---

name: ai-workflow-spine description: Use when designing, auditing, or improving an AI-native workflow, agentic project, coding-agent setup, skill, pipeline, scheduled automation, or LLM-backed product where deterministic code and AI judgment interact.

AI Workflow Spine

Use this skill to turn a fuzzy AI workflow into a maintainable system with explicit boundaries, replayable artifacts, and tests that let future agents iterate safely.

Goal

Produce or improve a workflow where deterministic engineering and AI judgment each do the jobs they are best suited for, with a clear artifact boundary between them.

When To Use

Use for:

• New AI/agent projects.
• Codex, Claude Code, OpenCode, or agent workflow setup.
• LLM-backed pipelines, scheduled automations, report generators, image/content generators, or research systems.
• Refactors where model prompts, tools, sources, and code are tangled together.
• Postmortems where bad AI output may be caused by stale or polluted inputs.

Do not use for small deterministic code fixes unless the bug crosses an AI/tool boundary.

Core Pattern

Default spine:

PRD -> RFC -> deterministic boundary artifact -> AI judgment step -> replayable output -> tests

Use the lightest version that fits the task. A one-day project may need a short README section instead of separate docs, but it still needs the boundary and verification.

Instructions

1. Identify the deterministic layer.

   • Collection, filtering, parsing, auth, config, scheduling, idempotency, retries, validation, and persistence should usually be code.
   • Make source enablement config-driven when possible.

2. Identify the AI judgment layer.

   • Interpretation, selection, synthesis, summarization, prompt writing, ranking, or dynamic research may belong to an LLM or agent.
   • Prefer a single API call when the input/output contract is fixed.
   • Upgrade to an agent only when the workflow must fetch extra context while deciding.

3. Define the boundary artifact.

   • Choose a plain, inspectable artifact: markdown context, JSON bundle, prompt file, CSV, rendered report draft, transcript, or query result bundle.
   • The artifact should be understandable without re-running private tools.
   • Save enough metadata to debug freshness and filtering: source, timestamp/window, config mode, and fallback reason when relevant.

4. Make outputs replayable.

   • Save the AI-facing input and the AI-produced intermediate output when they affect user-visible results.
   • Retry generation from the same intermediate artifact before recollecting data.
   • When output quality is poor, inspect the artifact before tuning prompts or switching models.

5. Define the control contract.

   • Use the lightest runner that fits: an ordinary turn for short or exploratory work, a persisted goal for bounded multi-turn work with verifiable completion, and a scheduled task only when inputs recur or external state must be revisited later.
   • For schedules, continue the same task when prior context matters; use standalone runs when each input window should be independent.
   • Name the trigger, one run's work unit, freshness or delta check, no-op condition, completion condition, pause/escalation conditions, and retry or cost boundary. Match polling cadence to how often the source can meaningfully change.
   • Separate completion of one run from retirement of a recurring routine. An unchanged source should normally produce a cheap no-op, not repeated work.
   • Pilot the exact prompt, tools, model route, and permissions on a representative slice before unattended execution; review early runs before widening scope or cadence.

6. Document decisions.

   • PRD: product philosophy, success criteria, target user, non-goals.
   • RFC: architecture boundary, alternatives rejected, coupling choices, fallback behavior, known failure modes.
   • Working log: changelog and lessons learned from real failures.

7. Test the boundaries.

   • Prioritize tests for parsers, time/window filtering, source unavailable behavior, fallback signals, retries, output shape, privacy scans, and artifact replay.
   • Keep default tests offline. Put live APIs, credentials, hardware, and expensive model calls behind opt-in integration tests.

Acceptance Criteria

A workflow shaped by this skill is done when:

• The deterministic layer and AI judgment layer are explicitly separated.
• There is a named boundary artifact with a stable schema or format.
• A future agent can inspect and replay the artifact without guessing hidden state.
• Long-running work has an explicit trigger, per-run work unit, no-op condition, completion/stop conditions, and a bounded retry, cadence, or cost policy where relevant.
• Unattended execution has been piloted manually with the same prompt, tools, model route, and permission boundary.
• Fallback and retry behavior is specified for empty inputs, stale sources, model failures, and downstream push/render failures.
• Tests cover the boundary and the failure modes most likely to break autonomous iteration.
• Docs record at least one rejected alternative when the architecture choice was non-obvious.

Output

Return one of these, depending on the task:

• A concise workflow audit with gaps and proposed changes.
• A PRD/RFC/doc patch.
• A skill or config patch.
• A test plan focused on artifact boundaries and agent-safe iteration.

Keep recommendations scoped. Do not introduce a framework, graph engine, or multi-agent system unless the workflow needs dynamic context gathering or parallel independent work.