AI Creation Workflow

AI creation workflow

Own the project plan and its execution. Ask skill-governor to allocate the best installed skill for each stage, then sequence those assignments, preserve their outputs, and keep one project state from planning through delivery.

Choose the smallest control mode

• Direct: one clear reversible step; invoke the known specialist and return.
• Light: two to four dependent stages; keep a short in-chat contract and checkpoints.
• Full: cross-stage production, paid generation, multiple deliverables, repeated runs, handoffs, or work that may resume later. Create a project control file from assets/templates/00-project-control.md and a state file from assets/templates/workflow-state.json. These assets travel with an independently installed Skill; do not depend on repository-level templates.

Run the project

1. Reuse current conversation, existing project files, and confirmed choices.
2. Define the outcome, deliverables, constraints, acceptance evidence and items explicitly outside scope. Resolve only missing facts that materially change goal, scope, cost, permission or an irreversible result.
3. Send each stage's required output, inputs, target tool, constraints, permissions and acceptance evidence to skill-governor. It returns the primary Skill, any necessary specialist, the boundary between them and missing inputs. If it returns partial or gap, let it finish the bounded local/online/create-or-improve loop before continuing this stage.
4. Build a dependency graph and write those assignments into the plan. Each stage has one primary skill, optional complementary skills, inputs, output contract, validator and retry rule.
5. Execute the next ready stage. Save material outputs before moving downstream.
6. Apply gates at creative lock, paid generation, external publication and other irreversible actions. A gate names the exact decision and its impact.
7. Validate the artifact itself. A successful command is supporting evidence, not proof that a media result is correct.
8. Close out with final files, validation evidence, known limits and state. Record skill-use evidence only when it reflects an actual task result.

Token and delegation

Use an adaptive context budget. Direct and light work starts compact; full work loads the project contract, decision index, current stage and its direct dependencies. Store stable facts once and pass references plus versions instead of copying artifact bodies into every handoff.

Skill allocation remains owned by skill-governor. Reuse its accepted decision while the lane, input type, tool, constraints and mapping fingerprint are unchanged. Rescan or reroute only for a new capability, competing candidates, mapping drift, a tool change or a failed assignment. Load the full instructions only for the skill executing the current stage; do not preload future skills.

Evaluate subagents only when a full workflow has at least two independent ready stages. Delegate a bounded artifact when inputs and acceptance are complete, writes and external resources do not overlap, no unresolved decision blocks it, parallel work is expected to reduce total cost, and current runtime rules allow delegation. Keep sequential work, shared-artifact edits, ambiguous requirements and small tasks in the root agent. The root owns user communication, integration and final validation.

Domain pack routing

Select a domain pack only after the project contract is clear. A pack contributes stage templates, specialist lanes, domain gates and runtime adapters; it never owns the overall project or replaces skill-governor. Validate installed packs with python scripts/domain_pack.py validate and read references/domain-packs.md before adding one.

Match the request to a pack's triggers, load only that pack and its linked guide, and send each declared specialist lane or runtime capability to skill-governor. The pack names required capabilities; the governor maps them to current installed Skills.

State rules

• A stage is pending, ready, running, blocked, review, accepted, failed, or stale; do not mark it accepted without its declared evidence.
• Store stable decisions once and reference them downstream. When a decision changes, mark only stages that consumed the changed input and their real dependants stale before rerunning them. Use scripts/workflow_state.py invalidate for a persisted state file.
• Retries preserve the original job identity and result directory. Never submit a second paid job merely because polling timed out; query by prompt or job ID.
• Keep secrets and private media outside project state. Store references to environment variable names and local files instead.

Read references/task-contract.md when creating a contract, references/project-control.md for a resumable project, and references/gates-and-evidence.md before executing or closing a costly stage. Read references/intake-and-gates.md when requirements are incomplete or a gate may be needed. Read references/role-model.md when allocating Skill ownership or handing a stage to the governor. Read references/token-and-delegation.md when creating a full project, deciding whether to delegate, or reducing context use.