Translating Intent into Agent Work

Resources

• Write the pitch — Shape Up’s brief combines the problem, appetite, solution, rabbit holes, and explicit no-gos. A strong starting point for translating existing pain and desired behavior into bounded agent work without prematurely prescribing every implementation detail. https://basecamp.com/shapeup/1.5-chapter-06

• The Value Proposition Canvas — Strategyzer’s original tool separates customer jobs, pains, and gains from the proposed offering. Helps teachers identify the product and market context an agent needs, while keeping assumed customer value distinct from validated knowledge. https://www.strategyzer.com/library/the-value-proposition-canvas

• Example Mapping — Cucumber’s technique organizes a requirement into rules, examples, and unanswered questions. Use it to expose ambiguity, contradictory examples, and missing decisions, then ask focused clarification questions before delegation. https://cucumber.io/blog/bdd/example-mapping-introduction/

• Gherkin reference — The official syntax reference explains concrete scenarios, Given/When/Then, and scenario outlines. Turns desired observable behavior, rejection cases, and boundary conditions into acceptance examples an agent can implement and a reviewer can verify. https://cucumber.io/docs/gherkin/reference/

• Documenting architecture decisions — Michael Nygard’s original ADR proposal records context, decisions, status, and consequences. Supplies the prior-decision context that prevents an agent from rediscovering rejected options or silently violating an architectural tradeoff. https://cognitect.com/blog/2011/11/15/documenting-architecture-decisions

• AGENTS.md — The format’s official site explains repository instructions for coding agents, including project guidance and testing commands. Useful for separating durable conventions and quality requirements from the task-specific brief, with clear local instructions for each assigned area. https://agents.md/

• Aider’s repository map — Aider documents a compact, ranked map of important symbols and their relationships. Demonstrates how to give an agent navigable codebase context instead of either a directory dump or an unsupported summary of where a change belongs. https://aider.chat/docs/repomap.html

• Spec Kit — GitHub’s specification-driven development toolkit includes clarification, planning, task decomposition, and analysis workflows. A concrete reference for deriving dependency-aware assignments and checkpoints from a specification, while checking consistency before implementation. https://github.com/github/spec-kit

• The Scrum Guide’s Definition of Done — The official guide distinguishes a usable increment meeting shared quality measures from work that is merely underway. Helps separate the enduring quality bar from feature-specific acceptance criteria and specify what completion evidence an agent must return. https://scrumguides.org/scrum-guide.html

• Claude prompting best practices — Anthropic’s guidance covers explicit instructions, context, examples, structured inputs, and agentic work. Apply it to briefs containing desired examples and clearly labeled unacceptable outcomes, evidence requirements, and escalation boundaries; prompt fluency does not replace requirements validation. https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices

• Architecture constraints in arc42 — arc42 distinguishes technical, organizational, and convention-related constraints. A concise checklist for stating what an agent may simplify or change and what remains fixed, so implementation freedom does not erase product or system boundaries. https://docs.arc42.org/section-2/