Okr Generator
Generates structured OKR plans (Objectives and Key Results) for teams and companies following Goo…
by onewave-ai · MIT
The skill
Generates structured OKR plans (Objectives and Key Results) for teams and companies following Google/Intel methodology. Takes company goals, team function, quarter, and current metrics to produce a comprehensive okr-plan.md with objectives, key results, scoring criteria, alignment mapping, tracking cadence, and retrospective templates.
OKR Generator
Generate a comprehensive, actionable okr-plan.md from user inputs, applying the Google/Intel OKR framework.
Contents
references/methodology.md-- Core OKR principles (non-negotiable rules to apply).references/inputs.md-- The four required inputs plus optional context.references/output-template.md-- Exact structure of the generatedokr-plan.md.references/best-practices.md-- Appendix content (common mistakes, OKR cycle, grading by KR type, reading list).
Workflow
- Read
references/inputs.md. Collect the four required inputs: company goals, team function, quarter/period, current metrics/baseline. If any are missing, ask for them. Never fabricate baselines. - Read
references/methodology.mdand apply every core principle. Objectives are qualitative and inspirational (no numbers); key results are quantitative and measurable (always a number). - Design 3-5 objectives. Attach 3-4 key results to each. Mark each as Committed or Aspirational, targeting roughly 60-70% committed and 30-40% aspirational across the plan.
- Map every objective to at least one company goal. Build both the alignment table and the cascade diagram. Drop any objective that does not ladder up.
- Write a 5-level scoring rubric (0.0, 0.3, 0.5, 0.7, 1.0) for every key result, with KR-specific descriptions.
- List initiatives separately from key results. Initiatives are the activities ("how"); key results are the measured outcomes ("what"). Never conflate them.
- Document dependencies and risks. Assign a single owner role to every objective and key result.
- Define the tracking cadence (weekly check-in, monthly scoring, quarterly retrospective) and include all three templates plus the appendix from
references/best-practices.md. - Generate the full document per
references/output-template.mdasokr-plan.mdin the working directory (or a user-specified path). Produce at least 400 substantive lines, no filler. - Summarize the plan: objective count and committed/aspirational split, the most ambitious key result, key dependencies/risks, and the recommended first action.
Generation Rules
- Validate inputs before generating; base all targets on the provided baselines (committed: ~10-30% improvement; aspirational: ~50-100%+ or breakthrough).
- Objectives qualitative, key results quantitative. Rewrite any KR lacking a number.
- Every objective maps to a company goal; every KR has a single owner and a scoring rubric.
- Never connect OKR scores to compensation, promotion, or performance reviews. Include the decoupling reminder in the scoring guide.
- Include all templates in full (weekly, monthly, quarterly). Do not abbreviate.
Edge Cases
- Multiple teams: Generate one plan per team in its own file (e.g.,
okr-plan-engineering.md,okr-plan-marketing.md). - Company-level OKRs: Remove team-to-company alignment; show department-level cascade instead.
- Mid-quarter adjustment: Preserve original OKRs, mark adjusted ones with
[ADJUSTED], and include rationale per change. - Previous-quarter scores provided: Calibrate ambition -- if the team scored 1.0 on everything, push harder; if below 0.4, investigate whether the cause was execution or target-setting.
Use this skill
Connect your agent to the Skill Store MCP server and ask for it by ref: onewave-ai/okr-generator-45532706308.
Details
- Publisher
- onewave-ai
- Category
- Business › Operations & productivity
- License
- MIT
- Updated
- 2026-10-19
References
More from onewave-ai
- Accessibility Auditor Audit websites for accessibility issues and WCAG compliance.
- Agent Army Deploy a 2-layer parallel agent hierarchy for large, parallelizable work — big refactors, multi-f…
- Agent Eval Harness Use to test whether an AI agent, prompt, or skill actually works before or after it ships — build…
- Agent Swarm Deployer Deploys swarms of sub-agents for massive parallel data processing tasks.