jentic/jentic-mini/init-design

Skill: init-design

Set up and carry out a new design in docs/design/designs/<name>/. Follows the conventions established in platform-actors — a brief README index, problem statement, scope, architecture decisions, open questions, and per-concern sub-documents. Ends with scaffolded stubs for each planned document.

Trigger

Invoked when the user asks to start a design for a new feature, module, or cross-cutting concern. Examples:

• "start a design for X"
• "set up the design docs for X"
• "run init-design on X"

Steps

1. Clarify scope (ask the user)

Before writing anything, ask two questions in a single message:

a. Pull in related context?

Should I scan the repo for existing code, specs, or design docs relevant to <feature>? (Relevant modules, OpenAPI specs, existing ORM models, etc.) This shapes the problem statement and identifies constraints before you commit to an approach.

b. Documents to create?

Besides README.md, which concern-level docs do you want scaffolded now?

Suggest a default list based on the feature — the platform-actors design used:

actors.md, access-scenarios.md, endpoints.md, openapi-sketch.yaml, per-actor subdirs (each with README.md, auth.md, endpoints.md), authorization-server/.

Say which you want, or say "suggest" and I'll propose a list with rationale.

Wait for the user's answers before proceeding.

2. Pull in related context (if yes)

Search for:

• Existing source modules that the feature touches (grep for the feature noun in src/, openapi/)
• Existing design docs that mention the feature or its entities
• OpenAPI specs that will need extending
• ORM models and migrations relevant to the new entities
• Existing permission scopes in admin/core/permissions.py

Summarise findings in a short bullet list to the user before writing any docs. This becomes the raw material for the "Relationship to existing work" section of the README.

3. Write the README

docs/design/designs/<name>/README.md — the index and anchor doc for the design. Structure:

# <Name> — Design
 
One-paragraph purpose statement.
 
## Problem statement
What is broken or missing today? Why does it matter?
 
## Scope
### In scope
- Bullet list
### Out of scope (deferred)
- Bullet list
 
## Design documents
Table: | Document | Status | Purpose |
Status values: Draft | TBD | Settled
 
## Architecture decisions
Key settled decisions with rationale. Use ### subsections per decision.
Name each decision with what was chosen and why, not just what the alternatives were.
 
## Key questions
Open questions that must be resolved before or during implementation.
Table: | ID | Question | Resolves in |
IDs: OQ-1, OQ-2, … (sequential; carry over if any are already settled)
 
## Request / interaction flow (optional)
Mermaid sequence or flowchart diagram if there is a clear primary flow.
 
## Relationship to existing work
Bullet list: what this design extends, replaces, or depends on.

4. Scaffold sub-documents

Create each planned sub-document as a stub. Each stub contains:

• A # <Title> heading
• A one-sentence purpose statement
• ## TBD section with the key questions or concerns it will address
• A ## Reference section pointing back to the README and peer docs

Do not write content that is not yet designed. Stubs are intentionally thin — they establish the document's existence and purpose so the README table is not broken, and so each doc can be filled in incrementally.

5. Confirm with the user

List the files created (with one-line summaries) and the open questions that need answers before implementation can start. Ask:

Ready to fill in any of these documents now, or shall we commit the skeleton first?

Conventions (enforced throughout)

These come from feedback on the platform-actors design work:

• No padding. Every sentence must earn its place. No meta-commentary ("this section will cover…"), no restatements of what the heading already says.
• Resolve stale OQ refs. When an open question is answered, move it out of the open-questions table and into a ### Resolved subsection below it. Each resolved entry uses the format #### OQ-N: <full original question text> followed by a one-paragraph resolution summary and a link to the sub-document where the full detail lives. The full question is the heading — do not shorten it to a topic label. The open table stays clean — only genuinely unresolved questions; resolved ones stay visible with their answers. Add a matching ## OQ-N: <title> — resolved section in the sub-document itself.
• Relative paths only. All cross-document links use relative paths. Verify paths on file creation and on any file rename/move.
• Mermaid for flows. Use sequenceDiagram for request flows, flowchart TD for entity/registration graphs. No ASCII flowcharts.
• Status column stays current. The README's document table has a Status column. Update it as documents graduate from TBD → Draft → Settled.
• Scope = decisions + open questions. The design is done when every OQ is resolved and every in-scope item has a settled document. Nothing else.

Reference pattern (platform-actors)

The platform-actors design is the canonical example of this pattern in this repo:

docs/design/designs/platform-actors/
  README.md                         — index, problem, scope, arch decisions, OQs
  access-scenarios.md               — concrete end-to-end access patterns (input to auth design)
  endpoints.md                      — shared OAuth/identity endpoint surface
  openapi-sketch.yaml               — design-time OpenAPI sketch (not production spec)
  actors/
    README.md                       — index for actor-type subdirs
    actors.md                       — entity model: Users, Agents
    agent/
      README.md
      auth.md                       — DCR + JWT Bearer flow; token model; key rotation
      endpoints.md                  — agent-specific HTTP surface
    users/
      README.md
      auth.md
      endpoints.md
  authorization-server/
    README.md                       — OAuth surface ownership, pluggable IDP scope
    implementation-guide.md         — endpoints, concerns, risks, phased delivery

The README is always the entry point. All other documents are reachable from it, and all cross-references use relative paths.