Workflow Writer is a single prompt that converts a syllabus, an SOP, a ticket list, a diagram, or plain prose into one declarative document: a Body that defines exactly when each task is done, and a Soul that defines why it exists and how it must be done well. That document is the input from which any Multi-Agent System or Agentic AI builds the application — using optimal resources, with no guessing, no waste, and no drift.
Instead of writing code first, you write the process first. Workflow Writer asks the right questions and emits one document. That document is format-agnostic on input and framework-agnostic on output — it plugs into any MAS, any orchestrator, any agent framework.
A single YAML algorithm spec. It is the source of truth — the app is compiled from it, not written around it.
Sequence, branch, loop, parallel, join. Every step has a precondition and a checkable postcondition.
A step is done only when its postcondition verifies true. Downstream never runs on unverified output.
A workflow without a Body is a wish. A workflow without a Soul is a dead checklist. Workflow Writer enforces both, and makes the Soul enforceable rather than decorative.
Inputs, outputs, success condition, steps with pre/postconditions, failure policy, termination, operators. This is what makes the workflow correct.
Intent, values, voice, quality-of-spirit, guardrails. Each Soul clause has an observable proxy and a verdict type, so it is checked by a QA operator. This is what makes the workflow worth doing.
The Body answers “is the step done?”. The Soul answers “was it done in the spirit the process exists for?” Neither is allowed to be vague.
Four stages. The writer never invents facts — it asks, mirrors back, and validates before emitting.
Prose, syllabus, SOP, Jira export, Mermaid/BPMN, existing prompt, README, table, or a mix. The writer maps each into the algorithm model — you don't rewrite your input.
Whole process → roles and boundaries → tasks and flow → per-task procedure. The critical question for every task: how do we know it is done correctly?
Intent, 3–7 values, voice and tone, quality-of-spirit with observable proxies, guardrails. Any clause that can't be observed is marked advisory — honestly.
Run both checklists. Fix before returning. Emit one YAML document plus short Assumptions and Open questions lists.
# Emitted document (abridged) algorithm: syllabus_to_tlm_site version: 0.1.0 soul: intent: "Help students learn the syllabus honestly and clearly." values: - id: V1 statement: "Never fabricate source content." observable_proxy: "every claim maps to a syllabus topic or cited source" verdict: pass|fail enforced_by: qa_verifier guardrails: - id: G1 statement: "Never shame a learner." purpose: "Turn a syllabus into a static TLM website." inputs: { syllabus: text, audience_level: string } outputs: { site_root: directory } success_condition: "quality_checklist passes AND lighthouse > 90" steps: - id: S1 operator: analyst precondition: "syllabus non-empty" postcondition: "every unit has >=1 topic and >=1 objective" soul_refs: [V1] on_fail: { retry: {max: 1}, then: escalate }
No reformatting required. The writer recognizes the shape and maps it into the algorithm model.
| Format | What it usually contains | How it is mapped |
|---|---|---|
| Free prose / interview | Narrative description | Structured interview questions |
| Syllabus / curriculum | Units, topics, objectives | Units → parallel branches, topics → steps, objectives → postconditions |
| SOP / policy doc | Ordered rules, approvals | Rules → steps, approvals → human operators, “must” → postconditions |
| Jira / Linear export | Issues, statuses, links | Issues → steps, links → depends_on, done-criteria → postconditions |
| Mermaid / BPMN | Nodes and edges | Nodes → steps, diamonds → branches, joins → join policy |
| Existing prompt | Role, requirements, checklist | Requirements → postcondition library, checklist → QA step |
| README / design doc | Goals, structure, conventions | Goals → intent + Soul, structure → roles, conventions → quality-of-spirit |
| Table / spreadsheet | Rows of tasks | Rows → steps, columns → fields, notes → dependencies |
| Mixed | Any combination | Parse each part, reconcile into one model |
Any place a work process must become an application — reliably, auditably, without wasted tokens or agent drift.
Faculty syllabus → static teaching-learning website: content, quizzes, case studies, chain-locked progression.
Policy documents → automated onboarding flows with human approvals where the SOP requires them.
Design docs and diagrams → build/test/deploy pipelines with explicit, verifiable completion gates.
Underspecified briefs → multi-agent research with a Soul that forbids fabrication and rewards clarity.
Ticket exports → triage, resolve, and escalate workflows where “done” is a checkable predicate.
Regulated procedures → replayable, auditable algorithm specs with a Soul that encodes ethics and guardrails.
The document is not just documentation — it is the resource plan for the application.
Agents receive only their declared input slice. No whole-context stuffing. Caching is keyed on (step_id, input_hash).
Control flow is deterministic. Branch decisions are engine-evaluated predicates, not agent opinions.
Every transition is a state change on a known step. Replay is possible from the document alone.
Feeds any MAS or agent framework — the document is the contract, not the runtime.
Soul clauses are bound to steps via soul_refs and checked by a QA operator. Values are not wallpaper.
Unenforceable clauses are marked advisory and surfaced as open questions — no silent hand-waving.
This is the entire mechanism — one prompt. Paste it into any capable model, then give it your process in whatever form you have (prose, syllabus, SOP, tickets, diagram, prompt, table, or mix) and one line of soul: why this exists and what it must never become.
Ctrl/Cmd + C. The text is a plain-textarea, so selection always works.
Then hand the document to any agentic framework — or compile it yourself. The Body guarantees correctness. The Soul guarantees it was worth doing.
Copy the prompt