anandamritraj
- 5 skills
- 0 followers
- 20 hours ago last updated
- ▌ Agency Design · anandamritraj bundleUse when designing an AI-native product, feature, or workflow where humans and AI agents share the work, to decide who does what and who holds which decisions. Walks each key step to set its level of autonomy (Level 1 human asks and AI answers, Level 2 human assigns and AI executes, Level 3 human and AI assign to each other), the decision rights, the approval gate, and how human agency is amplified, plus the team roster of who can task the agent and who holds each gate, then produces a one-page Agency Map. Use when the user asks about human-in-the-loop, agent autonomy, approval gates, trust, or the human-AI division of labor, or runs /agency-design.
- ▌ Forward Deployed Pod · anandamritraj bundleUse to review the output of any GOAL phase (a Goal Brief, a spec, a diff or PR, an eval report, a release plan, or telemetry) from four forward-deployed perspectives at once. Convenes a pod of four persona-agents (FD Product Manager, FD Tech Lead, FDE, FD Domain Expert/Business Owner), auto-selecting which lenses matter for the current phase (read from .goal/state.json), and returns one synthesized Pod Review with a verdict and a queue of proposed actions. The pod advises; the human decides and holds the gate. Use when the user asks for a pod review, a multi-perspective review, or to review a phase's work, and at GOAL phase gates.
- ▌ Goal And Constraints · anandamritraj bundleUse at the very start of any new product, feature, project, or decision, before brainstorming, speccing (OpenSpec), or building (Superpowers). Runs a short interview to name the user, their need, and the product context; the goal as a measurable objective function (outcome, not output); the data and evals it depends on; the single top constraint or bottleneck (discovered, not assumed); and the four product risks (value, usability, feasibility, viability), then produces a one-page Goal and Constraint Brief. Use when the user starts scoping something new, asks to define the goal, asks what the constraint or bottleneck is, or runs /goal-discover.
- ▌ Responsible Autonomy · anandamritraj bundleUse to govern how deployed AI autonomy matures over time, at the Learn-phase cycle review and as a standing Responsible AI Posture. Runs a short interview to set the posture (the current autonomy ceiling and the next cycle's proposed step up), write if-then maturity gates that tie each step up to pre-committed evidence (the eval pass^k bar), choose an independent verifier scaled to the stakes (the builder is never the verifier) and name a separate authorizer, set deception-aware held-out checks, and define trip-wires and a reversible brake that can lower autonomy, not only raise it, then produces a one-page Responsible Autonomy Sheet. Use when the user asks about responsible AI, an AI risk or maturity posture, maturing or rolling out agent autonomy, an ARB or review-board gate, independent verification, or who signs off before production, or runs /goal-govern. This governs how autonomy matures across cycles; setting a single step's autonomy level at a point in time is agency-design instead.
- ▌ Eval Driven Development · anandamritraj bundleUse to design the evals for a build before you build it, right after the Goal and Constraint Brief, and again to run them in the Evaluate phase. Runs a short interview to name the outcome the eval must prove, source an eval set from real failures, write outcome-graded pass/fail tasks with negative and distractor cases, choose graders (rules, LLM-as-judge, or human), set the reliability bar (pass@k versus pass^k), and wire the suite into the Learn-phase reward signal, then produces a one-page Eval Plan. Use when the user asks to define or write evals, decide how to measure or grade a model or agent, set a pass rate or reliability bar, prove an outcome before building, run the evals, or check whether the build passed, or runs /goal-evals. This grades model and agent outcomes; unit-testing code before writing it is test-driven-development instead.