Project Research

Overview

Researches concepts and ideas through multiple external sources: web search for market/competitor/trend data, Context7 for technical documentation, and optionally codebase analysis. Works with any concept input — from thinking pipeline output, existing documents, or direct input.

Flow: extract research questions → auto-select technique → execute multi-source research → synthesize → loop or generate report.

Workflow

Step 1: Parse Input

Auto-detect concept file:

1. Check if .project/project-seed.md exists (primary) or .project/project.json has non-empty concept.content (legacy fallback)
2. If found AND no inline input provided:
   • Show concept name and ask confirmation:

     header: "Load Concept"
     question: "Do you want to research this concept?"
     options:
       - label: "Yes, research this (Recommended)", description: "Use concept from project.json"
       - label: "Different concept", description: "I want to paste a different concept"
     multiSelect: false
     

Scope check — only if scope context exists (backlog, features, or page files):

Check .project/backlog.json, .project/features/, and app/**/page.tsx / src/pages/**/*.tsx. If at least one found:

header: "Scope"
question: "What do you want to research?"
options:
  - label: "Concept (Recommended)", description: "Work with concept from project.json"
  - label: "Feature from backlog", description: "Focus on a specific feature"
  - label: "Page / UX flow", description: "Focus on layout, UX or user flow"
  - label: "Standalone idea", description: "Standalone idea, not linked to the project"
multiSelect: false

Scope-specific input:

• Feature: Read backlog (see shared/BACKLOG.md), show features TODO/DEF, load 01-define.md or feature description
• Page/UX: Glob for page files, load as context
• Standalone idea: Ignore loaded concept, ask user for description

Output path follows scope:

• concept → .project/project-seed.md + update project.json metadata (name, pitch)
• feature → .project/features/{name}/research.md
• page/UX or standalone idea → .project/thinking/{topic}-research.md

Other input — if no concept found or user wants different input:

Accept input from user (thinking output, document, description, or chat context). If unclear: ask 2-3 targeted questions. Confirm the concept summarized before proceeding.

Enter Plan Mode

Follow shared/PLAN-MODE.md Entry protocol before Step 2. Steps 2-7 run in plan mode; the final report (Step 7) is written to the plan file for review. WebSearch, Context7 and Grep/Glob continue to work in plan mode.

Step 2: Extract Research Foundation

First extract four categories from the concept, then distill research questions.

2a. Extract foundation (4 categories, brief, one line per item):

• Assumptions — what are we taking for granted that we may need to validate?
• Knowledge gaps — what do we simply not know? (market, competitors, pricing, user behavior)
• Decisions — what choices do we need to make and what info do we need for them? (tech stack, pricing, target audience)
• Risks — what could cause this concept to fail? (technical, legal, scalability, adoption)

2b. Distill 3-5 research questions with the highest impact on feasibility. Questions can come from any category — usually a mix. Each question must be answerable via web search, documentation, or codebase analysis and be specific to THIS concept.

2c. Store assumptions separately (including those that don't make it into research questions) — these will be marked as Validated / Invalidated / Still Open in Step 5/7.

Present:

RESEARCH FOUNDATION

Concept: {concept title}

Assumptions:
- {assumption 1}
- {assumption 2}

Knowledge Gaps:
- {gap 1}
- {gap 2}

Decisions Needed:
- {decision 1}

Risks:
- {risk 1}
- {risk 2}

RESEARCH QUESTIONS (distilled from the above)

1. {specific research question — highest priority}
2. {specific research question}
3. {specific research question}
4. {specific research question} (optional)
5. {specific research question} (optional)

Confirm via AskUserQuestion (Yes/Adjust).

Step 3: Select Research Technique

Read references/research-techniques.md. Automatically select the most relevant technique based on:

• Which technique addresses the highest-priority unanswered questions?
• Which techniques have already been applied? (exclude)

Show the choice briefly:

RESEARCH TECHNIQUE: {Technique Name}

{1-2 sentences explaining why this technique was chosen}

Proceed directly to Step 4.

Step 4: Execute Research

Read the selected technique from references/research-techniques.md.

Execute research sources in parallel where possible:

• WebSearch: 2-4 targeted queries derived from the technique
• Context7: resolve-library-id → query-docs for technical questions
• Codebase: Grep/Glob if the concept involves an existing project

Present findings per source:

WEB|DOCS|CODE — {SOURCE NAME}

Query: "{search query}"

Findings:
- {key finding 1}
- {key finding 2}

Sources: {URLs or file paths}

Offer to go deeper if findings reveal unexpected angles.

Guidelines:

• Cross-reference between sources
• Flag contradictions and challenged assumptions
• Quantitative data where available
• Always cite sources

For Trend Analysis technique: Use the mental models from the technique as an operational lens during execution, not as post-analysis. Explicitly note temperature differences per source and label each signal with trajectory vocabulary (acute_rise / plateau / zombie / comeback).

