Agentic Spec

Use this skill to convert intent into a compact engineering contract before implementation. Keep the result proportional; do not create ceremony for work that only needs a direct fix.

Workflow

1. Classify rigor:
   • prototype: throwaway exploration, personal tool, or demo.
   • internal: team feature, workflow automation, or shared tool.
   • production: user-facing, security-sensitive, data-sensitive, or agentic system with tools/actions.
2. Gather the six context types:
   • Instructions: roles, constraints, repo rules, user priorities.
   • Knowledge: APIs, schemas, business rules, source docs.
   • Memory: relevant project state and previous decisions.
   • Examples: existing patterns, input/output examples, nearby code.
   • Tools: required APIs, MCP servers, CLIs, sandboxes, permissions.
   • Guardrails: explicit boundaries, non-goals, unsafe actions.
3. Generate SPEC.md from assets/templates/SPEC.md.
4. For AGENTS.md work, apply the filter test: omit anything discoverable from the codebase, keep project-level rules under 200 lines, and push scoped rules into subdirectories.
5. If requirements are materially ambiguous, ask only the blocking question. Otherwise record assumptions in the spec and continue.

Rigor Scaling

• Prototype: one-page spec, 1-2 edge cases, manual verification acceptable.
• Internal: focused spec, at least 3 edge cases, automated test plan expected.
• Production: full six-context spec, threat/data boundaries, measurable success criteria, eval/security/deploy follow-up links.

References

Read ../agentic-engineering-sdlc/references/day1_sdlc_and_context.md only when deeper context-engineering or AGENTS.md guidance is needed.

Done Criteria

• SPEC.md has goal, context, constraints, edge cases, out of scope, and success criteria.
• Rigor level is explicit and justified.
• Assumptions and user decisions are separated.
• Next recommended skill is named only when useful: agentic-evals, agentic-security-review, or agentic-production-readiness.