Research plan implement

Research → Plan → Implement Workflow

A structured three-phase approach for complex tasks. Instead of diving straight into code (which works fine for simple changes), this workflow front-loads understanding so that implementation goes smoothly the first time. Each phase produces a compact artifact that carries forward, keeping context clean as work progresses.

When to use this skill

This skill earns its overhead when the task has uncertainty — you don't fully understand the codebase area, the right approach isn't obvious, or mistakes would be costly to undo:

• Complex features — New functionality touching multiple files and components
• Bug investigation — Issues where the root cause isn't obvious and needs tracing
• Architectural refactoring — Changes with ripple effects across the codebase
• Unfamiliar codebases — When you need to build understanding before acting

Skip this skill for: Simple fixes, typos, single-file changes, or tasks where the path forward is already clear.

Workflow overview

flowchart TB
    A[Start: Task/Feature] --> B[Phase 1: Research]
    B --> C[Research Document]
    C --> D{Ready to plan?}
    D -->|More research needed| B
    D -->|Yes| E[Phase 2: Plan]
    E --> F[Implementation Plan]
    F --> G{Plan approved?}
    G -->|Needs revision| E
    G -->|Yes| H[Phase 3: Implement]
    H --> I[Execute phase]
    I --> J[Verify & update progress]
    J --> K{More phases?}
    K -->|Yes| L[Next phase]
    L --> J
    K -->|No| M[Done]

    style B fill:#e1f5ff
    style E fill:#ffe1f5
    style H fill:#f5ffe1
    style C fill:#ffebcd
    style F fill:#ffebcd
    style J fill:#ffebcd

Core ideas

• TodoWrite tracks progress throughout the workflow
• Compaction points between phases keep context windows clean — each phase produces a document that carries the essential knowledge forward
• Human review at phase boundaries is where you get the most leverage: catching a bad research conclusion or flawed plan is far cheaper than catching bad code
• Phases are iterative — you can always loop back when implementation reveals new insights

---

Initial interaction

When invoked, greet the user and orient them:

# Research → Plan → Implement

I'll guide you through a structured workflow:

**Phase 1: Research** — Understand the codebase and document findings
**Phase 2: Plan** — Design a phased implementation approach
**Phase 3: Implement** — Execute with verification at each step

Each phase produces a document that carries context forward cleanly.

---

**Let's start with Phase 1: Research**

Please share:
- Your task, feature request, or bug description
- Any relevant files, tickets, or documentation
- Context about what you're trying to accomplish

Wait for the user's response before proceeding.

---

Phase 1: Research

Goal: Understand the codebase area relevant to the task and produce a structured research document with specific file references.

Output: A markdown research document (saved to the project's preferred docs location, or a sensible default like docs/research/ or the project root).

Process

1. Read user-mentioned files first — If the user points to specific files, read them fully before doing anything else. This gives you the context needed to ask the right questions of the codebase.

2. Decompose the research question — Break the task into distinct areas to investigate. Create a TodoWrite list to track research subtasks.

3. Investigate the codebase — Use sub-agents scaled to the task's complexity (see Adaptive Sub-Agent Usage below). The goal is to document what exists — describe how things work without critiquing or suggesting improvements.

4. Synthesize findings — Wait for all sub-agents to complete, then compile results. Prioritize what you found in the live codebase. Include specific file:line references for all significant findings.

5. Generate research document — Run ./scripts/spec_metadata.sh to collect git/date metadata, then write a structured document with: research question, summary, detailed findings, code references, and open questions.

6. Present findings — Share a concise summary with the user, highlighting key discoveries.

Phase transition

**Phase 1 Complete: Research Document Created**

I've documented findings in `[path to document]`

**Key findings:**
- [2-3 bullet summary]

Ready to create an implementation plan? Type "plan" to proceed, or ask for more research.

For the full research process, see references/phase1-research-guide.md

---

Phase 2: Plan

Goal: Transform research into an actionable implementation plan with clear phases, specific file changes, and testable success criteria.

Output: A markdown implementation plan with numbered phases and separated automated/manual verification criteria.

Process

1. Build on research — Load the Phase 1 document. If the user corrects anything, verify the correction in the codebase before proceeding.

2. Design interactively — Present design options with pros/cons. Get feedback on approach before writing details. The planning phase is a conversation, not a monologue.

3. Write the plan — Document phases with specific file changes. Separate success criteria into automated (commands that can be run) and manual (things requiring human judgment). Resolve all open questions — a plan with unknowns isn't ready.

