Big Feature Workflow

Big Feature Workflow

This skill provides a structured workflow for AI coding agents working on large features or large refactors that may span multiple iterations.

The workflow is centered around persistent documentation in docs/ai/ so that planning, implementation, and later review can build on shared artifacts.

Feature name

Before doing any work, identify a short and stable feature name.

Use a concise lowercase kebab-case identifier, for example:

• info-layout-refactor
• query-rework
• bulk-import-api

Derive the feature name from the user request when it is obvious. If multiple names are plausible, propose one and ask the user to confirm.

Files to use

Use these files under docs/ai/:

docs/
  ai/
    <feature_name>_plan.md
    <feature_name>_tasks.md
    <feature_name>_devlog.md

• docs/ai/<feature_name>_plan.md: implementation plan and execution model
• docs/ai/<feature_name>_tasks.md: task graph for task-driven execution (optional for direct execution)
• docs/ai/<feature_name>_devlog.md: session log (Implement mode) and summary (Review mode)

Create docs/ai/ if missing.

File lifecycle

• _devlog.md: permanent record - append only, never rewrite past entries
• _tasks.md: ephemeral tracking artifact - exists only while work is in progress
• _plan.md: design record - captures rationale and scope boundaries that the devlog does not

Resuming work

When the user asks to continue work on a feature that already has artifacts in docs/ai/, read all existing files for that feature (_plan.md, _tasks.md, _devlog.md) before entering any mode. Use past session log entries and task state to understand what has already been done. If multiple features have artifacts in docs/ai/, ask which one. Then proceed with the requested mode.

Three modes of operation

When using this skill, always operate in exactly one of these modes:

1. Plan
2. Implement
3. Review

If the user does not clearly specify the mode, stop and ask.

The user may refer to the modes indirectly. For example:

• "Use the big feature workflow to plan the query rework" -> Plan mode
• "Now implement the next unblocked tasks" -> Implement mode
• "Check whether the implementation matches the plan" -> Review mode

Terminology

A session is one complete unit of work in Implement mode: selecting tasks (or a slice), implementing them, and finalizing tracking artifacts. A session may include back-and-forth with the user (clarifications, corrections, rework) but ends when the selected scope is done and the devlog entry is written.

Universal rules

Do not perform git operations (commit, branch, push, etc.) unless the user explicitly asks.

Mode changes are user-initiated.

Announce the active mode and feature name on a single line:

• when entering or switching modes (e.g. Plan mode, feature query-rework``)
• when tracking artifacts are finalized at the end of work in that mode:
  • Plan mode: after _plan.md (and _tasks.md if applicable) is written
  • Implement mode: after task statuses and the devlog ## Session Log entry are written
  • Review mode: after task markers are corrected and the devlog ## Summary is updated

Do not repeat the announcement on every response within a mode.

Stop and ask the user if:

• the mode is unclear
• the feature name is unclear

Plan mode

When the user asks you to plan the work, you are operating in Plan mode.

In this mode, your responsibility is to analyze the problem and write the workflow documents needed for later implementation.

Allowed edits

In Plan mode, you may only create or modify:

• docs/ai/<feature_name>_plan.md
• docs/ai/<feature_name>_tasks.md

Do not edit source code, tests, configs, or any other files in Plan mode.

Goal

The goal of Plan mode is to produce a concrete, implementation-oriented plan and choose the execution model for later implementation.

Execution model decision

In Plan mode, choose exactly one execution model and record it in docs/ai/<feature_name>_plan.md:

• task-driven (default): create docs/ai/<feature_name>_tasks.md
• direct (small scoped work): do not create tasks file

Use direct only when work is small, low-risk, and can be implemented in a single session without dependency tracking. Direct mode is strictly single-session: if the work does not fit, return to Plan mode and switch to task-driven.

