dev-design
/dev-design
Use when given a large, multi-part feature/refactor/initiative idea (often a pasted proposal or brain-dump) that needs to be understood, questioned, and broken into sub-projects, each captured as a PRD (a short PM-spec + architecture doc, design only). Stops at approved PRDs; task generation is a separate step (dev-tasks).
- Author
- tejitpabari99 ↗
- Repository
- tejit-plugin ↗
- Licence
- MIT
- Last repository change
- 10.11.2026
Assessment and security
- Files in this package
- 2
- Scripts
- 0
- HTML comments
- 0
- URLs in SKILL.md
- 0
No signals detected. Repository changed on 10.11.2026, without references to retired models or APIs.
MIT, open: use and modify while retaining author attribution.
What it does and when to use it
Use when given a large, multi-part feature/refactor/initiative idea (often a pasted proposal or brain-dump) that needs to be understood, questioned, and broken into sub-projects, each captured as a PRD (a short PM-spec + architecture doc, design only). Stops at approved PRDs; task generation is a separate step (dev-tasks).
Original author description
Use when given a large, multi-part feature/refactor/initiative idea (often a pasted proposal or brain-dump) that needs to be understood, questioned, and broken into sub-projects, each captured as a PRD (a short PM-spec + architecture doc, design only). Stops at approved PRDs; task generation is a separate step (dev-tasks).
How it works
- Overview
- When to Use
- Workflow
- Output Convention
- PRD.md format
- Common Mistakes
SKILL.md preview
---
name: dev-design
description: Use when given a large, multi-part feature/refactor/initiative idea (often a pasted proposal or brain-dump) that needs to be understood, questioned, and broken into sub-projects, each captured as a PRD (a short PM-spec + architecture doc, design only). Stops at approved PRDs; task generation is a separate step (dev-tasks).
metadata:
version: "0.1.0"
---
# dev-design — Idea → Sub-Project PRDs
Before anything else, follow `../_shared/update-check.md`.
## Overview
Turn a big, messy idea into **sub-projects**, each captured as a tight `PRD.md` (PM spec +
architecture doc, aim ~800–1,500 words). Core principle: **understand and challenge before
you split; split before you write; let parallel sub-agents write — tight on the first
draft.**
Design only — PRDs carry Open Questions, not tasks; `dev-tasks` derives tasks later from
the *approved* PRD so it can be revised without rotting a tasks file.
Shared dispatch/output conventions for this whole pipeline live in
`../_shared/dev-common.md` — read it once; this file only covers what's unique to PRDs.
## When to Use
- Human pastes a proposal / list of improvements / refactor brief and wants it planned.
- Work spans multiple subsystems or many independent changes.
- The ask is "design this", "break this down", "create PRDs".
- **Not for:** a single well-scoped change (use `dev-code` light mode); generating tasks from an
existing PRD (`dev-tasks`); writing code (planning only).
## Workflow
```dot
digraph d {
"Explore codebase" [shape=box];
"Brainstorm + ask questions" [shape=box];
"Propose phased decomposition" [shape=box];
"Human approves?" [shape=diamond];
"Dispatch 1 agent per sub-project (parallel) -> PRD.md" [shape=box];
"Verify PRDs + write index" [shape=box];
"Explore codebase" -> "Brainstorm + ask questions" -> "Propose phased decomposition" -> "Human approves?";
"Human approves?" -> "Propose phased decomposition" [label="no / revise"];
"Human approves?" -> "Dispatch 1 agent per sub-project (parallel) -> PRD.md" [label="yes"];
"Dispatch 1 agent per sub-project (parallel) -> PRD.md" -> "Verify PRDs + write index";
}
```
0. If `auto` is in the request, print the Auto mode text from `../_shared/dev-common.md` first thing, if it has not been printed yet.
1. **Workspace guard.** Apply the Workspace guard in `../_shared/dev-common.md` before
anything else (pull, then a worktree off the latest default branch for new work).
If this is the first dev-loop skill of this piece of work, show once: "Say `auto` any
time to run all remaining stages without stopping." Follow Auto mode if on.
2. **Invoke the `brainstorm` skill first.** Skip if `brainstorm.md` exists or brainstorm
just ran. It explores context, then asks questions in
rounds (several at once, each with a default), then presents the design for approval.
3. **Explore the codebase before asking anything.** Read real files. Verify the proposal
against the code — surface dead code, contradictions, stale assumptions.
4. **Ask + challenge.** Follow Question rules in `../_shared/dev-common.md`: ask in
rounds, only the blocking, decomposition-changing decisions, plain words, a
recommended default on each (`defaults` accepts all). Never ask what the code can
answer. Challenge over-engineering. Park non-blocking unknowns for each PRD's Risks /
Open Questions — don't ask them now.
