dt-workflow

name: dt-workflow description: >- Runs Claude Code dynamic workflows (ultracode) on partitionable Digitaltableteur tasks with objective verification bars. Use when the user says "workflow", "ultracode", "bulk sweep", "parallel migration", "promote all components", or wants autonomous multi-agent work with tests as the done gate. Do NOT use for single-file edits, subjective design decisions, or shipping PRs (use dt-ship-pr). metadata: version: 1.1.0 category: orchestration documentation: https://www.digitaltableteur.com/.well-known/agent-skills/dt-workflow

Dynamic workflow orchestration

Use Claude Code dynamic workflows for large, partitionable, verifiable sweeps. The runtime plans fan-out, runs subagents, and verifies results — your job is to specify the task, partition key, area skill, and done bar.

Requires: Claude Code Team/Enterprise/Max with dynamic workflows (research preview).

Copy-paste templates: references/templates.md

Instructions

Step 1: Decide if this is a workflow task

Use a workflow when ALL are true:

Do not use workflows for: homepage visual redesign, copy tone, architecture debates, or single-file fixes.

Step 2: Choose partition + area skill

Step 3: Pick invocation mode

Prefer the keyword trigger until you trust cost and behavior on this repo.

Before the first real run, configure permissions so subagent edits are not interrupted (/permissions → auto mode for the session). Monitor cost with /cost after the pilot.

Step 4: Write the prompt (required shape)

Every workflow prompt MUST include:

1. workflow: prefix (or ultracode:)
2. Partition scope — explicit paths or list source (manifest, grep, migration plan)
3. Per-slice rules — which area skill and conventions apply
4. Done bar — exact npm commands that must pass
5. Pilot size — start with 3–5 slices unless user says full run
6. Confirmation gate — plan-only first; wait for user "go" before edits; /cost after pilot

Example skeleton:

workflow: [task summary]

Partition: [how to split work]
Per slice: follow [skill] and [AGENTS.md path]
Done when: [npm commands — must all pass]
Pilot: first [N] items only

Confirmation:
- Plan only first — show partitions; wait for my "go".
- /cost after pilot before full fan-out.

Do not: commit secrets, edit .env*, force-push main

Staged migrations (large Vite → Next.js work): run T3a → T3b → T3c from references/templates.md as separate sessions — map, port, fix loop. Do not combine stages in one prompt.

Step 5: Set verification gates

Run these after the workflow returns (or include in done bar):

# Default PR gate
npm run typecheck && npm run lint && npm test && npm run build

# Domain-specific (add as needed)
npm run validate:components      # design system changes
npm run validate:agent-docs      # skill / AGENTS.md changes
npm run validate-translations    # i18n changes
node scripts/design-system/agent-eval/run.mjs   # agent-manifest

Invoke subagents when touching sensitive areas:

• security-reviewer — app/api/chat/, guardrails, auth
• accessibility-expert — component a11y / axe changes
• translation-language-checker — FI/SV copy

Step 6: Ship results

When the consolidated answer looks good:

1. Review diff (workflow hides intermediate steps)
2. Run verification gates locally
3. Use dt-ship-pr only when user asks to commit / open PR

Examples

Example 1: "Workflow to fill missing Finnish translations"

Actions:

1. Load template T2 from references/templates.md
2. Review plan-only key list; reply "go" for pilot (10 keys)
3. Paste prompt into Claude Code with workflow: prefix
4. After result: npm run validate-translations && npm test

Example 2: "Ultracode promote WIP components to beta"

Actions:

1. Load template T1 (component beta promotion)
2. User reviews plan-only output; replies "go" for 5-component pilot
3. Check /cost; expand only if budget OK
4. Done bar includes npm run validate:components
5. Review consolidated diff before PR

Example 3: "Migrate next batch of Vite pages"

Actions:

1. Run T3a (map only) — review migration table
2. Reply "proceed to T3b" with approved routes
3. Run T3b pilot (3 routes); /cost; then remaining routes
4. Run T3c fix loop if gates fail after port
5. Manual browser check before deleting src/ legacy

Example 4: User asks for one Button fix

Actions:

1. Do not invoke this skill — single-file scope
2. Use dt-design-system in normal paired session instead

Troubleshooting

Workflow still at Level 1 / checks fail after deploy

Cause: done bar omitted or production not deployed yet.

Solution: wait for Vercel, re-run objective verifier (curl scan API, npm run validate:*). See docs/AGENT_READINESS.md.

Cost higher than expected

Cause: full-repo partition without pilot or skipped confirmation gate.

Solution: re-run with explicit Confirmation: plan only first; wait for my "go" and Pilot: first 5 items. Check /cost after pilot.

Results disagree with repo conventions

Cause: area skill not named in prompt.

Solution: always cite dt-* skill + AGENTS.md path in the workflow prompt.

Workflow vs GSD confusion

Cause: both orchestrate multi-step work.

Solution: GSD for phased planning with .planning/ artifacts and human checkpoints; dynamic workflows for bulk execution with tests as gate.

Boundaries

• MUST include objective done bar (npm scripts) in every workflow prompt
• MUST start with a pilot unless user explicitly requests full sweep
• MUST NOT use for commit/PR without user request (dt-ship-pr)
• MUST NOT edit .env*, credentials, or bypass hooks
• Ask before destructive git operations

Discovery for external LLMs

Machine-readable skill index (all dt-* skills including this one):

https://www.digitaltableteur.com/.well-known/agent-skills/index.json

This skill markdown:

https://www.digitaltableteur.com/.well-known/agent-skills/dt-workflow

Copy-paste workflow prompts (HTTP):

https://www.digitaltableteur.com/.well-known/agent-skills/dt-workflow/references/templates.md

Human index: AGENT_INDEX.md