How to work in Plan mode

To produce a high-quality plan:

• Always analyze first
  • inspect the relevant code
  • inspect existing documentation, ADRs, and related artifacts
  • identify affected components, boundaries, dependencies, and risks
• Clarify ambiguity before writing the plan
  • ask the user questions if requirements, constraints, or scope are unclear
• Keep the plan concrete and concise
  • include only context that helps future implementation
  • avoid repeated summaries and generic filler
• Focus on design and execution guidance
  • describe the intended end state
  • identify likely files, modules, or subsystems that will need changes
  • call out important constraints, patterns, or migration concerns
• Split larger work into phases
  • unless the task is genuinely small, the plan should describe multiple phases or implementation steps

What the plan should contain

Read templates/plan-template.md and use it as the structural reference for docs/ai/<feature_name>_plan.md. Sections marked optional in the template may be omitted when empty.

Do not turn the plan into full code. Your responsibility is to provide high-level implementation guidance, not to pre-write the entire change line by line.

What the tasks file should contain

Read templates/tasks-template.md and use it as the structural reference for docs/ai/<feature_name>_tasks.md. The status legend and dependency conventions defined in that template are normative and must be followed so that later modes can reliably parse task state.

Tasks should be:

• small and actionable
• independently reviewable where possible
• dependency-aware
• aligned with the phases in <feature_name>_plan.md so progress traces back to the plan
• detailed enough that implementation can proceed incrementally without re-planning the whole feature

Avoid vague tasks such as:

• "implement backend"
• "finish frontend"
• "do refactor"

Prefer concrete tasks such as:

• "Refactor query builder to support grouped filters"
• "Add request validation for bulk import payload"
• "Update info layout state handling for async loading"

Stop conditions

Stop and ask the user if:

• the scope is too ambiguous to produce a reliable plan
• the user asks for source-code changes while in Plan mode

Done criteria

Plan mode checklist:

• docs/ai/<feature_name>_plan.md exists, is actionable, and declares execution model (task-driven or direct)
• docs/ai/<feature_name>_tasks.md exists for task-driven plans
• no source files modified

Implement mode

When the user asks you to implement work from the plan, you are operating in Implement mode.

In this mode, your responsibility is to execute implementation work from the plan. In task-driven execution, complete one or more unblocked tasks from docs/ai/<feature_name>_tasks.md and keep task state in sync with reality. In direct execution, implement the entire plan in a single session without task tracking. In both cases, record what you did in the devlog.

Allowed edits

In Implement mode, you may create or modify:

• source code, tests, configs, and any other project files needed to complete the selected implementation scope
• docs/ai/<feature_name>_tasks.md (status updates only, task-driven mode)
• docs/ai/<feature_name>_devlog.md (append to ## Session Log only)

You must not:

• modify docs/ai/<feature_name>_plan.md
• modify the ## Summary section of docs/ai/<feature_name>_devlog.md
• rewrite past entries in ## Session Log
• restructure the tasks file (no renaming, reordering, or removing tasks); only update the status marker when the tasks file exists

Goal

Complete one small, reviewable implementation increment from the plan.

How to work in Implement mode

1. Read docs/ai/<feature_name>_plan.md in full.
2. Resolve execution model:
   • if docs/ai/<feature_name>_tasks.md exists: task-driven
   • if tasks file does not exist: require Execution Model with Mode: direct
   • if ambiguous: stop and ask
3. Choose session scope:
   • task-driven: read tasks file, find unblocked [ ] tasks (dependency rules from templates/tasks-template.md), and either use user-requested unblocked tasks or pick the next unblocked logical group; prefer 1-4 tightly related tasks per session, erring toward fewer when they involve complex logic or cross-component changes; if multiple plausible groupings exist, stop and ask
   • direct: the session scope is the entire plan (single session)
4. Mark selected tasks [~] before coding (task-driven only).
5. Implement within selected scope. Do not silently expand scope.
6. Run the narrowest relevant verification commands per AGENTS.md / CLAUDE.md.
7. Finalize tracking artifacts:
   • task-driven: set each selected task to [x], [~], [!], or [ ] as appropriate; [!] is for external blockers only; if a task remains [~], devlog must include done, remaining, and why
   • ensure docs/ai/<feature_name>_devlog.md exists (read templates/devlog-template.md and create from it if missing)
   • append a new ## Session Log entry; for direct execution use Tasks completed: n/a - direct mode

Deviation policy

• Minor deviations are acceptable. Record them under Deviations from plan.
• Significant deviations (design change, new affected components, phase changes) require returning to Plan mode. Stop and tell the user.
• Newly discovered work goes to Follow-ups. Do not add tasks in Implement mode.

