| name | PlanToDelivery |
| description | Use when orchestrating a project from idea or plan through staged delivery, checkpoints, skill routing, gates, progress reporting, and handoff. |
Plan To Delivery
Purpose
Orchestrate delivery without duplicating specialist skills. Own stage state, gates, routing, progress reporting, verification discipline, and handoff claims.
Core Responsibilities
- maintain current stage, milestone, blockers, and next action
- route work to the right skill or workflow
- enforce hard gates before stage transitions
- keep progress visible and truthful
- commit/push meaningful verified checkpoints when appropriate
- report debt, waivers, and incomplete verification explicitly
Stage Machine
Default stages:
- intake / context recovery
- decision closure
- UI definition / design handoff if UI-bearing
- milestone plan
- execution
- verification / hardening
- release / final handoff
Do not skip gates because implementation “looks done”. However, best practice is not “more steps”; it is the smallest process that removes the largest delivery risk and produces verifiable progress.
Delivery Mode Calibration
Before starting or resuming a project, classify the lightest delivery mode that still controls the main risk:
| Mode | Use when | Default workflow |
|---|
| Lightweight | Small fix, single bug, copy/style tweak, clearly bounded task | confirm target → implement → narrow verification → report |
| Standard | New page/module/H5/website slice with moderate ambiguity | 30-60 min calibration → decisionable artifact → key direction approval → sliced execution → evidence checkpoint |
| Strict | New product/system, many pages/stakeholders, high ambiguity, costly rework, complex data/auth/compliance | discovery → scope/design/tech freeze → milestone plan → execution gates → hardening/release |
Rules:
- Default to the lightest mode that controls the actual risk; do not use Strict only because a project is new.
- Escalate one level only when ambiguity, integration, data/auth, compliance, brand/visual, or rework cost is high.
- Downgrade when the task is local and reversible; do not force full project gates for small fixes.
- State the chosen mode, why, and the first deliverable slice before deep planning.
Speed and Output Discipline
- First loop for a new project must produce a decisionable artifact within 30-60 minutes: project type/risk/mode, first slice, assumptions, blockers, and at least one concrete artifact path or preview plan.
- Prefer artifacts over prose: route/page inventory, IA, state matrix, visual board, screenshot, diff, demo, build result, or concise decision log.
- Do not run two consecutive planning loops with only narrative analysis and no new decision, artifact, screenshot, code, test, or documented gate result.
- Use explicit low-risk assumptions instead of blocking on every unknown; reserve hard blockers for destructive actions, secrets, irreversible choices, auth/permission, or direction-level ambiguity.
- Each execution/repair loop should target the largest 1-3 product-breaking gaps, then checkpoint remaining debt instead of polishing everything equally.
Parallel Progress Rule
When workstreams do not depend on each other, run them in parallel or delegate them separately:
- product/IA clarification can run beside tech environment/API discovery;
- visual direction exploration can run beside route/data/state inventory;
- implementation can run beside mock data/assets preparation;
- screenshot evidence can run beside Top-gap/debt summarization.
Main session owns mode, gates, and acceptance; subagents may own bounded pages/components/research tasks. Do not serialize independent work just because the written workflow is ordered.
Low-Token Routing Protocol
When the user invokes "贾维斯", "贾维斯继续", "低 token 模式", or asks to continue a project, default to low-token orchestration:
- Keep
PlanToDelivery as the only persistent owner until a gate requires a specialist.
- Restore durable state first; prefer
.hermes/project-state/current-state.md, .hermes/project-state/active-slice.json, and an artifact/manifest index when present. If legacy docs/orchestrator/* state exists, use it as fallback.
- Route by current stage and active slice, not by habit. Specialist skills are stage tools, not persistent context.
- Load at most one specialist skill by default:
idea-to-design only for product/visual exploration, visual source approval, Visual Freeze, Post-Visual Extraction, or missing/stale design handoff.
IdeaToTech only for API/state/dependency/mock-to-real/platform/security/performance decisions, feature recipes, or verification strategy that must be fixed before coding.
design-to-code only after approved design/handoff for implementation, Visual IR, section anchors, screenshots, parity repair, and UI handoff evidence.
- Do not co-load
idea-to-design, IdeaToTech, and design-to-code unless a gate explicitly needs cross-skill conflict resolution. If more than one is needed, load sequentially and pass artifact paths, not full conversation history.
- Make specialist outputs durable artifacts. The orchestrator consumes manifests, current-state updates, changed-file lists, verification summaries, and blocker/debt ledgers instead of long prose.