4. Iterate until approved — Present the draft, incorporate feedback, repeat until the user is satisfied.

Phase transition

**Phase 2 Complete: Implementation Plan Created**

Plan is at `[path to document]`

**Summary:**
- [N] phases with specific file changes
- Automated and manual success criteria for each phase

Ready to implement? Type "implement" to proceed, or suggest plan adjustments.

For the full planning process, see references/phase2-planning-guide.md

---

Phase 3: Implement

Goal: Execute the plan phase-by-phase with verification after each phase.

Output: Working implementation with all automated checks passing.

Process

1. Load complete context — Read the plan and research document fully. Check for existing checkmarks to see what's already done.

2. Execute one phase at a time — Make changes, run automated verification, fix failures. Update checkboxes in the plan as items pass.

3. Pause for manual verification — After automated checks pass, stop and list the manual verification steps for the user. Wait for confirmation before proceeding (unless the user asked you to do multiple phases).

4. Handle mismatches — If reality doesn't match the plan, stop and explain: what the plan expected, what you found, why it matters, and options for how to proceed. Don't improvise without approval.

5. Manage context — For plans with 4+ phases, compact progress after each phase by adding implementation notes to the plan file.

After completion

**Implementation Complete**

All phases implemented and verified:
- All automated verification passing
- Implementation matches the plan

**Changes made:**
- [Brief summary]

**Next steps:**
- Perform final manual testing
- Review changes
- Create commit/PR when satisfied

For the full implementation process, see references/phase3-implementation-guide.md

---

Adaptive sub-agent usage

Not every task needs the same level of investigation. Scale your sub-agent usage to match the task's complexity:

Light investigation (1-2 agents)

For tasks where the affected area is small or mostly known:

• A single codebase-locator to find relevant files, then read them yourself
• Or a codebase-analyzer on a specific component you've already located

Standard investigation (2-4 agents in parallel)

For most tasks — run agents concurrently on different aspects:

• codebase-locator to find where things live
• codebase-analyzer to understand how specific code works
• codebase-pattern-finder to find similar existing patterns to model after

Deep investigation (4-6 agents in parallel)

For large, unfamiliar, or high-stakes tasks:

• All the standard agents, plus:
• web-search-researcher when external docs or API references are needed
• Multiple analyzer agents targeting different subsystems

Choosing the right level

Ask yourself: How much do I already know about the relevant code? If the user pointed you to specific files and the change is localized, light investigation is enough. If you're exploring unfamiliar territory across multiple subsystems, go deep.

Available agents

All agents are in the agents/ directory. Each knows its job — just tell it what you're looking for:

Agent	What it does	Tools
codebase-locator	Finds WHERE files and components live	Grep, Glob, LS
codebase-analyzer	Understands HOW specific code works	Read, Grep, Glob, LS
codebase-pattern-finder	Finds similar patterns to model after	Grep, Glob, Read, LS
thoughts-locator	Discovers relevant docs in thoughts/ directory	Grep, Glob, LS
thoughts-analyzer	Extracts insights from specific documents	Read, Grep, Glob, LS
web-search-researcher	Finds external docs and resources	WebSearch, WebFetch, Read

All agents are documentarians — they describe what exists without suggesting changes. This keeps their output focused and useful as reference material.

---

Why this workflow is structured this way

Front-loading understanding prevents rework

The cost of mistakes compounds through phases:

• Wrong research → Wrong plan → Hundreds of lines of misdirected code
• Wrong plan → Dozens of lines of incorrect implementation
• Wrong code → Localized fix in one component

By investing time in research and planning, you reduce the most expensive kind of rework.

Compaction points preserve context quality

AI context windows degrade as they fill. Each phase produces a compact artifact (research doc, plan) that carries essential knowledge forward without the noise of intermediate exploration. This means Phase 3 starts with clean, focused context rather than a window full of grep results.

Human review at boundaries, not in the middle

Reviewing a plan takes minutes. Reviewing 500 lines of code takes much longer. By gating phases on human approval, the user gets maximum leverage from their review time.

Iteration is cheap, not wasteful

Because each phase produces a standalone document, returning to an earlier phase doesn't throw away work — the artifacts from the first attempt still exist and inform the next iteration.

---

Workflow variations

Existing research available

If the user already has a research document, skip directly to Phase 2. Read the existing document and build on it — don't redo work that's already done.

Simple tasks that just need a plan

If the codebase area is well-understood and no exploration is needed, skip Phase 1 and go straight to planning.

Implementation reveals surprises

When Phase 3 uncovers something the plan didn't account for, communicate what you found and loop back to research or planning as needed. This is expected and healthy — reality is messy.