dge-design

DGE Design

Use this skill when starting a new piece of work. It is the brainstorm and gap-exposure step for Delivery Graph Engineering.

Purpose

Turn a raw demand into testable requirements without designing the implementation. The output feeds /dge-plan-graph.

Core rules

1. Ask exactly one question at a time and wait for the answer. When an answer is vague, incomplete, or contradictory, push back with one follow-up before moving on — do not batch questions and do not proceed on a fuzzy answer.
2. For every question, offer your own grounded recommendation and let the user confirm or redirect — interrogate, but never leave the user to invent the answer from nothing.
3. Lock in each concrete answer by restating it before it becomes a requirement.
4. Expose gaps explicitly; do not hide uncertainty in prose.
5. Do not produce work nodes or implementation tracks here.
6. Do not mark the design ready while blocker gaps are unresolved.
7. Use the canonical DGE IDs: DEM-###, REQ-###, GAP-###.

Inputs

#$ARGUMENTS

If no demand is provided, ask the user what demand, problem, or request they want to shape.

Preflight: require the DGE CLI

Before writing anything, run the shared preflight (one callable place for every dge-* skill and the /dge-deliver conductor). Design runs before dge init, so skip the graph check:

dge preflight --no-graph || npx --no-install dge preflight --no-graph

If it exits non-zero, stop and tell the user to install the DGE CLI first — do not proceed:

DGE CLI not found. The dge CLI is a separate npm package from the /dge-* slash commands. If you installed the skills via the plugin marketplace, you still need the CLI — the marketplace ships prompts, not the binary. Install it in this project:

npm install --save-dev github:rafaelolsr/delivery-graph

Then re-run /dge-design.

The dge CLI is the only writer of delivery-graph/graph.json. Never hand-write or hand-edit graph.json. If the CLI is missing, the correct action is to install it, not to emulate it.

Workflow

1. Capture the raw demand

Record:

• source
• requester, if known
• problem statement
• summary — a single-sentence TL;DR (≤ ~25 words: what changes and why). This becomes the bold lead line every gate/report surface opens with, so write it as the one line you would keep if the reader read nothing else. Capture it through dge add-demand --summary "..." (or backfill later with dge edit-demand).
• desired outcome — keep it to at most 3 sentences; lead with the single-sentence result and push detail into constraints/requirements. A wall-of-text outcome is a smell: if it runs long, the extra sentences are usually constraints or requirements in disguise. (The summary field is where the one-liner lives, so the outcome does not have to carry it.)
• urgency
• constraints
• non-goals
• known stakeholders

2. Survey what already exists

Before writing any requirement, read the relevant parts of the existing codebase. Do not design against a blank slate.

• Locate code, modules, config, or prior demands that already address part of this demand (grep/find/read; check delivery-graph/demands/ and requirements/ for overlap).
• For each candidate, decide: reuse as-is, extend, or replace with justification.
• Record findings as a short prior-art note in the demand's problem/constraints capture. If the survey reveals an unknown that blocks scoping, emit it as a GAP of type: scope.

Default position: the simplest change that reuses existing code beats a new abstraction. If nothing exists to reuse, say so explicitly — that statement is the evidence you looked.

Read prior learnings (compound loop)

Past demands leave behind learnings — bugs, gotchas, conventions, and failed approaches — under delivery-graph/learnings/. Query them before scoping so this demand does not re-solve or re-break something already learned:

dge learnings --about "<a few keywords from this demand>" --json   # or: npx dge learnings ...
dge learnings                                                       # list all when unsure

For each relevant learning: read the file it names, and either fold its guidance into a requirement/constraint or record why it does not apply. This is the read side of the compound loop that /dge-compound feeds — skipping it means the toolset stops getting smarter with each demand. If there are no learnings yet, say so; that statement is the evidence you looked.

3. Grill the demand

Run a structured grill-me: a relentless but constructive interrogation that exposes every unresolved decision before any requirement is written. This is the step that turns a fuzzy ask into concrete, testable requirements — do not shortcut it.

3a. Map the decision tree, and show it

Pick the branches relevant to this demand from the list below, then state the map to the user up front: "Here is what I want to pin down. I'll go branch by branch." As you work, announce which branch you're on so the user can see progress and steer.

• Problem — what problem, for whom, and what are they doing today instead?
• Users / stakeholders — who is the primary user? who signs off?
• Scope — what is explicitly in, and what is explicitly out?
• Data — what does this read, write, and display?
• Actions — what can a user do, and what happens after each?
• Edge cases — empty, loading, error, partial-data, and concurrency states?
• Constraints — tech stack, existing systems to respect, performance, a11y?
• Success — how do we know it works? what evidence proves it?

