STOP — DO NOT READ THIS FILE. You are already reading it. This prompt was injected into your context by Claude Code's plugin system. Using the Read tool on this SKILL.md file wastes ~7,600 tokens. Begin executing Step 1 immediately.

Step 0 — Immediate Output

Before ANY tool calls, display this banner:

╔══════════════════════════════════════════════════════════════╗
║  PLAN-BUILD-RUN ► DISCUSSION                                 ║
╚══════════════════════════════════════════════════════════════╝

Then proceed to Step 1.

/pbr:discuss-phase — Pre-Planning Discussion

References: @references/questioning.md, @references/ui-brand.md

Reference: skills/shared/context-budget.md for the universal orchestrator rules. Reference: skills/shared/context-loader-task.md for efficient context loading patterns.

You are running the discuss skill. Your job is to help the user think through a phase BEFORE planning begins. You identify gray areas where the user's preference matters, ask structured questions, and capture every decision in a CONTEXT.md file that the planner must honor.

This skill runs inline (no Task delegation).

Multi-Session Sync

Before any phase-modifying operations (writing CONTEXT.md, updating STATE.md), acquire a claim:

acquireClaim(phaseDir, sessionId)

If the claim fails (another session owns this phase), display: "Another session owns this phase. Use /pbr:progress to see active claims."

On completion or error (including all exit paths), release the claim:

releaseClaim(phaseDir, sessionId)

Core Principle

Decisions made here are LOCKED. The planner cannot override them, even if research suggests a different approach. The planner must implement the locked decision and may note the discrepancy, but must follow the user's choice exactly.

Flow

Step 1: Parse Phase Number and Check for Existing Plans

CRITICAL (hook-enforced): Write .active-skill NOW. Write the text "discuss" to .planning/.active-skill using the Write tool.

Parse $ARGUMENTS:

If argument is --project: enter PROJECT mode (see Step 1-project below). Skip Steps 2-8.
If argument is a phase number: enter PHASE mode (existing flow — continue with Step 1 as-is).
If no argument: existing logic applies (read STATE.md for current phase).
If --auto is present in $ARGUMENTS: set auto_mode = true. Log: "Auto mode enabled — batching gray area questions"

Argument	Meaning
`3`	Discuss phase 3
`3 --auto`	Discuss phase 3 with auto mode — batch all gray areas into a single presentation
`--project`	Discuss project-level cross-cutting decisions

Validation (PHASE mode):

Must be a valid phase number (integer or decimal like 3.1)
If no argument provided, read STATE.md to get the current phase
If no current phase and no argument: "Which phase do you want to discuss? Run /pbr:progress to see available phases."

Phase directory resolution:

List directories in .planning/phases/
Find directory matching pattern {NN}-* where NN matches the phase number
If not found, check ROADMAP.md for the phase name
If still not found, display:

╔══════════════════════════════════════════════════════════════╗
║  ERROR                                                       ║
╚══════════════════════════════════════════════════════════════╝

Phase {N} not found.

**To fix:** Run `/pbr:progress` to see available phases.

Check for existing plans (after resolving the phase directory):

Check for PLAN.md or PLAN-*.md files in the phase directory
If plan files exist:
- Warn: "Phase {N} already has plans. Decisions from this discussion won't retroactively change them. Consider re-planning with /pbr:plan-phase {N} after."
- This is a warning only — do not block the discussion

Step 1-project: Project Discussion Mode (--project)

When invoked with --project, discuss project-level cross-cutting decisions. This mode writes to .planning/PROJECT.md ## Context section (project-level), NOT a phase directory.

Check for existing project context:

Read .planning/PROJECT.md and check if it has a ## Context section.
If it has a ## Context section: CRITICAL -- DO NOT SKIP: Present the following choice to the user via AskUserQuestion before proceeding: Ask the user (using the context-handling pattern from skills/shared/gate-prompts.md): question: "PROJECT.md already has a ## Context section. How should we handle it?" options: Overwrite | Append | Cancel
If Cancel: stop and display the existing PROJECT.md path.
Backwards compat: If .planning/CONTEXT.md exists (legacy), read it as the existing context source for Append mode, then write merged result to PROJECT.md ## Context.

