design
The `/design` skill validates whether a proposed product goal aligns with strategic direction by analyzing it against a PRODUCT.md file across five dimensions: gap alignment, persona fit, competitive differentiation, precedent, and scope fit. Use it before discovery begins or when pressure-testing user value to ensure planned work supports the product's core mission and identified roadmap gaps rather than pursuing tangential features.
git clone --depth 1 https://github.com/boshu2/agentops /tmp/design && cp -r /tmp/design/images/gemini/skills/design ~/.claude/skills/designSKILL.md
# /design -- Product Validation Gate > **Quick Ref:** Validates that a proposed goal aligns with the product's strategic direction before discovery begins major work. Council-gated with `--preset=product`. --- ## Quick Start ```bash /design "add caching layer for CLI" # validate goal against PRODUCT.md /design --quick "add caching layer" # inline check, no council spawning /design --strict "redesign hook system" # higher threshold (avg >= 2.5) /design # infers goal from recent context ``` --- ## Execution Steps ### Step 0: Check for PRODUCT.md Locate `PRODUCT.md` in the repo root. ```bash ls PRODUCT.md 2>/dev/null ``` **If absent:** Output a warning and return PASS with note: "No PRODUCT.md found -- skipping product validation gate. Run `/product` to generate one." **If present:** Continue to Step 1. ### Step 1: Load Product Context Read `PRODUCT.md` and extract: - **Mission** -- the product's core purpose - **Personas** -- defined user types and their needs - **Gaps** -- known product gaps or roadmap items - **Competitive landscape** -- how the product differentiates If any section is missing, note it as unavailable and score that dimension conservatively (1). ### Step 2: Score Alignment Matrix Evaluate the proposed goal against five dimensions. Use the scoring rubric in [references/alignment-matrix.md](references/alignment-matrix.md). | Dimension | Score (0-3) | Rationale | |-----------|-------------|-----------| | Gap Alignment | | Does this goal address a known product gap? | | Persona Fit | | Does this serve defined personas? | | Competitive Diff | | Does this strengthen competitive position? | | Precedent | | Has similar work been done before? What can we learn? | | Scope Fit | | Is this appropriately scoped for the current phase? | Compute the average score across all five dimensions. ### Step 3: Run Council Invoke council with the product preset. See [references/product-council-preset.md](references/product-council-preset.md) for judge configuration. ``` Skill(skill="council", args="--preset=product validate design alignment for: <goal>") ``` Pass the alignment matrix from Step 2 as context to the council judges. If `--quick` flag is set, skip council spawning and perform inline assessment instead. ### Step 4: Write Design Artifact Write the design artifact to `.agents/design/`: ```bash mkdir -p .agents/design ``` Filename: `<date>-design-<goal-slug>.md` (e.g., `2026-03-30-design-add-caching-layer.md`) Artifact contents: - Goal statement - Alignment matrix with scores and rationale - Council verdict (or inline assessment if `--quick`) - Final verdict and recommendation ### Step 5: Output Verdict Determine the final verdict based on scores: | Condition | Verdict | |-----------|---------| | Average score >= 2.0 AND no dimension at 0 | **PASS** -- goal aligns with product direction | | Average score >= 1.5 OR one dimension at 0 with others strong | **WARN** -- goal has alignment concerns, review before proceeding | | Average score < 1.5 OR multiple dimensions at 0 | **FAIL** -- goal does not align with product direction | When `--strict` is set, raise the PASS threshold to average >= 2.5. Output the verdict in this format: ``` DESIGN VERDICT: <PASS|WARN|FAIL> Gap Alignment: <score>/3 -- <one-line rationale> Persona Fit: <score>/3 -- <one-line rationale> Competitive Diff: <score>/3 -- <one-line rationale> Precedent: <score>/3 -- <one-line rationale> Scope Fit: <score>/3 -- <one-line rationale> Average: <avg>/3.0 Artifact: .agents/design/<filename>.md ``` --- ## Flags | Flag | Effect | |------|--------| | `--quick` | Inline check without council spawning. Faster but less thorough. | | `--strict` | Raise PASS threshold from avg >= 2.0 to avg >= 2.5. Use for high-stakes changes. | --- ## Reference Documents - [references/design.feature](references/design.feature) — Executable spec: goal-vs-PRODUCT.md council verdict before discovery, --quick inline, --strict threshold (soc-qk4b) - [references/alignment-matrix.md](references/alignment-matrix.md) -- Scoring rubric for the five alignment dimensions - [references/project-reality-check.md](references/project-reality-check.md) -- Quick reality-check prompts before committing to a project direction - [references/product-council-preset.md](references/product-council-preset.md) -- Council `--preset=product` judge configuration and verdict rules --- ## See Also - [skills/council/SKILL.md](../council/SKILL.md) -- Multi-model consensus council - [skills/product/SKILL.md](../product/SKILL.md) -- Generate PRODUCT.md - [skills/discovery/SKILL.md](../discovery/SKILL.md) -- Discovery phase orchestrator - [skills/pre-mortem/SKILL.md](../pre-mortem/SKILL.md) -- Plan validation gate
Switch a caller-selected coding-agent Triggers: "switch account", "rotate coding-agent account".
Use when operating ACFS flywheel health checks, init, and agent loop tooling from ~/acfs/bin/acfs.
Use Agent Mail as an optional messaging and Triggers: "coordinate writers", "reserve files".
Operate explicit orchestrator, implementer Triggers: "agent-native factory", "role-shaped agent panes", "persistent workers".
Run AGY headlessly via scheduled ticks or `agy -p`, capture agentapi JSONL evidence, and validate automated AGY loops or event streams.
|-
Use an explicitly selected AGY runtime for Triggers: "agy", "antigravity", "AGY evidence".
|-