Operational workflows for AI development

Use Plan when an idea needs shaping before implementation: brainstorm, refine, challenge, promote to spec, and execute with traceable scope.

Use Caveman when speed and low-token communication matter. It strips filler while preserving exact technical names, commands, risks, and code.

A lightweight documentation pattern for AI coding tools: keep the root contract small, then route agents to architecture, content, UI, and task guides only when needed.

Use this loop when an agent can edit files, run commands, call external tools, or work across long-running sessions. It separates the control plane from disposable compute, keeps credentials outside the workspace, and makes approval plus verification explicit before anything leaves the sandbox.

Use this workflow when a team wants to start or steer an agent from Slack, Teams, or another shared discussion. It filters the thread into a durable brief, separates participation from write authority, pins the repository and branch, and carries the work through an independently reviewed pull request.

Use this loop to instrument production agents without treating transcripts as observability. It defines stable identities, traces workflow and tool boundaries, limits payload capture, emits low-cardinality metrics, and joins each run to evidence such as review, merge, deployment, or rollback status.

Use this workflow for agent-authored or AI-reviewed changes that need more than a generic diff pass. It establishes a risk contract, bounds review context, separates blocking findings from suggestions, records resolution evidence, and keeps final acceptance with an independent owner.

Use this workflow when a provider or AI platform deprecates a model, alias, or request parameter. It turns the notice into an owned inventory, tests the replacement against an explicit behavior contract, stages a reversible rollout, verifies that retired traffic reaches zero, and removes the old path before shutdown.

Use this workflow when moving a long-running tool-using workload onto the OpenAI Agents API or another managed agent harness. It separates the provider-managed session from application-owned identity, authorization, side-effect state, retention policy, evaluation, and release control so managed execution does not become unmanaged product risk.

Use this workflow when Sentry, another error tracker, or an incident record can provide a coding agent with stack traces and runtime context. It separates impact control from code repair, preserves an exact evidence packet, tests competing fix hypotheses, limits the agent to an isolated change, and requires deployment plus a post-fix observation window before the error is resolved.

Use this workflow when a repeated task needs deterministic stages around agent judgment rather than another open-ended chat. It separates ordinary code from agent steps, versions every input and result contract, limits concurrency and spend, pauses before consequential actions, and proves that resume or retry behavior cannot duplicate external effects.

Use this workflow when one genuinely hard reasoning, research, proof, coding, or evaluation task needs more than 65,536 tokens of thinking and output. It treats Gemini 4 Argon long decode as an asynchronous production job: the application chooses an explicit shared token ceiling, persists reconnect state, handles every terminal status, joins all output steps, and verifies completeness, quality, usage, storage, and cost before publishing the result.

Use this workflow when a team wants a prompt-generated voice for a GPT-Live product. It separates acoustic intent from conversation behavior, binds one approved voice ID to a release manifest, tests disclosure and listener experience, reviews resemblance and retention risks, and canaries the change with a known fallback voice.