Step 5: Synthesize Findings

Map findings to the research questions from Step 2:

SYNTHESIS — {Technique Name}

### Q1: {research question}
**Status:** {Answered / Partially answered / Needs more research}
- {key finding with source reference}
**Implication:** {what this means for the concept}

### Q2: {research question}
...

### Assumption Tracking
- {assumption}: **Validated** — {evidence + source}
- {assumption}: **Invalidated** — {evidence + source}
- {assumption}: **Still Open** — {what's needed to validate}

### Decision Tracking
- {decision}: **Ready** — {recommended option + rationale}
- {decision}: **Blocked** — {what's still needed}

### Risk Tracking
- {risk}: **Severity High/Med/Low** — {mitigation OR "still open"}

### New Insights
- {unexpected finding or new angle discovered}

Proceed to Step 6.

Step 6: Next Action

First build an open-items overview based on Step 5 tracking:

Open items overview:
- Unanswered questions: {n}
- Still-Open assumptions: {n}
- Blocked decisions: {n}
- Unaddressed High/Med risks: {n}

Total open items: {sum}

Decision logic:

• Total = 0 or no relevant techniques remaining → go directly to Step 7
• Total > 0 → show AskUserQuestion. Determine the recommended technique by:
  1. Select the most urgent type of open item in order: High-severity risks > Blocked decisions > Still-Open assumptions > Unanswered questions
  2. Match the type with the Addresses: mapping from references/research-techniques.md
  3. First option remains "Generate report (Recommended)" so the user can always close out

Show the reasoning briefly before the question:

Recommendation: {Technique} — addresses {N} open {item type}(s)

Ask via AskUserQuestion:

header: "Next Step"
question: "How do you want to continue?"
options:
  - label: "Generate report (Recommended)", description: "Create the final result with all findings"
  - label: "{Technique}", description: "{rationale — addresses {N} open {item type}}"
multiSelect: false

Step 7: Generate Final Report

Generate a structured markdown report:

# Research: {concept title}

## Summary

{2-3 sentence executive summary}

## Key Findings

### {Finding Category}

{findings with evidence and source references}
**Implication:** {what this means for the concept}

## Assumptions

**Validated:**

- {assumption} — {evidence}

**Invalidated:**

- {assumption} — {evidence, what to reconsider}

**Still Open:**

- {assumption} — {what's needed to validate}

## Decisions

**Ready to decide:**

- {decision} — {info gathered, recommended option, rationale}

**Still blocked:**

- {decision} — {what's still needed before you can choose}

## Risk Assessment

| Risk   | Severity     | Likelihood   | Mitigation                     |
| ------ | ------------ | ------------ | ------------------------------ |
| {risk} | High/Med/Low | High/Med/Low | {strategy or "needs research"} |

## Recommendations

Recommendations are always actor-specific (e.g. "for a B2B SaaS brand", "for a solo creator"). No generic "keep monitoring" advice — every recommendation contains a concrete action or decision.

1. {actionable recommendation}

## Open Questions

- {unanswered or emerged question}

## Sources

- [{source title}]({url})

Include Competitive Landscape table and/or Technical Feasibility assessment where relevant.

End of analysis phase: follow shared/PLAN-MODE.md Exit protocol — write the research report to the plan file, then ExitPlanMode. After approval the skill continues with Step 8 (Save & Sync, .project/ writes).

Step 8: Save & Sync

Scope = concept (default):

Auto-save without extra questions:

1. Write report to .project/thinking/{concept-name}-research.md (full report as user archive)
2. Add ## Research Findings section to .project/project-seed.md (this is what project-plan/dev-ship's define phase see as context)
3. Confirm:

RESEARCH SAVED

Report: .project/thinking/{concept-name}-research.md
Concept: project-seed.md updated with key findings
Applied techniques: {list}

Next steps:
- /project-seed critique - Critically analyze with research context
- /project-seed brainstorm - Creatively expand with new insights
- /project-plan - Convert to feature backlog

Scope = feature or page:

1. Write to scope path (.project/features/{name}/research.md or .project/thinking/{topic}-research.md)
2. Optionally ask if key findings should also be added to project-seed.md

The markdown is the source of truth — no project.json thinking[] append. Skills that want to consume research (like /dev-ship (define phase)) read directly from .project/thinking/*.md or .project/features/{name}/research.md.

Scope = standalone idea:

1. Write to .project/thinking/{topic}-research.md
2. Offer "Copy to clipboard" per shared/CLIPBOARD.md

The markdown is the source of truth — no project.json thinking[] append.

Guidelines

Flow:

• No AskUserQuestion between research execution steps — just execute and present
• After synthesis always go to Step 6 for next action
• One technique at a time: select → execute → synthesize → decide

Flexibility:

• User can adjust research focus at any time
• With unexpected findings: offer to explore that path

Formatting:

• NEVER use blockquote syntax (>) — unreadable background in dark terminals
• NEVER use backticks for emphasis on regular words — use bold
• Backticks only for code, file paths, and command references

Language: Follow the Language Policy in CLAUDE.md.