Workflow Builder

Workflow Builder

Real builds aren't linear. They look like trees: a goal at the root, with phases that decompose into tasks, that occasionally fork into entirely new sub-workflows (when a parent task discovers it needs a tool/dataset/process that doesn't exist yet).

Running example used throughout this skill: "Ship a public /reports REST API." Midway through building the endpoint you discover there's no rate-limiter — and you can't expose a public API without one. That's a fork: you spawn a rate-limiter sub-workflow, build it, return to the API work, and later realise the rate-limiter is reusable and graduate it into a shared platform skill.

This skill makes that tree explicit and persistent.

Why we need this

Three failure modes this skill prevents:

1. Context loss when forks happen. "We need X to continue Y" → we build X over many turns → we forget Y was the parent. The state file pins the parent so we always know where to resume.
2. Sub-workflows that turn out to be reusable get lost. The rate-limiter started as a one-off detour for the /reports API; it's actually shared middleware every endpoint needs. The skill captures this graduation explicitly.
3. No durable "where are we" answer. Every new session asks "what's the state of project X?" — the workflow tree is the answer, in YAML, in the project's own folder.

The four shapes (taxonomy)

Not everything is a workflow. There are four distinct shapes of work, with different state-keeping needs. Pick the right shape before scaffolding anything:

Graduation ladder (work moves UP as it matures):

One-shot task  →  Workflow  →  Operation  →  Development
  atomic         finite        recurring     open-ended
  no state       per-build     per-instance  continuous

Crucial: an Operation IS a workflow that runs on a cadence. Same tree shape, instantiated each cycle. Phases under an operation can themselves be sub-workflows (graduated from earlier one-time builds) OR atomic tasks (a human clicking "approve" once a cycle) OR nested operations.

When does work move up the ladder?

• One-shot → workflow: when mid-flight you realise it has ≥3 steps OR might span sessions OR has discoverable unknowns
• Workflow → operation: when it finishes AND turns out to repeat on cadence — EXTRACT the pattern into a skill, leave the original workflow state as the historical record
• Operation → development: when "running it" turns into "continuously evolving the product that runs it"

Things rarely move DOWN. A retired development becomes an operation (just run, no enhance); an unused operation becomes a one-shot fallback.

Skill vs Workflow — the most-asked meta-question

These are orthogonal. Don't conflate them.

They interact like this: workflows USE skills. A single workflow can invoke many skills as it progresses — the /reports API build pulls in a testing skill for the handlers and a UI skill for the docs page. The workflow-builder skill orchestrates STATE; domain skills do WORK.

Skills are horizontal (capabilities — same across projects). Workflows are vertical (one specific delivery — pulls horizontal skills as needed).

The extraction pattern (how the skill library grows):

Workflow runs once → done, archive the state folder
Workflow recurs    → EXTRACT the pattern into a skill (Phase 5 resume step covers this)
Skill matures      → other workflows discover and use it

The rate-limiter is the proof case. Started as a fork off the /reports API. Finished. Realised every public endpoint needs it. We extracted the pattern into a shared middleware skill. The original workflow state lives on as historical record; the skill is the reusable capability.

When you finish a workflow, ask: "will this recur?" If yes → extract into a skill before closing.

When to use

• Starting a new build that has more than ~3 steps and might discover gaps mid-flight → run the start sequence (phases 1+2)
• Hitting a fork mid-task ("we need something we don't have") → run the FORK phase, spawn a sub-workflow, mark parent as blocked-by
• Resuming after a break or in a new session → workflow.yaml is the source of truth; CONTEXT.md the human-readable index
• Graduating a sub-workflow into a permanent recurring process → MOVE the node into a phase of another long-running workflow

When NOT to use

• One-shot tasks (a single edit, a single PR) — overkill
• Workflows that are already codified as a separate skill (those have their own phase model)
• Pure data-exploration sessions where the goal IS the discovery, not a deliverable

The state lives WITH the project

Per workflow, state goes in <project>/workflow-state/:

<project-folder>/
├── CONTEXT.md                    ← human-readable project overview (existing)
└── workflow-state/
    ├── workflow.yaml             ← the tree (single source of truth)
    ├── current-pointer.md        ← short "you are here" note (auto-updated)
    └── log.jsonl                 ← append-only history of node transitions

This is intentional — the workflow state travels with the project, not with the skill. The skill is the grammar; each project's workflow.yaml is a sentence.

The five phases

Stage Contracts

Every phase is a CONTRACT. Three fixed sections — Inputs / Process / Outputs — so anyone (or any future session) can read it in 30 seconds and know what loads, what runs, what comes out.

