Skip to content

Sathvikar01/Agent-engineering-skills / agent-development-workflow

agent-development-workflow

Verified

Coordinate the complete engineering lifecycle for substantial AI-agent or model-assisted workflow builds. Use when building, productionizing or substantially redesigning an agent, adding major tools/state/RAG/orchestration, or evaluating and hardening a whole agent system. Do not use for one bug, tiny prompt/schema/API changes, ordinary app work, simple deterministic utilities, conceptual agent explanations, or a narrow task owned by one specialist.

FileKindSize
SKILL.mdskill md7.7 KB
evals/evals.jsonconfig4.3 KB
evals/trigger-evals.jsonconfig3.4 KB
references/decision-and-defence.mdmarkdown1.3 KB

evals/evals.json

View on GitHub
{
  "skill_name": "agent-development-workflow",
  "evals": [
    {
      "id": 1,
      "prompt": "Productionize a prototype that reads ambiguous receipts, calculates exact reimbursement, answers with citations and may send a payment after approval. It passed one happy example. Team wants regex-only receipt parsing and a five-agent framework. Before coding produce the engineering path, uncertainty map, baseline/eval gates, contract/authority/evidence flow, vertical slice and meaningful decision/defence artifacts. No repository or dataset is supplied; distinguish proposed checks from executed evidence.",
      "expected_output": "Full lifecycle with task-fit models, authority and measured acceptance.",
      "files": [],
      "expectations": [
        "Maps semantic receipt uncertainty to bounded model interpretation and arithmetic/permissions to deterministic authority; rejects unsupported framework/workers.",
        "Defines baseline/multidimensional critical gates before large implementation and a repeatable run/debug/correct/review path with vertical slice.",
        "Covers evidence reconciliation and model-failure/security/recovery behavior without claiming unexecuted results.",
        "Gives material decision status and direct technical defence with limitations/fallback; one happy example is insufficient."
      ]
    },
    {
      "id": 2,
      "prompt": "Build a local Python vertical slice, no provider or payments. Scripted model proposes reimbursement from observed receipt R1, amount=30; deterministic policy cap=20; a second permitted supported proposal amount=20 exists and user authorized any amount<=20 for this fake. Starter admission is def admit(p): return p[\"amount\"]>=20 (bug). PLAN expected invariant, implement fixture tests, RUN to expose bug, DEBUG/CORRECT it, rerun and REVIEW. Also test invented R9, timeout, schema-valid inconsistent explanation_amount and no valid candidate. Save code, actual command logs and a concise evidence/defence report.",
      "expected_output": "Executed correction loop and truthful safe vertical slice.",
      "files": [],
      "expectations": [
        "Actually executes a test exposing starter bug, identifies root cause, corrects it and reruns meaningful regression checks.",
        "Denies amount30/unknown evidence/inconsistent explanation without effects; supported authorized20 reaches fresh admission once.",
        "Bounds timeout/alternative paths and reports no-admissible status; trace links source/decision/effect and evidence.",
        "Reviews results/limitations and defends the implementation without claiming provider or production reliability."
      ]
    },
    {
      "id": 3,
      "prompt": "Fix one typo in this production model prompt: Return insufficent_evidence when facts are absent. Supply the corrected sentence only.",
      "expected_output": "Specialist-sized direct edit; no meta-workflow overhead.",
      "files": [],
      "expectations": [
        "Returns corrected insufficient_evidence sentence.",
        "Does not impose architecture/approval/eval/lifecycle documentation for this typo."
      ]
    },
    {
      "id": 4,
      "prompt": "Our receipt-to-reimbursement agent passes 18/20 examples. One output has the wrong currency field and the reviewer asked 'is it robust?'. Plan the next debugging step, including what you will paste to the AI coding assistant, and answer the reviewer's question about the currency handling in the format you would use in a technical interview.",
      "expected_output": "Evidence-bearing debug request and a direct, anchored technical defence.",
      "files": [],
      "expectations": [
        "States an observation and a root-cause hypothesis before investigating or delegating, and includes the actual failing case, output field and error/trace evidence in the AI request instead of a bare 'please investigate'.",
        "Traces the failing request from input/evidence to the exact final currency field.",
        "Answers the reviewer directly first, then implementation location, reason, evidence and limitation, and replaces 'robust' with the named run/test result (18/20, the failing case) rather than an adjective.",
        "States the limitation as gap, current fallback and next step, and separates what the engineer chose or verified from what the AI tool generated."
      ]
    }
  ]
}
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 470kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall