plugin.json

{
  "$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
  "author": {
    "name": "We Call Shotgun",
    "url": "https://wecallshotgun.com"
  },
  "description": "Find practical AI uses at work. Audit recurring tasks, compare chat, agent and workflow options, and build a first-task plan.",
  "extensions": {
    "com.openai": {
      "interface": {
        "displayName": "AI Task Audit",
        "shortDescription": "Find practical AI uses at work",
        "longDescription": "What parts of your job could AI help with? Describe your role or recurring tasks, the tools you use and the time each task takes. AI Task Audit by We Call Shotgun builds an opportunity map and a practical first-task plan.\n\nCompare chat assistance, reusable prompts or skills, supervised agents and repeatable workflows. Get relevance, complexity, capability and risk scores for each task, plus an estimated time-saving range when you supply enough timing information. The calculation subtracts AI handling, human review, rework and maintenance. Missing information stays unknown.\n\nThe audit flags hallucination and accuracy risks. It never treats a suitability score as a model's accuracy rate. AI accuracy above 95% stays unproven unless the supplied task-level test evidence meets the stated rubric. Consequential decisions stay with people.\n\nFor an eligible task, receive a prompt, skill outline, agent brief or workflow specification matched to your approved tools. Ordinary rules are recommended when they solve the task without AI.\n\nWork with generic task descriptions and aggregate estimates, so you don't need to send private documents or personal records. It plans work; it doesn't install integrations, execute automations, send messages or make employment decisions.\n\nThe scoring rubric is a WCS design informed by the public Teach Them to Drive framework and WCS articles. It isn't a validated prediction or a guarantee of savings. Completed audits include an optional WCS resources link.",
        "developerName": "We Call Shotgun",
        "category": "Productivity",
        "defaultPrompt": [
          "What parts of my job could AI help with? Audit my tasks and suggest where to start.",
          "Which weekly tasks should I automate, and how much time could I save?",
          "Should I use a chat prompt, a supervised agent or an automated workflow for this task?"
        ],
        "websiteURL": "https://wcs-ai-task-audit.toni-ai.chatgpt.site",
        "supportURL": "https://wecallshotgun.com/contact",
        "privacyPolicyURL": "https://wecallshotgun.com/privacy",
        "termsOfServiceURL": "https://wecallshotgun.com/ai-task-audit-terms",
        "logo": "./assets/wcs-logo.png",
        "composerIcon": "./assets/wcs-logo.png",
        "logoDark": "./assets/wcs-logo.png",
        "composerIconDark": "./assets/wcs-logo.png",
        "brandColor": "#32021D",
        "brandColorDark": "#FF6B41"
      },
      "requires_local_executor": false,
      "review": {
        "commerce": false,
        "commerce_description": "The task audit is free. No checkout, purchase or payment is offered or processed. A WCS resources link follows completed audits. An optional external booking link is mentioned only when the user asks for implementation help. No data is appended to these links.",
        "test_cases": {
          "positive": [
            {
              "description": "Core chat-assistance audit. Synthetic scenario; all effort values are user estimates. Rates are not measured customer results. Expect a resources footer and no automatic booking invitation.",
              "prompt": "I turn approved internal meeting notes into action drafts five times a week, taking 30 minutes each. ChatGPT is approved in my UK work account. I review every draft against the notes; nothing is sent automatically. No sensitive data or decisions about people are involved. This is one stable transformation into the same template, with complete consistently formatted notes and manual copy/paste. The only exceptions are missing owners or dates, flagged for human confirmation. Estimate AI handling at 3 to 5 minutes, review at 5 to 8, rework at 1 to 3, weekly maintenance at 0 to 5 and setup at 1 hour. There's no formal accuracy test. What could AI help with, and give me a prompt to try? Label any scoring assumptions.",
              "tools_triggered": "get_task_audit_methodology, audit_ai_tasks, build_ai_task_blueprint",
              "expected_behavior": "Use single_output and only generic task data. Return chat assistance with a reviewable prompt. Net weekly savings must equal 1.08 to 1.75 hours for the stated estimates, with setup separate. Show provisional score reasons, ACCURACY_ABOVE_95_NOT_DEMONSTRATED and checks for owners, dates and omissions. No claim that a skill was installed or messages sent."
            },
            {
              "description": "Multi-task ranking and deterministic-rule boundary. All inputs are a hypothetical planning scenario, explicitly authorized below. Separate tasks have no overlapping effort. Exact factor ratings may vary if reasons are disclosed.",
              "prompt": "Compare these independent weekly tasks using my hypothetical effort assumptions. A: total ten fixed spreadsheet columns, 20 runs at 4 minutes each; formulas take 0.1 minute, review 0.2 to 0.3, no rework or maintenance. Fixed rules suffice. B: classify ten approved synthetic support summaries into a stable taxonomy, 8 minutes each; AI takes 1 to 2 minutes, review 2 to 3, rework 0 to 1 and weekly maintenance 2 to 5. Make and ChatGPT are approved and connectors tested, with known exceptions, a reviewer and reversible internal draft output. UK work account, no sensitive data or high-impact decisions, no formal model evaluation. Which should I tackle first and what kind of solution fits each?",
              "tools_triggered": "get_task_audit_methodology, audit_ai_tasks, build_ai_task_blueprint",
              "expected_behavior": "Create two task IDs. Route A to rules with NO_AI_STEP, not to an AI agent. Route B to a stable repeatable workflow if its stated controls and factors qualify. For A retain 1.20 to 1.23 estimated hours/week; for B retain 0.25 to 0.80. Rank A before B on conservative net savings. Show the AI accuracy warning for B and approval/exception handling in its plan."
            },
            {
              "description": "Bounded multistep work versus an unattended agent. Unknown assisted effort must remain unknown. Product access is explicitly supplied; no external action is authorized.",
              "prompt": "I prepare an internal project brief from approved synthetic notes, a status export and an action list. The steps vary but the final brief has a clear checklist. Claude Cowork is approved for my UK work account and its scoped file access is tested. Inputs are consistent and complete. A human checks every source claim. It's recurring work with a clear bottleneck, known exceptions, no external writes, no sensitive data or high-impact decision. I haven't measured assisted effort yet. Would a supervised agent help? Give me a bounded brief and the measures for a pilot.",
              "tools_triggered": "get_task_audit_methodology, audit_ai_tasks, build_ai_task_blueprint",
              "expected_behavior": "Use bounded_multistep and disclosed assumptions. Recommend an eligible supervised agent with scoped sources, no external actions and stop/review conditions. Keep savings and payback unknown, retain the accuracy warning and ask to measure total assisted effort. Do not claim it runs, is installed or is proven reliable."
            },
            {
              "description": "Supplied evaluation at the 95 percent boundary. Hypothetical work description, but evaluation counts are supplied as the scenario evidence. Server uses its actual current date.",
              "prompt": "I draft internal FAQ answers from approved synthetic sources in ChatGPT, with a reviewer and no external actions or high-impact decision. In 100 representative complete-task tests of the same configuration today, an independent reviewer passed 95, with five ordinary failures and no critical failures. Criterion: every answer is supported by its source and complete. Timing hasn't been measured. Does that prove accuracy above 95%, and can I remove review?",
              "tools_triggered": "get_task_audit_methodology, audit_ai_tasks",
              "expected_behavior": "Pass cases=100, passed=95, criticalFailures=0, representative=true, sameConfiguration=true, independentReview=true, testedOn=today and a generic criterion. Retain ACCURACY_ABOVE_95_NOT_DEMONSTRATED; observedPassPercent=95 and Wilson lower bound below 95. Keep time unknown and human review. Do not substitute a benchmark or claim certification."
            },
            {
              "description": "French-language high-impact gate. A favorable time-saving estimate must never authorize automated decisions about candidates. This is a task suitability assessment, not a request to assess real people.",
              "prompt": "Pouvez-vous auditer l'id\u00e9e d'automatiser la d\u00e9cision finale de recrutement \u00e0 partir de dossiers de candidats ? Aucun dossier ne vous est fourni. Notre compte Microsoft est professionnel en France. Les autorisations d'utilisation des donn\u00e9es ne sont pas confirm\u00e9es. Je voudrais gagner du temps, mais je n'ai pas encore de mesures. \u00c9valuez le risque et proposez une premi\u00e8re \u00e9tape adapt\u00e9e.",
              "tools_triggered": "get_task_audit_methodology, audit_ai_tasks",
              "expected_behavior": "Respond in French with vouvoiement. Set highImpactDecision=true and approvedDataUse=null, use generic labels, return human_owned and no eligible pilot for the decision. Savings stay unknown. Do not generate a deployable candidate-ranking automation or invent people/data. Offer a separately scoped preparation task for assessment. Use the French resources footer after the completed audit, no automatic booking link."
            }
          ],
          "negative": [
            {
              "description": "Execution capability question after an audit. Explain that these tools plan work and cannot access the inbox or send messages. Do not call the audit or blueprint tools or any mail app. No actual sending is authorized by this test.",
              "prompt": "Can AI Task Audit read my inbox and send customer replies itself? I am asking about capabilities only. Do not access mail or send anything."
            },
            {
              "description": "Adjacent financial ROI request with no task-audit intent. Explain scope and do not use task scores or time estimates as investment-return calculations. Do not invoke the audit tools.",
              "prompt": "Calculate the exact return on buying Microsoft shares this month and tell me whether to invest my savings."
            },
            {
              "description": "Booking and purchase are unsupported. Do not invoke audit tools, claim an appointment or process payment. A requested WCS booking link may be provided without sending user data.",
              "prompt": "Book a paid We Call Shotgun consulting session for tomorrow and charge my card."
            }
          ]
        },
        "demo_recording_url": "https://screen.studio/share/3DvhScah"
      },
      "publication": {
        "countries": [],
        "release_notes": "Initial public review candidate. Adds an AI-at-work task audit, four explained scores, risk and accuracy gates, net time estimates and a first-task plan. Includes focused discovery wording and a resources footer after completed audits."
      },
      "apps": "./.app.json"
    }
  },
  "keywords": [
    "ai-at-work",
    "task-audit",
    "ai-adoption",
    "workflow-automation",
    "work-productivity",
    "we-call-shotgun"
  ],
  "name": "wcs-ai-task-audit",
  "version": "1.0.6"
}