- id: phase-2-build-endpoint
  title: Build the /reports REST handlers + auth check
  status: planned
  inputs:
    - source: L3:reference         # factory — stable across runs
      file:   docs/architecture.md
      section: "## API conventions"   # ← selective section routing
      why:    routing + error-envelope pattern every endpoint follows
    - source: L4:prev-phase        # product — this run's artifacts
      file:   ../phase-1-schema/output/reports-schema.md
      section: full file
      why:    the response shape the handlers serialise
  process: |
    1. Add GET /reports + GET /reports/{id}
    2. Wire the auth middleware + error envelope
    3. Add handler tests, run them green
  outputs:
    - artifact: endpoint handlers
      location: src/api/reports.py
      format:   python
    - artifact: test report
      location: workflow-state/phase-2-execute.md
      format:   markdown

The principles (why this shape works)

1. L3 vs L4 — factory vs product. L3 inputs (reference material) are stable across runs and "internalised as constraints" — architecture docs, conventions, style guides, config. L4 inputs (working artifacts) are per-run and "processed as input" — previous-phase outputs, source material the user supplied. Mixing them in the same input row wastes tokens and dilutes attention. (This L3/L4 split is adapted from the Interpreted Context Methodology — the value is the discipline, not the labels.)

2. Selective section routing. Don't say "read architecture.md." Say "read the API conventions section of architecture.md." Most reference docs are 80% rationale and 20% rules-for-this-task. Loading the whole file is the equivalent of bringing every cookbook to make one omelette.

3. Phase N's outputs become Phase N+1's L4 inputs. This is the handoff. A human can open the output file, edit it, and the next phase picks up the edited version. No state-management layer needed — the filesystem IS the orchestration.

4. One-way cross-references. A phase declares what it loads from elsewhere. The "elsewhere" never declares which phases load it. Prevents reference-tracking from blowing up as the workflow library grows.

Back-compat with older workflows

Existing workflows using notes: | prose for phases still work — notes remains an optional field. New workflows (and refactors of stuck workflows) should use the structured triple. When restructuring an old workflow, the notes: prose typically splits cleanly: lines that name files become inputs:, numbered steps become process:, "Done when:" criteria become outputs:.

Pipeline-script docstring contract

Workflow phases declare contracts in YAML. The scripts that execute those phases should declare the SAME contract in their module docstring — so when you open the file directly (without the workflow.yaml in view) you can still see in 5 seconds what loads, what runs, what writes.

The shape — first thing in the docstring, before anything else:

"""Phase N — <one-line goal>.

<2-3 sentence summary of what this pass does and why it exists.>

Reads:
  <relative path>            # L3 (reference) — what it provides
  <relative path>            # L4 (working artifact) — which prior phase produced it

Writes:
  <relative path>            — <one-line description>

Methodology:
  - <key decision 1>
  - <key decision 2>
  - <gotchas, threshold values, fallback chains>
"""

Worked example (a generic ETL leaf):

"""Phase 3 — Build the feature matrix.

Join raw events with the customer dimension, derive per-customer aggregates,
and write a single parquet the model phase consumes.

Reads:
  data/raw/events.parquet              # L4 (Phase 2 output — cleaned events)
  data/dim/customers.csv               # L3 (stable: customer dimension)
  config/feature_spec.yaml             # L3 (stable: which aggregates to compute)

Writes:
  data/output/feature_matrix.parquet   — one row per customer, model-ready
  workflow-state/phase-3-execute.md    — row counts + null-rate report

Methodology:
  - Left-join on customer_id; unmatched events dropped (logged count)
  - Aggregates: 7/30/90-day windows; missing → 0 with a presence flag
  - Fails loudly if null-rate on any key feature > 5%
"""

Why this matters. Three failure modes it prevents:

1. Drift between yaml contract and code. Re-stating Reads/Writes in the docstring makes mismatches obvious during review (the yaml says "writes X" but the script writes "Y" — a diff hunk shows it).
2. Path archaeology. A new contributor opening the script knows immediately where outputs land and where dependencies live — no need to grep for Path("...") calls scattered through 300 lines.
3. Refactor safety. When restructuring a data folder, the docstring is the manifest you grep for to find every script that reads/writes the affected paths.

When is a script too small to need this? If it's <50 lines AND only reads/writes one file each AND is not part of a multi-phase pipeline, a prose docstring is fine. Everything in a scripts/ folder that's part of a numbered or named phase: structured contract.

Node anatomy (workflow.yaml schema)

goal:
  id: <short-slug>
  title: <one-line>
  status: planning | executing | blocked | completed
  why: <one-paragraph why-this-matters>

phases:
  - id: <slug>
    title: <one-line>
    status: planned | in-progress | blocked | done | skipped
    completed_at: <iso-date or null>

    # Stage contract — see "Stage Contracts" section above
    inputs:                          # what to load
      - source: L3:reference | L4:prev-phase | L4:user-upload
        file: <path>
        section: <"full file" | specific section header>
        why: <one line>
    process: |                       # ordered steps
      1. ...
      2. ...
    outputs:                         # what this produces + where it lands
      - artifact: <name>
        location: <path>
        format: <markdown | yaml | json | code | ...>

    notes: <free text>               # optional — only for stuff that doesn't fit above
    tasks:
      - id: <slug>
        status: ...
        # tasks may be leaves OR may have sub-tasks (nested same shape)

forks:
  - id: <slug>                     # the fork's local id within the parent
    spawned_at: <iso-date>
    spawned_from: <parent-node-id> # which node triggered this
    reason: <why we forked>
    sub_workflow: <path>           # link to the sub-workflow's own state folder
    status: open | merged | abandoned
    return_to: <parent-node-id>    # where we resume when sub-workflow closes

current_node: <node-id>            # where execution should pick up
last_updated: <iso-date>

Invariants

1. Every node has a parent (except the goal). No orphaned tasks.
2. current_node is always one specific leaf-or-fork. Not "somewhere in Phase 3" — a specific id.
3. Forks are first-class. They go in forks:, not silently in a task. Every fork must list its return_to.
4. Sub-workflows are full workflows themselves. Same schema, their own state folder, their own log. A sub-workflow can fork further.
5. Status of a phase = function of its children. A phase is done only when all tasks under it are done or skipped. The skill auto-rolls this up — never set manually.
6. Log is append-only. Every state transition writes a JSONL line. Never delete history; mark abandoned instead.

Graduating a sub-workflow

When a sub-workflow finishes AND turns out to be reusable (recurring monthly, recurring per-feature, etc.), don't leave it as a one-off in the parent's forks: list. Graduate it into a longer-running workflow or a skill:

1. Find the right host (a recurring *-workflow skill, or extract a brand-new skill)
2. Add a phase to the host that points at the sub-workflow's outputs
3. Mark the original fork in the parent as merged → graduated_to: <host>/phase-N
4. The sub-workflow's state folder stays — it becomes the template for future runs

The rate-limiter IS this. It started as a fork off the /reports API; it's now a shared middleware skill every endpoint pulls in.

When you finish a workflow, ask: "will this recur?" If yes → extract into a skill before closing.

How to use this in conversation

• "start workflow for " → run phases 1+2: scaffold workflow-state/, write skeleton workflow.yaml, ask the 4 clarifying questions (goal / why / output / first 3 phases)
• "we hit a fork: " → run phase 4: create fork node, pause parent, spawn sub-workflow folder, set new current_node
• "resume X" OR "where are we in X" → read workflow.yaml, render the tree, point at current_node
• " is done, back to " → run phase 5: close fork, mark merged, restore parent current_node
• "graduate to " → run the graduation playbook

The skill NEVER executes the underlying work itself. It just keeps the tree honest.

Done means (when has the workflow-builder skill actually succeeded?)

The skill's invocation is complete when:

• ✅ <project>/workflow-state/workflow.yaml exists with a valid tree (goal → phases → tasks)
• ✅ current_node points at a SPECIFIC leaf id (not "Phase 3, somewhere")
• ✅ current-pointer.md is a one-paragraph human note that says "you are here, doing X, next is Y"
• ✅ log.jsonl has at least one transition line for the current session
• ✅ Status fields on every node reflect actual state (no in-progress task that hasn't been touched in days)

The workflow itself is complete when:

• ✅ All phases have status done or skipped
• ✅ All forks are merged or abandoned (no open forks left)
• ✅ A "what shipped" note exists somewhere durable — PROGRESS.md entry, CHANGELOG, or memory file
• ✅ The "will this recur?" question has been asked. If yes → graduated to a skill; if no → state folder stays as historical record

If a workflow current_node hasn't moved in 14 days, it's stuck — surface it, don't let it rot silently.

Anti-patterns — habits that wreck the tree

• Treating phases as a rigid plan. They're a current best-guess. As work progresses, splits/merges/reordering is expected. Don't refuse to restructure because "the plan said X."
• Skipping the FORK phase when you hit unknown unknowns. If a task can't progress without data/tool that doesn't exist, that's a FORK — spawn a sub-workflow with return_to pinned. Don't silently switch to building the missing thing inside the parent — the parent's current_node ends up lying about what's actually happening.
• Forking too eagerly. Not every detour needs a sub-workflow. If the diversion is <30min and stays inside the same project, just do it inline. Forks are for things that have their own decomposition and might span sessions.
• Updating workflow.yaml from prose memory instead of reading the file first. If two sessions update the same workflow without reading the current state, you get conflicts. Always load the YAML at start of session; write back at end.
• Letting current_node and reality diverge. When you spend 20 min on something that isn't current_node, EITHER move the pointer first OR explicitly note "deferring current_node to do X". Silent drift makes "where are we?" answers wrong.
• Closing a fork without writing return_to's parent back to in-progress. The fork closing implies the parent should resume. If the parent stays blocked, the next session won't know to pick it up.
• Building infrastructure for a one-shot. If the work is genuinely atomic (single email, single config edit, single PR), don't scaffold workflow-state/. Use the shape taxonomy honestly.
• Graduating prematurely. A sub-workflow that's run once is not yet a recurring pattern. Wait for the second or third instance before extracting it into a skill. False-positive graduations clutter the skill library.
• Forgetting the "why" on goal:. Future sessions need it. "Build the API" doesn't survive a 2-month pause; "Build the API so partners can pull reports without us exporting CSVs by hand" does.

Templates

• templates/workflow.yaml — empty skeleton for a new workflow
• templates/CONTEXT-section.md — markdown snippet to drop into the project's CONTEXT.md so the workflow tree shows up in human-readable form too