Load project context:

Read .planning/PROJECT.md (if exists) — project vision, scope, and context (## Context section)
Read .planning/REQUIREMENTS.md (if exists) — requirements for constraint awareness
Backwards compat: If PROJECT.md has no ## Context but .planning/CONTEXT.md exists, read from the legacy file

Run gray areas for project-level decisions (Steps 2.5-5 pattern):

Identify 3-4 cross-cutting architectural decisions across ALL phases
Focus on: technology stack choices, infrastructure, security posture, observability approach, deployment strategy, data storage decisions
Follow the same Steps 3-5 pattern (gray areas → options → follow-ups)

Write project context to PROJECT.md:

Read ${CLAUDE_SKILL_DIR}/templates/project-CONTEXT.md.tmpl
Fill in from discussion decisions:
- Locked Decisions table: all decisions the user made (not "Let Claude decide")
- User Constraints: budget, team, timeline, hosting from the conversation
- Deferred Ideas: anything explicitly ruled out for this project
- Claude's Discretion: decisions marked "Let Claude decide"
If Append mode: merge new decisions with existing PROJECT.md ## Context content CRITICAL: Write the discussion results to PROJECT.md NOW.
Write to .planning/PROJECT.md under the ## Context section with subsections:
- ### Locked Decisions
- ### User Constraints
- ### Deferred Ideas

Update STATE.md pointer: Add under Accumulated Context:

Project context: .planning/PROJECT.md ## Context ({N} locked decisions, {N} deferred, {N} discretion)

After writing: Auto-sync to CLAUDE.md is handled by the post-write-dispatch hook. Display: ✓ Project context written to PROJECT.md — locked decisions will auto-sync to CLAUDE.md

Skip to Cleanup step. Do NOT run Steps 2-8 of the phase flow.

Step 2: Load Phase Context

Reference: skills/shared/context-loader-task.md for efficient context loading patterns.

Read the following files to understand what this phase needs to accomplish:

ROADMAP.md — Find the phase entry. Extract:
- Phase name and description
- Goal / objective
- Requirements to satisfy
- Success criteria
- Dependencies on prior phases
Prior SUMMARY.md files — Scan .planning/phases/ for phases with lower numbers that have SUMMARY.md files. Read their frontmatter to understand:
- What's already been built (provides field)
- What technologies are in use (tech_stack field)
- What patterns have been established (patterns field)
- What decisions were already made (key_decisions field)
REQUIREMENTS.md (if exists) — Read project requirements relevant to this phase. CRITICAL: After reading, display to the user which specific requirements map to this phase. Flag any requirements that could constrain decisions. If the user's discussion decisions later contradict a stated requirement, surface the contradiction immediately — don't wait until planning.
CONTEXT.md (if exists in the phase directory) — Check if a prior discussion already happened
- If CONTEXT.md exists, inform the user and: CRITICAL -- DO NOT SKIP: Present the following choice to the user via AskUserQuestion before proceeding: Use the context-handling pattern from skills/shared/gate-prompts.md: question: "Phase {N} already has a CONTEXT.md from a prior discussion. How should we handle it?"
Handle responses:
- "Overwrite": Replace CONTEXT.md entirely with new decisions
- "Append": Add new decisions below existing ones, marked as "Amendment ({date})"
- "Cancel": Stop the discussion, keep existing CONTEXT.md

Step 2.5: Open Exploration Phase

Before jumping into specific gray areas, give the user space to share their mental model unprompted.

Present the phase goal (from ROADMAP.md) and ask an open question:
- "Before we get into specifics — what's your mental model of how this phase should work? What excites you about it? What concerns you?"
Follow the user's response with 2-3 domain-aware follow-ups from skills/shared/domain-probes.md:
- Match what the user mentioned to the relevant domain tables
- Pick the 2-3 most insightful probes based on what they said, not a generic checklist
- Ask them conversationally, not as a numbered list
Surface implications from what the user shared:
- "You mentioned X — that usually means we'd need Y. Is that part of your vision?"
- Connect their ideas to concrete technical or design consequences
- Flag anything that would affect scope, complexity, or dependencies
Let the conversation reveal gray areas naturally. The user's responses here often surface the real gray areas better than top-down analysis. Carry any themes, concerns, or preferences forward into Step 3.

Step 3: Identify Gray Areas

Analyze the phase goal, requirements, and what's already built. Identify 3-4 gray areas where the user's preference matters. Gray areas fall into these categories:

Read ${CLAUDE_SKILL_DIR}/templates/decision-categories.md for the category reference table.