5. **Propose a phased decomposition and get approval.** Each sub-project = one coherent
chunk with scope, dependencies, a phase. Show a dependency graph. Do not dispatch until
the human says go. In auto mode, skip this approval and go.
6. **Dispatch one PRD-authoring agent per sub-project, in parallel** — dispatch rules in
`../_shared/dev-common.md`. Use `design-agent-prompt-template.md`. Agents write
`PRD.md` only, tight on the first draft — no TASKS.md.
7. **Verify + index.** Confirm each `PRD.md` exists. Write `README.md` in the branch
folder: sub-project table, dependency graph, locked decisions, consolidated
cross-cutting Risks / Open Questions.
8. **Hand off.** Human reviews PRDs, resolves/triages Open Questions *in the PRD*
(`[RESOLVED: …]` or `[DEFERRED]`), then runs `dev-tasks`. Don't generate tasks here.
In auto mode, every OPEN item takes the recommended default, logged as
`[RESOLVED: auto default - <choice>]`, and `dev-tasks` runs next without stopping.
- Tell the user in one line: "You can run `dev-critique` on this design if you want a
second opinion; it will not run unless you say so." (Not in auto mode; auto never
runs it.)
- Write the Handoff section per `../_shared/dev-common.md` and end with its "Continue
here" message (skipped in auto mode).
## Output Convention
Path grammar in `../_shared/dev-common.md` (find/create the branch folder first):
```
docs/agent_files/<branch-slug>-<YYYYMMDD-HHMM>/<NN-kebab-name>/PRD.md
docs/agent_files/<branch-slug>-<YYYYMMDD-HHMM>/README.md # index, written last
```
## PRD.md format
Length: aim for **~800–1,500 words** (soft target — going over is fine; never trim after
writing). Full annotated skeleton + frontmatter is in `design-agent-prompt-template.md`;
shape:
```
frontmatter(status/summary/date) → TL;DR → Problem → Goals/Non-Goals → Requirements →
Architecture (high-level; tiny snippet only if essential, never full code) →
Decisions (# | Decision | Choice | Alternatives considered | Why) → Manual steps →
Risks / Open Questions ([OPEN]/[RESOLVED: …]/[DEFERRED] — the dev-tasks gate) →
Acceptance Criteria
```
## Common Mistakes
- **Generating tasks here.** Stop at PRDs; `dev-tasks` owns task files.
- **Dispatching before approval.** Get the decomposition signed off first (auto mode
skips the approval).
- **Running `dev-critique` on your own.** It only runs when the user asks.
- **Full code in the PRD.** Architecture states the interface/contract delta, not a
component's full body — that belongs to `TASKS.md`/the implementation.
- **Thin agent prompts.** Cold agents need full context — see `../_shared/dev-common.md`.
- **No status tag on a Risk/Open Question.** Without `[OPEN]/[RESOLVED]/[DEFERRED]`,
`dev-tasks` can't tell a settled PRD from an unsettled one.
- **No index.** Without the README, cross-cutting decisions stay scattered.
Files in this package (2)
- SKILL.md 6 KB
- design-agent-prompt-template.md 4 KB
Versions
| Repository | Score | Licence |
|---|---|---|
| This page · tejitpabari99/tejit-plugin | 100 | MIT |
| pi-ixel/awesome-agent-workflow | 100 | MIT |
| lflish/devflow | 100 | MIT |
| FOURTEEN1416/academic-agent-toolkit | 96 | CC-BY-4.0 |
| codeaholicguy/ai-devkit | 90 | Apache-2.0 |
Database snapshot: 13.11.2026 · package version 1eb3774875cc (SHA-256 digest of the entire package)
Downloads in the last 30 days: 0 · ranking
Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.
Similar skills
- 100dev-tasks
Use when one or more approved PRDs already exist (from dev-design or otherwise) and you need to generate or refresh the high-level TASKS.md for each — a tracking table plus a short per-task spec, not code. Re-run any time a PRD changes — it regenerates tasks from the current PRD… - 100initiative
[Project Management] Use when a workflow step or the user asks for an initiative: a structured product idea or feature request ready for task refinement. - 100initiative
[Project Management] Use when a workflow step or the user asks for an initiative: a structured product idea or feature request ready for task refinement. - 100initiative-breakdown
Cut an initiative, an idea bigger than one PRD — a brief, a context package, a pile of design notes, or a draft PRD that outgrew itself — into an ordered list of right-sized planned PRDs, each one finishing with something that can be promoted to production. Consolidates the… - 100writing-feature-prds
Use when starting a new feature, defining requirements before implementation, or when the user says "new feature", "create a spec", "create a PRD", or "feature PRD". - 100prd
Use when the user wants a PRD, feature spec or requirements write-up - "write me a PRD", "spec this feature", "turn this idea into requirements", "document how this flow works", "draft the requirements doc" - or wants to update an existing PRD (new sections, changed plans,…