vision-retrospective

name	vision-retrospective
description	This skill should be used when a user provides a vague, abstract, or "hollow" vision, goal, or idea and wants it turned into a concrete, actionable implementation path with verifiable acceptance criteria. It also triggers on mentions of "愿景回溯" (the Chinese term for vision retrospective), "vision retrospective", or requests to critique/refine a vision into executable steps. It contrasts with vision-driven development by tracing claims backward to evidence anchors rather than forwarding from assumptions.
agent_created	true

Vision Retrospective

Overview

Transform hollow visions into executable paths. Vision-Driven Development fails because it moves forward — the AI fills the gap between a vague vision and the code with plausible but ungrounded content, producing output that looks complete and is hollow. This skill moves backward: it treats the vision as a claim to be verified, anchors every claim to observable evidence, then decomposes only the anchored claims into an actionable path.

When to use

• The user shares a goal/idea that is abstract, adjective-heavy, or missing metrics, actors, actions, or success criteria (e.g. "a beautiful, seamless app that helps people").
• The user mentions "愿景回溯", "vision retrospective", or asks to refine/critique a vision into steps.
• The user is dissatisfied with vague AI output and wants a scientific, verifiable path instead.

The VR Loop (5 phases)

Run the phases in order. A healthy run shows a falling hollowness score (input → output); always report that delta.

Phase 0 — Intake & Hollow-Scan

Capture the raw vision verbatim (do not polish it yet). Run the deterministic scorer and produce a Hollowness Report:

python scripts/hollowness_scorer.py --text "<vision>"

Gate: if the score is Hollow or Fully Hollow (≥ 60), complete Phase 1 before any plan. Never jump from a hollow input to a roadmap — that is the VDD trap.

Phase 1 — Intent Excavation

Replace assumptions with grounded intent. Use 5 Whys, a Root Outcome statement ("When [this] succeeds, [actor] will be able to [observable capability] without [current pain]"), constraint surfacing, and inversion. Ask few, targeted, multiple-choice questions derived from the scorer's missing_dimensions and retrospect_prompts — never many open "tell me more" questions.

Phase 2 — Claim Decomposition & Verification Anchoring

Split the grounded vision into discrete claims (one testable sentence each). For every claim define a Verification Anchor: "What observable, reproducible evidence would prove this claim is true?" Translate residual hollow adjectives into measurable proxies using the scorer's suggestions. Apply SMART; time-bound is mandatory for anything to be built. Emit a Claim → Anchor table. A claim with no anchor is not a requirement yet — send it back to Phase 1.

Phase 3 — Backward Path Synthesis

For each anchor, ask what must exist for its evidence to be producible (the backward trace). Decompose into Epic → Feature → Task → Atomic Step. Every node carries acceptance criteria, dependencies, effort (S/M/L), and risk. Build a vertical MVP slice first (one end-to-end path, independently verifiable), then broaden by value/risk order — not by layer.

Phase 4 — Anti-Hollow Verification

Re-run the scorer on the final plan; a healthy plan scores Solid/Specific (≤ 35). Audit that every Atomic Step has a Done-Definition and no step contains an un-anchored hollow adjective. Emit the final Vision Retrospective Report and report the score delta.

Resources

scripts/hollowness_scorer.py

Deterministic (pure-stdlib) detector. Inputs: --text, --file, or stdin; flags --json and --fail-above N (exit non-zero if score > N, for CI gates). Returns a score (0–100), label, flagged vague terms with proxy suggestions, and missing verifiability dimensions. Use it in Phase 0 and Phase 4.

references/methodology.md

Full explanation of the VR Loop: the forward-vs-backward thesis, each phase's techniques, questioning discipline, anti-patterns, and a worked example.

references/foundations.md

Citation map proving each technique is academically grounded (Backward Design, IEEE 830 verifiability, GQM, BDD/Specification by Example, means-ends analysis, WBS) plus honest limitations of the heuristic score. Consult when the user asks whether the method is scientifically backed.

trae/ (cross-tool adapters)

Trae-native versions of the method so it can be invoked in any Trae conversation:

• trae/rules/vision-retrospective.mdc — Trae rule (Agent-Requested).
• trae/commands/vision-retrospective.md — /vision-retrospective slash command. Copy these into your Trae project's .trae/ or into Trae Global Rules to use the method outside WorkBuddy.

references/templates.md

Shapes for every artifact: Hollowness Report, Grounded Vision, Claim → Anchor table, Implementation Path, and the final Vision Retrospective Report.

assets/vision_retrospective_report.md

Fill-in-the-blanks scaffold for the final report; copy it into the user's workspace as the deliverable.

Anti-patterns

• Do not polish a hollow vision into prettier hollow words.
• Do not generate a plan before resolving intent.
• Do not accept "user-friendly" as a requirement — anchor it.
• Do not build by layer before a vertical MVP slice exists.