{"schema_version":"2.0","record_type":"article","canonical_url":"https://marketingwiki.ai/articles/prompts-vs-skills-vs-agents","id":"prompts-vs-skills-vs-agents","slug":"prompts-vs-skills-vs-agents","title":"Prompts vs Skills vs Agents for Marketing Work","description":"A file-level architecture for deciding when marketing instructions should stay a prompt, become a reusable skill, or join an agent workflow.","dek":"Prompts request outputs. Skills preserve methods. Agents coordinate decisions and tools across a goal.","category":"Architecture","topics":["prompts","skills","agents"],"publishedAt":"2026-08-10","updatedAt":"2026-08-10","lastVerifiedAt":"2026-08-10","readingMinutes":3,"author":"Marketing Wiki Editors","reviewer":"Marketing Wiki Editors","featured":false,"sources":[{"title":"Agent Skills specification","url":"https://github.com/agentskills/agentskills/blob/main/docs/specification.mdx"}],"wordCount":481,"body":"Use a prompt for one request, a skill for a repeatable method, and an agent when work requires several decisions or tools.\n\nConfusing these layers creates giant prompts that are hard to test, reuse, or update.\n\n## Prompt\n\nA prompt describes current task and desired result.\n\n```text\nReview this landing-page draft for clarity. Audience is operations leaders at 50-200 person SaaS companies. Return five specific edits.\n```\n\nPrompt works when context is local, stakes are low, and method does not need to survive current conversation.\n\n## Skill\n\nA skill packages a method so several agents and people can apply same standard.\n\n```text\nskills/landing-page-review/\n├── SKILL.md\n├── references/rubric.md\n└── examples/review.md\n```\n\nSkill defines:\n\n- When to use it\n- Required inputs\n- Review sequence\n- Evidence rules\n- Output format\n- Failure and stop conditions\n\nAgent Skills specification uses `SKILL.md` as entry point and supports progressive disclosure through linked references. That keeps core instruction small while detailed material remains available when needed.\n\n## Agent\n\nAgent owns goal requiring several steps:\n\n```text\nGoal: improve landing-page conversion hypothesis quality.\n\n1. Read audience context.\n2. Analyze page and current evidence.\n3. Run landing-page review skill.\n4. Generate test hypotheses.\n5. Score hypotheses against experiment rubric.\n6. Prepare review packet.\n```\n\nAgent decides which skill or tool to use and when to stop. Permissions should match goal. Drafting hypotheses does not require access to deploy site changes.\n\n## Workflow\n\nWorkflow describes operating process around agent:\n\n- Trigger\n- Owner\n- Inputs\n- Agent or skills used\n- Approval points\n- Destination\n- Measurement\n- Recovery path\n\nWorkflow may be fully deterministic or contain an agentic step. Calling every workflow an agent hides where decisions occur.\n\n## Decision table\n\n| Need | Use |\n| --- | --- |\n| One output from current context | Prompt |\n| Repeatable task with stable standard | Skill |\n| Several decisions or tools toward a goal | Agent |\n| Team process with owners and approvals | Workflow |\n\n## Common failure: giant prompt\n\nLarge prompts often mix brand context, task instructions, scoring rules, examples, tool policy, and output schema. One edit can break unrelated behavior.\n\nSplit by ownership:\n\n- Context states facts and constraints.\n- Skill states method.\n- Prompt states current request.\n- Agent states sequence and decisions.\n- Workflow states people, approvals, and system boundaries.\n\n## Common failure: agent without eval\n\nAn agent that runs repeatedly needs a fixed way to judge results. Keep a small benchmark set of representative inputs and expected properties. Test instruction changes against same set before adopting them.\n\n## Practical migration\n\nStart with prompts already used every week. Extract shared method into one skill. Add examples only when they teach a concrete boundary. Compose agent after two or more skills need coordination.\n\nArchitecture should grow from repeated work. Empty agent folders and generic prompt libraries create maintenance without improving output."}