Stop conditions

Stop and ask the user if:

• docs/ai/<feature_name>_plan.md does not exist
• docs/ai/<feature_name>_tasks.md does not exist and the plan does not set Execution Model to Mode: direct
• in task-driven mode, there are no unblocked tasks
• in task-driven mode, user-requested tasks are blocked or have unmet dependencies
• the implementation requires significant deviation from the plan
• in direct mode, the work cannot be completed in the current session (return to Plan mode and recommend switching to task-driven)
• verification fails in a way you cannot resolve within the selected scope

Done criteria

Implement mode checklist:

• in task-driven mode, each selected task has an accurate final status ([x], [~], [!], or [ ])
• one new ## Session Log entry appended in the devlog
• relevant verification commands run and results recorded

Review mode

When the user asks whether implementation matches the plan, you are in Review mode.

Review mode verifies plan fidelity and completion state. Check the plan against actual implementation, validate task markers if a tasks file exists, and update the devlog summary. This is not a general code review.

Allowed edits

You may:

• update ## Summary in docs/ai/<feature_name>_devlog.md (replace content)
• correct inaccurate status markers in docs/ai/<feature_name>_tasks.md (if it exists)

You must not:

• modify source code, tests, configs, or other project files
• modify docs/ai/<feature_name>_plan.md
• rename, reorder, add, or remove tasks
• rewrite past ## Session Log entries

If docs/ai/<feature_name>_devlog.md does not exist, read templates/devlog-template.md and create the devlog from it before updating ## Summary.

Goal

Determine whether implementation matches the documented plan, and report whether the current state is complete, partial, blocked, or divergent.

How to work in Review mode

1. Read docs/ai/<feature_name>_plan.md in full.
2. Read docs/ai/<feature_name>_tasks.md (if present) and docs/ai/<feature_name>_devlog.md (if present).
3. Inspect relevant implementation artifacts (source, tests, configs).
4. Compare implementation to the plan:
   • phase status: complete, partial, or missing
   • key design and behavioral decisions: followed or materially changed
   • task markers: accurate or inaccurate (if tasks file exists)
5. Correct inaccurate task markers (if tasks file exists).
6. Update devlog ## Summary status:
   • complete: planned scope implemented
   • partial: some planned work missing
   • blocked: remaining work is externally blocked
   • divergent: implementation materially differs from the plan
   • in progress: implementation is active but not review-complete
7. Report findings to the user with missing work and material deviations.

The Summary template is oriented toward service architectures. For other project types (frontend apps, CLI tools, libraries), adapt:

• "External changes" → public-facing changes (UI, routes, CLI flags, library API)
• "Consumer impact" → downstream impact on dependents, or omit if none
• "Unchanged surfaces worth noting" → omit if not relevant

Review focus

Prioritize plan fidelity and completion state: scope, phases, key decisions, missing work, and task-marker accuracy (if tasks file exists). Deprioritize style nits, unrelated cleanup, and speculative improvements.

Deviation policy

• Minor differences that preserve plan intent are acceptable. Note them.
• Significant differences must be called out explicitly; mark status divergent when material.
• If tasks claim completion not reflected in code, call out the mismatch and correct task markers.

Stop conditions

Stop and ask the user if:

• docs/ai/<feature_name>_plan.md does not exist
• the relevant code or artifacts needed for review are missing
• the plan is too vague to verify reliably against the implementation
• the repository state is inconsistent enough that implementation completeness cannot be determined with confidence

Done criteria

Review mode checklist:

• plan compared against actual implementation
• missing, partial, blocked, or divergent work clearly identified
• inaccurate task markers corrected (if tasks file exists)
• devlog ## Summary updated to current state
• user told final state: complete, partial, blocked, or divergent