How to identify gray areas:

Look at the phase requirements — where are there multiple valid approaches?
Look at prior phase decisions — where might this phase need to diverge?
Look at the success criteria — what's ambiguous about how to satisfy them?
Consider the user's likely pain points — what would they want a say in?

Important: Do NOT identify gray areas that are purely implementation details (e.g., variable naming, file organization). Focus on areas that affect user experience, system behavior, or long-term maintainability.

Step 4: Present Gray Areas

Smart batching (auto_mode): When auto_mode is true, collect ALL gray areas and present them in a single batch rather than one at a time. Present all gray areas together with their options so the user can respond to everything at once. This reduces interaction rounds while still capturing user decisions. After the user responds, proceed to Step 5 with all decisions.

Standard mode (auto_mode is false or absent):

CRITICAL -- DO NOT SKIP: Present the following choice to the user via AskUserQuestion before proceeding: Present each gray area to the user using the gray-area-option pattern from skills/shared/gate-prompts.md. For each gray area:

Use AskUserQuestion: question: "Gray Area {N}: {Title} — {Why this matters}" header: "Decision" options: (generate 2-4 concrete options from analysis, each with pros/cons in the description) - label: "{Option A}" description: "{Pros: ..., Cons: ...}" - label: "{Option B}" description: "{Pros: ..., Cons: ...}" - ...up to 4 options total, with "Let Claude decide" always as the last option - label: "Let Claude decide" description: "Mark as Claude's Discretion" multiSelect: false

If more than 3 concrete options exist for a gray area, present only the top 3 plus "Let Claude decide" (max 4 total). Mention the omitted option(s) in the question text.

Rules for presenting options:

Each option must be concrete and implementable
Include honest pros/cons (don't bias toward one option)
"Let Claude decide" should always be available as a choice
If the user gives a preference not in the options, accept it
If the user says "I don't care", mark it as Claude's Discretion

Step 4b: Advisor Research (optional, for technology decisions)

If any gray area involves a technology choice (library selection, framework decision, architecture pattern) and the user expresses uncertainty or asks "which is better?", spawn a focused research agent:

Task({
  subagent_type: "pbr:advisor-researcher",
  prompt: "<files_to_read>
    1. .planning/PROJECT.md — project context and constraints
    2. .planning/CONTEXT.md — existing locked decisions (if exists)
  </files_to_read>
  Decision area: {gray area title}
  Context: {project constraints, existing stack}
  Options to compare: {options from Step 4}
  Return a structured comparison table with recommendation."
})

After the advisor returns, present the comparison table to the user and let them make the final decision. The user's choice becomes the locked decision — the advisor's recommendation is advisory only.

Skip this step if: No gray areas involve technology choices, or the user already made confident decisions in Step 4.

Step 5: Deep-Dive Each Selected Area

For each gray area where the user made a decision (not "Let Claude decide"), ask 4 follow-up questions to fully capture their intent.

CRITICAL — STOP: Do NOT skip ANY of the 4 follow-up areas below. All 4 MUST be asked for each gray area where the user made a decision.

Follow-up question types:

CRITICAL — STOP: Do NOT skip this follow-up area.

Scope boundary: "Should {feature} also handle {edge case}?" CRITICAL -- DO NOT SKIP: Present the following choice to the user via AskUserQuestion before proceeding: Use the yes-no pattern — this is a binary decision.

CRITICAL — STOP: Do NOT skip this follow-up area. 2. Quality level: "How polished should this be?" Do NOT use AskUserQuestion — this is freeform. Let the user describe their quality expectations in their own words.

CRITICAL — STOP: Do NOT skip this follow-up area. 3. Integration: "How should this interact with {existing component}?" Do NOT use AskUserQuestion — this is freeform. The answer depends on the specific component and context.

CRITICAL — STOP: Do NOT skip this follow-up area. 4. Future-proofing: "Should we design this to support {potential future need}, or keep it simple?" CRITICAL -- DO NOT SKIP: Present the following choice to the user via AskUserQuestion before proceeding: Use the yes-no pattern: question: "Design {feature} to support {future need}, or keep it simple for now?" options: - label: "Future-proof" description: "Add extensibility for {future need}" - label: "Keep simple" description: "Build only what's needed now"

Rules for follow-ups:

Ask all 4 questions for each area
Record exact answers (don't paraphrase)
If the user gives a short answer, capture it as-is
If the user says "you decide" on a follow-up, move that specific sub-decision to Claude's Discretion

Completion check: After all follow-up questions for all gray areas, verify that all 4 follow-up areas produced output for each decided gray area. If any area was missed, go back and ask it before proceeding to Step 6.

