ai-agentic-workflow

ai-agentic-workflow

Follow a structured three-phase workflow — analyse, plan, execute — for every non-trivial task to ensure accurate, well-reasoned outcomes.

When to use

Use this skill when:

• Receiving a task that requires more than a single-step action
• The request involves multiple files, components, or concepts
• The problem is ambiguous and needs clarification before acting
• A failure could result from acting without understanding the full context

Do not use this workflow for trivial requests (e.g., "rename this variable", "fix this typo") — act directly instead.

Instructions

Phase 1: Analyse

Understand the problem completely before proposing any solution.

1. Read the request carefully. Identify what the user explicitly asked for and what they likely expect but did not state.
2. Gather context. Explore relevant files, dependencies, and existing patterns in the codebase. Do not guess — use tools to verify assumptions.
3. Identify constraints. Note technology choices, architectural patterns, coding conventions, and any referenced skills that apply.
4. Clarify ambiguity. If the request has multiple valid interpretations and you cannot determine the correct one from context, ask a single focused question. Otherwise, proceed with the most reasonable interpretation and state your assumption.
5. Summarise findings. Produce a brief internal summary:
   • What needs to change
   • What already exists that is relevant
   • What constraints or patterns must be respected
   • What risks or edge cases exist

Phase 2: Plan

Design the solution before writing any code.

1. Break the task into discrete steps. Each step should be independently verifiable.
2. Order steps by dependency. Do foundational work first (types, interfaces, models) before implementation that depends on it.
3. Identify affected files. List every file that will be created or modified.
4. Consider side effects. Will this change break existing tests, imports, or integrations? Plan for those.
5. Choose the minimal path. Prefer the smallest change that fully satisfies the request. Do not refactor unrelated code unless explicitly asked.
6. Present the plan. Output a concise numbered checklist of what you will do. If the task is large or risky, pause for user confirmation before executing. For smaller tasks, proceed immediately.

Phase 3: Execute

Implement the plan step by step.

1. Follow the plan in order. Do not skip steps or change the approach mid-execution without re-planning.
2. Implement one logical change at a time. Complete each step fully before moving to the next.
3. Validate as you go. After each significant change:
   • Check for errors in modified files.
   • Run relevant tests or build commands if available.
   • Verify the change works in context (imports resolve, types align, logic flows correctly).
4. Handle unexpected issues. If a step fails or reveals a problem:
   • Stop and re-analyse the specific issue.
   • Adjust the remaining plan if needed.
   • Do not force through errors with workarounds unless no better option exists.
5. Confirm completion. When all steps are done, verify the full solution:
   • All checklist items completed
   • No introduced errors or regressions
   • The original request is fully satisfied

Applying the phases

Task complexity	Analyse	Plan	Execute
Trivial (single obvious action)	Skip	Skip	Act directly
Small (2–5 steps, clear path)	Brief mental model	Inline checklist	Execute immediately
Medium (multiple files, some ambiguity)	Thorough context gathering	Explicit numbered plan	Step-by-step with validation
Large (cross-cutting, architectural)	Deep exploration, ask if unclear	Detailed plan, wait for confirmation	Incremental with checkpoints

Anti-patterns to avoid

• Jumping to code without understanding. Never edit files before completing analysis.
• Planning without context. A plan based on assumptions instead of verified facts will fail.
• Deviating from the plan silently. If you need to change course, acknowledge it and re-plan.
• Over-engineering. Implement what was asked, not what might be needed someday.
• Incomplete execution. Do not stop at "mostly done" — finish all steps including validation.