CONDUCTOR — the adaptive entry point
Trigger: /pds <your idea>
The conductor is the one command that runs the whole cycle for you. Instead of remembering the
sequence of agents and the exact validation rituals, you describe what you want and /pds asks the
lane, then walks direction → scope → production → judgment, handing off to BOB, RAY, and ANALYZER
at the right moments. Production goes to Figma by default, to code only with modules.code: true.
It’s the answer to “how does anyone — junior or expert — use this stack?”
One dial: user_level
The conductor adapts to a single setting in STACK.md — user_level: junior | expert.
| junior | expert | |
|---|---|---|
| Narration | explains each step and why each gate exists | terse — agent signals only |
| Judgment | proposes 2–3 pre-argued options with a recommendation | waits for your decision |
| Vocabulary | glossed inline the first time it appears | assumed known |
Same system, two experiences. expert is the default — zero change for existing users.
A second dial, output: short | full, sets how much lands in the chat. short (default): the
result first, the file path, the decision you owe — the detail stays in the file. Say « détaille »
or pass --full for the long form once.
One phase, one conversation
At each gate crossed, the conductor writes agent-system/sessions/state_<feature>.md and closes
with → nouvelle session · /pds reprendre <feature>. The next phase starts from that file, not
from the old history — measured on our own cycles, the history re-read on every turn was ~75 % of
the spend; one phase per session brought it to 10 % in pulse run 6.
For a project with several features, /pds cadrer <project> runs a session 0 first: it collects
the context, writes project_<project>.md and one state file per feature, crosses no gate, and
ends on a dispatch plan — a title and a prompt per session, which you open yourself. A session that
runs long is relayed to a fresh one through the state file, never forked (ADR-017) — and it does
not have to watch the size itself: a hook injects the reminder at 120k and the relay at 150k (ADR-018). It also proposes the effort for each phase: medium by default, one step up when a
problem stalls twice, straight back down after (ADR-014).
What it does
- Bootstraps context. If
client_vision,roadmap, ordesign_guideare still template stubs, it interviews you and auto-detects your stack (framework, theme, fonts, tokens) to fill them — the real onboarding cost, removed. - Orchestrates the cycle. It calls
/bob --brief,/ray,/bob --build,/analyzeras-is, named handoffs at every step — never improvised — and manages the human gates. - Proposes judgment for newcomers. In junior mode it offers complete, pre-argued aesthetic directions instead of a blank brief — and explains each gate before asking you to cross it.
The conductor never crosses a gate for you. It proposes; you decide. It doesn’t alter the gates, the scoring, or the system prompts — it only adapts how they’re presented.
When to use it
- You’re new to the stack, or onboarding someone who is → run
/pds. - You want the whole cycle without juggling commands → run
/pds. - You’re an expert who prefers to drive each agent → keep using
/ray,/bob,/analyzerdirectly.
Designing in Figma instead of code? That track has its own guided flow — see /design-workflow.