Step 6: Capture Deferred Ideas

Reference: The deferred section prevents scope creep in downstream planning. See CONTEXT.md template.

During the conversation, the user may mention ideas they want but not in this phase. Track these separately:

Ideas explicitly deferred by the user ("let's do that later")
Ideas that are out of scope for this phase based on ROADMAP.md
Ideas the user considered but rejected

Step 6.5: Decision Gate (inline)

Before creating the CONTEXT.md artifact, confirm readiness with the user per @references/questioning.md decision gate pattern:

AskUserQuestion:
  question: "I think I have a clear picture of your decisions for this phase. Ready to capture them in CONTEXT.md?"
  header: "Ready?"
  options:
    - label: "Yes, proceed"    description: "Write CONTEXT.md with captured decisions"
    - label: "Keep exploring"  description: "I want to share more or discuss further"

If "Keep exploring" -- ask what they want to add or identify gaps and probe naturally. Loop until "Yes, proceed" is selected.

Freeform rule: If the user selects "Keep exploring" and responds with freeform text (not selecting an option), switch to plain text conversation per @references/questioning.md freeform rule. Do NOT present another AskUserQuestion until you have processed their freeform response.

Step 7: Write CONTEXT.md

CRITICAL — DO NOT SKIP: Write CONTEXT.md NOW. This file is the sole output of the discuss skill. Skipping it means all decisions are lost. Write the CONTEXT.md file to the phase directory:

Path: .planning/phases/{NN}-{phase-name}/CONTEXT.md

Content:

Read ${CLAUDE_SKILL_DIR}/templates/CONTEXT.md.tmpl for the template structure.

Placeholders to fill:

{N} -- the phase number
{ISO date} -- today's date in ISO format
{Decision title} / {User's exact choice and reasoning} -- from Step 4-5 decisions
{Idea} / {Why deferred} -- from Step 6 deferred ideas
{Area} / {What Claude can choose} -- items marked "Let Claude decide"
{2-3 sentences} -- user's vision summary from the conversation
{Concern} / {Context} -- concerns raised during discussion
<specifics> -- concrete technical details, patterns, or conventions from the conversation
<code_context> -- relevant files and modules referenced during discussion

Decision Summary generation: The ## Decision Summary section at the top of CONTEXT.md is a compact digest (~300 tokens) of all decisions. For each locked decision, write only the title and the user's choice in one phrase (no scope/quality/integration details). List deferred and discretion items as comma-separated titles only. This summary is injected into agent prompts by the plan skill -- keep it concise.

Downstream awareness: The CONTEXT.md written here is consumed by:

/pbr:plan-phase -- planner reads locked decisions as constraints
/pbr:execute-phase -- executor receives decisions via planner's plan
/pbr:verify-work -- verifier checks deliverables against phase intent

Ensure all locked decisions are written clearly enough for mechanical consumption.

Step 7.5: Update STATE.md Pointer

CRITICAL — DO NOT SKIP: Update STATE.md pointer to new CONTEXT.md NOW. Update .planning/STATE.md's ## Accumulated Context section to add a reference to the new CONTEXT.md:

Add under the ### Decisions subsection:

Phase {N} discussion: .planning/phases/{NN}-{slug}/CONTEXT.md ({count} locked, {count} deferred, {count} discretion)

This creates a pointer so /pbr:resume-work and progress-tracker.js know that phase-specific decisions exist and where to find them.

Step 7.7: Write Discussion Log (audit trail)

Write .planning/phases/{NN}-{slug}/DISCUSSION-LOG.md as an audit trail of the discussion. This file is NOT consumed by agents — it's for human audit only.

Read ${CLAUDE_PLUGIN_ROOT}/templates/DISCUSSION-LOG.md.tmpl for the format.

For each decision area discussed:

Record all options that were presented
Mark which option was selected
Include the user's verbatim choice and any notes/rationale
Record items delegated to Claude's discretion
Record deferred ideas

This log is separate from CONTEXT.md (which is the implementation artifact for agents). The discussion log preserves the alternatives considered, not just the final decisions.

If planning.commit_docs is true, include in the next commit.

