An Agent Skill that turns your coding agent into a technical design partner for the discovery phase of software projects — the architecture and design conversation that happens before Spec-Driven Development.
Works with Claude Code, Cursor, Codex, GitHub Copilot, and any agent that supports the Agent Skills specification.
Turns your agent into a technical design partner that:
- Explores the problem with you through structured brainstorming — one question at a time, never a questionnaire.
- Researches the state of the art with current, cited costs (web search, never stale memory) when the discussion touches volatile facts like vendor pricing or platform capabilities.
- Presents multiple approaches with explicit tradeoffs — honest pros, cons, and costs. Never a single recommendation without alternatives.
- Aligns decision by decision, encouraging "let me validate my understanding" checkpoints — and correcting wrong premises precisely, without condescension.
- Pushes back when it disagrees. If you lean toward a suboptimal option, it points to the technically correct path and explains why. Agreement that isn't earned is a disservice.
- Consolidates everything into an ADR-style decision document when — and only when — you signal that alignment is done.
Spec-Driven Development (Constitution → Specify → Clarify → Plan → Tasks → Implement → Validate) formalizes the cycle after you know what to build. This skill is how you discover what to build. Its output feeds SDD at two points:
| Skill output | SDD artifact |
|---|---|
| Engineering principles distilled from decisions | Constitution (AGENTS.md) |
| Context, decisions, and suggested vertical slices | Input for Specify (PRDs / feature specs) |
The final artifact is a set of consolidated Architecture Decision Records, with:
- Purpose preamble (the what and the why; PRDs record the how)
- Context, premises, and negative definition of the system
- Numbered decisions — each with rationale, discarded alternatives, and a reversibility note
- Out of scope, with discarded and deferred explicitly separated
- Known risks → mitigations
- Next steps framed as vertical slices (PRD candidates)
- Engineering principles as testable assertions, ready for a Constitution
A hard fidelity rule applies: the document records only decisions actually made in the conversation. Open questions are content (deferred items, risks) — never gaps to be "helpfully" filled in.
- Take over one-off technical questions ("how do I clear my git stash?") — those get a direct answer.
- Run when your decisions are already made and you just want implementation — that's the SDD flow, not discovery.
- Generate the document prematurely. Before you ask for consolidation, the conversation is the artifact.
# skills.sh CLI
npx skills add HiIamZeref/design-partner
# or GitHub CLI (v2.90+)
gh skill install HiIamZeref/design-partnerClaude.ai: upload the .skill file in a conversation and click Save skill.
Just start a design conversation — the skill triggers on intent, not keywords:
- "I want to brainstorm automating document data extraction for a mid-size company. How does the market solve this today?"
- "Should we route by document type here? PDFs to a cheap model, spreadsheets to a robust one?"
- "Let me validate my understanding: we chose X because Y, right?"
And when alignment is done:
- "We're done — give me the full review of everything we decided."
skills/design-partner/ the installable skill (SKILL.md + ADR decision-document template)
MIT