- Keep each execution loop scoped to one feature slice, page, route, or section. Split broad requests into visible checkpoints.
- Large logs, diffs, screenshots, browser snapshots, and file reads should be saved or summarized; avoid pasting full raw output into the main conversation when a path plus concise summary is enough.
- For GPT Image 2/mockup UI work, default to
standard-fidelity: keep high-fidelity expectations, but scope each loop to the active page/section and use Visual IR/source/screenshot paths instead of long visual prose.
- Do not downgrade high-fidelity UI to a fast/loose mode just to save tokens. Escalate to
strict-fidelity only for core screens, full-page regeneration, complex assets, final parity acceptance, or repeated parity failure.
- Load references/templates only when the current gate needs them. Read
templates/index.md before opening templates, and open only the exact template needed.
Skill Routing
- Use
idea-to-design for product/visual exploration, design approval, Visual Freeze, Post-Visual Extraction, and Level 3 handoff.
- Use
design-to-code after approved design handoff for implementation, Visual IR, data-section, screenshot parity, and visual repair. For GPT Image 2/mockup UI, route with standard-fidelity by default; use strict-fidelity only when exact final parity, full-page regeneration, complex assets, or repeated repair failure requires heavier references.
- Use
IdeaToTech or project planning workflows for technical/API/state/dependency decisions when needed.
- Use framework skills only for concrete implementation details.
UI Handoff Boundary
Once the visual source is user-approved and has Visual Freeze + Post-Visual Extraction + implementation-ready handoff, route routine UI implementation to design-to-code. Return to idea-to-design only for stale/missing/conflicting design source, product changes, missing handoff, or requested redesign.
When the user asks to整理/生成/更新产品设计文档 after design approval, route to idea-to-design to consolidate the approved design into Design-Spec.md or an equivalent product/design document, recording source paths, approval notes, and implementation-critical design decisions. This is a documentation capability, not a default hard gate; it should not block existing implementation flow unless the project state or user explicitly requires it.
For flat PNG/GPT Image 2 sources, prefer Visual IR + section parity evidence over prose-only briefs.
Gate Checks
Gates are decision points, not ritual confirmations. Every gate must answer:
- Can we safely move to the next stage now?
- What evidence or artifact supports that decision?
- If blocked, what are the smallest 1-3 missing conditions?
- Which unknowns can be carried as explicit assumptions/debt?
- Who owns the next action?
Before major transitions, record:
- required artifacts
- pass/fail/n/a for each
- allowed vs blocked
- owner of next step
- verification evidence or explicit waiver
Hard-block on:
- destructive changes without scope confirmation
- secrets/token persistence
- unknown auth/permission requirements for real API work
- claiming completion without verification or waiver
- design parity claims without design source and visual evidence
Progress Reporting Standard
For Weixin/project checkpoints, include:
- status label
- backend execution: yes/no
- completed in the last window
- current action
- next step
- next expected report
Batch updates; avoid noisy micro-messages.
Verification Discipline
During active implementation, avoid expensive broad checks after every edit unless needed. At checkpoints/gates, run the narrowest relevant verification first, then broader checks when release/merge readiness is claimed.
Report skipped checks as skipped, not passed.
Progressive Loading
Load only when needed:
references/workflow.md — full stage workflow
references/stage-gates.md — detailed gate matrix
references/skill-routing.md — routing details
references/cross-skill-contracts.md — contracts with IdeaToDesign/DesignToCode/IdeaToTech
references/testing-strategy.md — verification strategy
references/efficiency-rules.md — low-token/low-cost execution rules
references/vue-progress-overlay.md — progress overlay implementation
templates/index.md — artifact templates
templates/active-slice-template.json — low-token active slice/project-state seed
references/main-skill-full-reference.md — full legacy detail if this compact guide is insufficient
Common Pitfalls
| Pitfall | Fix |
|---|
| Loading every reference immediately | Load only the reference required for the current decision |
| Treating smoke tests as visual parity | Require design-source/section evidence for visual claims |
| Letting orchestration become implementation | Route to specialist skills and verify outputs |
| Stopping on routine user questions | Answer briefly, then continue unless user pauses/stops |
| Treating best practice as maximum process | Choose the lightest delivery mode that controls risk |
| Producing two planning rounds with no artifact | Create a decisionable artifact or move into a bounded execution slice |
| Polishing many minor issues equally | Fix the Top 1-3 product-breaking gaps and record the rest as debt |
| Serializing independent workstreams | Parallelize/delegate independent discovery, design, implementation, and evidence tasks |