Step 8: Confirm and Route

After writing CONTEXT.md, display branded output per @references/ui-brand.md:

╔══════════════════════════════════════════════════════════════╗
║  PLAN-BUILD-RUN ► DISCUSSION CAPTURED ✓                      ║
╚══════════════════════════════════════════════════════════════╝

**Phase {N}: {name}**

✓ Decisions: {count} locked, {count} deferred, {count} discretion

╔══════════════════════════════════════════════════════════════╗
║  ▶ NEXT UP                                                   ║
╚══════════════════════════════════════════════════════════════╝

**Plan this phase** — your decisions will be honored

`/pbr:plan-phase {N}`

<sub>`/clear` first → fresh context window</sub>

**Also available:**
- `/pbr:progress` — see project status
- `/pbr:explore` — explore ideas further

Cleanup: Delete .planning/.active-skill before finishing. This must happen on all paths (success, cancel, and error).

Decision Categories Reference

LOCKED Decisions

These come from:

User selecting a specific option (not "Let Claude decide")
User answering follow-up questions with specific preferences
User volunteering a strong opinion during discussion

The planner MUST:

Implement exactly as the user specified
Not modify, optimize, or "improve" the decision
Note in the plan if research suggests a different approach
Still follow the locked decision regardless

DEFERRED Ideas

These come from:

User explicitly saying "not now" or "later"
Ideas that are out of scope per ROADMAP.md
User rejecting an option but saying it's a good idea for later

The planner MUST NOT:

Include deferred ideas in the plan
Sneak deferred ideas in as "nice to have" tasks
Combine deferred ideas with in-scope work

CLAUDE'S DISCRETION

These come from:

User selecting "Let Claude decide"
User saying "I don't care" or "whatever you think"
Follow-up questions answered with "you decide"

The planner MAY:

Choose any reasonable approach
Use research results to inform the choice
Change their mind during planning if they learn something new
Must document their choice and rationale in the plan

Edge Cases

Phase already has CONTEXT.md

CRITICAL -- DO NOT SKIP: Present the following choice to the user via AskUserQuestion before proceeding: Use the context-handling pattern from skills/shared/gate-prompts.md (same as Step 2)
If "Overwrite": replace entirely
If "Append": add new decisions below existing ones, marking them as "Amendment"
If "Cancel": stop the discussion, keep existing CONTEXT.md

Phase already has plans

Handled by Step 1 — warn but do not block
See Step 1 "Check for existing plans" for the exact warning message

User wants to discuss multiple phases

Handle one at a time
After completing one, suggest: "Want to discuss Phase {N+1} too? Run /pbr:discuss-phase {N+1}."

User disagrees with all options

Ask: "What would you prefer instead?" — this is freeform text, do NOT use AskUserQuestion.
Accept any answer and lock it as a decision.
The options were suggestions, not constraints

User wants to skip follow-ups

If user says "that's enough" or "skip the details", respect that
Write what you have — partial follow-ups are fine
Mark missing follow-up areas as Claude's Discretion

Error Handling

Reference: skills/shared/error-reporting.md for error output patterns.

If any step fails (file read error, directory not found, write failure):

Display branded error using the pattern from error-reporting.md
Clear .planning/.active-skill before exiting
Do NOT leave partial CONTEXT.md files -- write atomically or not at all

State Integration

This skill updates STATE.md's Accumulated Context section with a pointer to the phase CONTEXT.md file. It does NOT change the project position (current phase/plan). STATE.md position is updated when /pbr:plan-phase runs.

Git Integration

Reference: skills/shared/commit-planning-docs.md for the standard commit pattern.

If planning.commit_docs: true in config.json:

Commit CONTEXT.md: docs(planning): capture phase {N} discussion decisions

Memory capture: Reference skills/shared/memory-capture.md — check researcher output for <memory_suggestion> blocks and save any reusable knowledge discovered during research.

Anti-Patterns

DO NOT bias the user toward a particular option
DO NOT skip the follow-up questions (unless the user asks to skip)
DO NOT paraphrase the user's decisions — capture them verbatim
DO NOT add your own opinions to LOCKED decisions
DO NOT mark something as Claude's Discretion unless the user explicitly chose that
DO NOT include implementation details in CONTEXT.md — it captures WHAT, not HOW
DO NOT skip the deferred ideas section — this prevents scope creep later
DO NOT present more than 4 options per gray area — decision fatigue is real