3b. Walk each branch, one question at a time

Start with the most foundational branch (usually Problem — everything depends on it). Within a branch:

1. Ask one question and wait.
2. Offer a recommendation grounded in the codebase survey and prior learnings: "My recommendation: . <one-sentence rationale, citing prior art where possible>." Then ask the user to confirm or redirect.
3. If the answer is vague, incomplete, or contradictory, push back with one concrete follow-up — do not accept it and move on. Useful pushes:
   • "What does that look like specifically?"
   • "Give me a number — how many is 'a lot'?"
   • "What's the worst thing that happens if we get this wrong?"
4. If the answer is concrete, lock it in: "Locked: ." and move on.
5. If it depends on an earlier decision, flag it: "This depends on — does that still hold?"

Challenge, as you walk, every: unclear problem · vague success criteria · missing owner · hidden dependency · untestable requirement · scope creep · speculative feature (actually asked for, or only imagined?) · premature abstraction or robustness beyond the stated outcome · requirement with no traceable need · everything defaulting to must when it is really should/could · contradiction · validation ambiguity · unhandled edge/empty/error state.

Anything the user cannot resolve becomes a GAP (Step 4), not a silently-assumed answer.

4. Emit gaps

Represent each unresolved issue as:

- id: GAP-001
  type: validation
  severity: blocker
  question: "What evidence proves this requirement is complete?"
  blocks: [REQ-001]
  resolution: null

5. Write requirements

Before writing any requirement, replay the Confirmed decisions recap: list each locked-in decision from Step 3 back to the user and give them a last chance to correct a misread. Only then convert decisions into requirements — a requirement must trace to a locked decision, not to an assumed answer.

Each requirement must be testable.

Assign priority honestly: must only for what the stated outcome fails without. Push nice-to-haves to should/could, and record anything deliberately excluded as a wont priority or a demand non-goal. Every requirement must trace to a stated outcome; if it cannot, drop it or demote it.

- id: REQ-001
  demand_id: DEM-001
  statement: "..."
  priority: should
  acceptance:
    - "..."
  validation:
    method: automated-test
    required_evidence:
      - "..."

6. Save outputs

Author the canonical store only through the dge CLI (see CLI contract below). The CLI writes delivery-graph/graph.json and the demands/ and requirements/ markdown for you:

• dge add-demand ... writes delivery-graph/demands/DEM-###.md and the graph entry
• dge add-requirement ... writes delivery-graph/requirements/REQ-###.md and the graph entry
• dge add-gap ... / dge resolve-gap ... record gaps in the graph

Do not write graph.json yourself. The CLI owns its schema.

Readiness gate

Return ready_for_graph: true only when:

1. The problem is clear.
2. The intended outcome is explicit — a one-line summary exists and the outcome is at most 3 sentences (not a wall of text).
3. Non-goals are named.
4. Constraints are known or marked as gaps.
5. Requirements are testable.
6. Acceptance criteria exist.
7. No blocker gaps remain unresolved.
8. Existing code was surveyed; each requirement either reuses/extends prior art or states why new work is needed.
9. Priorities are differentiated (not everything is must) and speculative scope is either justified or recorded as a non-goal / wont.
10. Edge cases were probed — empty, error, partial-data, and concurrency states are each either covered by a requirement or explicitly out of scope.
11. Every locked-in decision was replayed to the user (Confirmed decisions recap) and each requirement traces to one.

If blocker gaps remain, stop and report them instead of invoking /dge-plan-graph.

Output

Follow the shared output convention (see skills/README.md): lead the final reply with a bold one-line synthesis — ready for /dge-plan-graph, or blocked on N gaps — then the demand's progress indicator (see skills/README.md), then the detail (requirements captured, gaps open), then a ## Next block: /dge-plan-graph when ready, or the specific blocker gaps to resolve when not. After a successful save, include the stable local viewer link delivery-graph/view/index.html so the user can open the graph without running a separate viewer command.

CLI contract

The DGE CLI is required (see Preflight) and is the only writer of the canonical store. Use dge if it is on the PATH (the plugin ships it), otherwise npx dge:

dge add-demand --title "..." --source "..." --summary "one-line TL;DR" --outcome "..."   # or: npx dge add-demand ...
dge edit-demand DEM-### --summary "..."   # backfill the TL;DR on an existing demand
dge add-requirement ...
dge add-gap ...
dge resolve-gap ...

Never edit graph.json by hand. If the CLI is unavailable, stop and install it (Preflight).