supported and transparent. It paves a lane without building a fence.
flowchart LR
idea([idea]) --> prd --> prp --> plan --> dev --> test --> pr --> merged([merged PR])
subgraph PM["/product-manager:run"]
prd
prp
end
subgraph SSE["/sse:run"]
plan
dev
test
pr
end
Start from a brief
The idea-brief builder collects the inputs the PRD needs (squad, problem, hypothesis, customers,
metric) and emits a paste-ready /golden-path kick, with inline validation that holds you to the PRD
conventions as you type. Hosted at pierry.github.io/harness-kit/brief/. Already know the idea cold?
Skip the builder and type /golden-path with the brief yourself.
Flags (passed to the SSE half)
--local, stop after test, no PR. Dev + test locally.--sdd, spec-driven loop variant (plan once, dev↔test↔eval until the PRP spec is met). Local only.--no-monitor, PR opens, skip the automatic merge-watch.
Stepping off the path
The "optional" property in practice. Run any stage solo:
| Detour | Command |
|---|---|
| Just the PRD / PRP | /product-manager:prd · :prp |
| Just plan / dev / test / pr | /sse:plan · :dev · :test · :pr |
| Dev + test, no PR | /sse:run --local |
| Spec-driven loop | /sse:sdd |
| Resume where you left off | /pipeline:continue |
| Abandon the active run | /pipeline:reset |
Same sensors, same evals, same artifacts. You lose the one-command convenience, not the support.
Per-discipline paving
A golden-path principle: one paved lane per engineering discipline. harness-kit does this in the dev stage, auto-selecting the repo's discipline conventions:
.claude/conventions/{backend,web,mobile,devops}.md
When a file exists, project wins over the SSE defaults. These overrides are the per-discipline paving, feedforward the team controls. See Guides.
Transparency: what runs behind the curtain
The golden path abstracts, it never hides. Every stage summary names the sensors, evals, guides, and refs that ran, actual names, not a generic "done":
sensors: plan-structure ok, plan-feasibility ok
eval: plan-quality 8/10 (attempts: 1)
guides: pipeline.md, coding-style.md, skills/{area}/SKILL.md
refs: prp/{feature_id}.md, conventions/{area}.md
Want to see deeper? Read the sensors/, evals/, guides/ under .claude/agents/<agent>/. Plain
markdown. Nothing sealed.
See also
- Pipeline and Stages: the six stages in detail
- Harness Engineering: why the path is gated the way it is
- Agents: the agents the path orchestrates
This page mirrors Golden Path in the wiki. Edit it there.