Generate balanced numerical values from your system designs' 7B inputs and the frozen data schema. The balance generator reads all 7B inputs, produces 4 balance documentation files (difficulty, economy, monetization, tuning), updates CSV config files with balanced values, then runs a quality gate to validate everything before completion.

Step 1: Load Project State and Check Prerequisites

Run:

node bin/gf-tools.cjs init progress

If project_exists is false: Display "No Game Forge project found. Run /gf:new-game to start." Stop.

Auto-mode detection: Check $ARGUMENTS for --auto flag. Store as AUTO_MODE (true/false). When AUTO_MODE is true, all user interaction is skipped and the entire balance stage runs in a single autonomous pass.

Check stages.data_schema -- if not complete and NOT AUTO_MODE: Display "Stage 3B requires data schema to be complete (frozen). Run /gf:data-schema first." Stop.
- If AUTO_MODE: Skip prerequisite check (the auto pipeline just completed data schema stage).

Check freeze status:

node bin/gf-tools.cjs data-schema freeze-status

If frozen is false and NOT AUTO_MODE: Display "Stage 3B requires a frozen schema. Run /gf:data-schema and freeze your schema first." Stop.
- If AUTO_MODE: Skip freeze check (the auto pipeline just froze the schema).

Load configuration:

node bin/gf-tools.cjs config-get language
node bin/gf-tools.cjs config-get genre

Store as LANGUAGE and GENRE.

Step 1.5: Determine Session Granularity

If AUTO_MODE: Skip session granularity calculation and split prompt entirely. Auto mode always uses a single pass regardless of system count. Proceed directly to Step 2.

Run:

node bin/gf-tools.cjs system-design system-status

Count confirmed systems from the response.

Adjust generation strategy:

3-8 systems: Single pass (current default) -- generate all balance docs at once
9+ systems: Consider splitting. Display to user: "Your game has {N} systems with 7B inputs. For best quality, consider generating balance in focused passes. Or continue with single pass for all systems." Wait for user preference.

Step 2: Detect State and Branch

Run:

node bin/gf-tools.cjs balance status

If AUTO_MODE: Always take the Generate path (not_started -> Step 3), then automatically continue through Review (Step 5) and Complete (Step 6) without stopping. The entire generate-review-complete pipeline runs in one pass.

Branch based on the returned status value:

not_started -> Go to Step 3 (Generate path)
in_progress -> Go to Step 5 (Review path)
complete -> Display completion message: "Stage 3B Numerical Balance is complete. Quality gate has passed. To re-run the quality gate, reply 'review'. To start fresh, reply 'regenerate'." Wait for user input.
- If user says "review" -> Go to Step 5
- If user says "regenerate" -> Go to Step 3
- Otherwise -> Display progress and Stop.

Step 3: Extract 7B Inputs and Gather Context

3a. Extract 7B Inputs

Run:

node bin/gf-tools.cjs balance extract-7b

Parse the JSON response. Display summary: "Found {N} systems with 7B balance inputs."

List each system briefly: system ID, system name.

3b. Read Frozen Schema

Read the frozen schema files for context:

.gf/stages/03a-data-schema/tables.md -- table structures and current sample data
.gf/stages/03a-data-schema/relationships.md -- table relationships

3c. Read Genre for Target Ranges

The GENRE value determines which column in the genre-specific target ranges (lifecycle-phases.md) the balance generator uses:

Casual, Puzzle, Match-3 -> Casual / Puzzle column
Action, Idle, Clicker, Arcade -> Action / Idle column
RPG, Strategy, Simulation, Roguelike -> RPG / Strategy column

Step 4: Spawn Balance Generator Agent

Spawn .claude/agents/gf-balance-generator.md via the Task tool with:

You are generating complete numerical balance documentation and updating CSV config files with balanced values.

**7B Balance Inputs:**
{paste full JSON from balance extract-7b}

**Genre:** {GENRE} -- use the appropriate genre column from lifecycle-phases.md for target ranges
**Language:** {LANGUAGE} -- ALL balance content must be in this language (descriptions, rationales, risk assessments). Parameter names, field names, system IDs, and enum keys remain in English.

**Frozen schema files to read:**
- .gf/stages/03a-data-schema/tables.md (table structures, field definitions, current sample data)
- .gf/stages/03a-data-schema/relationships.md (table relationships)

**System design files to read for 7B cross-reference:**
{list all .gf/stages/02-system-design/systems/*.md files}

**Reference files to read:**
- .claude/skills/gf-balance/references/lifecycle-phases.md (lifecycle phase definitions with genre-specific targets)

**Template files to read (output format specifications):**
- .claude/skills/gf-balance/templates/difficulty-template.md
- .claude/skills/gf-balance/templates/economy-template.md
- .claude/skills/gf-balance/templates/monetization-template.md
- .claude/skills/gf-balance/templates/tuning-template.md

**Output directory:** .gf/stages/03b-balance/
**CSV update location:** .gf/stages/03a-data-schema/configs/

**CLI commands available:**
- node bin/gf-tools.cjs balance update-csv --table "TABLE_NAME" --data '[...]'  (update CSV data rows)
- node bin/gf-tools.cjs balance validate-economy --flows '{...}'  (validate economy flows)
- node bin/gf-tools.cjs balance validate-freeze  (verify schema integrity after updates)

**Output files to produce:**
1. difficulty.md -- numerical target matrix + per-segment difficulty tables with inline adjustment strategies
2. economy.md -- resource flow tables + cross-resource dependencies + prohibited drought points
3. monetization.md -- protected parameters list + experience protection rules + commercialization boundary summary
4. tuning.md -- tuning priority table + validation thresholds + rollback strategy + self-check conclusions

**Constraints:**
- You may ONLY modify: sample data row values, default value fields, validation range fields.
- You may NOT add/remove/rename tables, fields, types, PKs, or FKs.
- Every balanced value MUST trace to a 7B input category. Values without upstream justification will be rejected by the quality gate.
- All output uses structured tables, not narratives.
- 7B Category 8 (no-go zones) defines problems that must NOT be solved by tuning numbers -- do not generate values for these.
- After updating CSV files, run: node bin/gf-tools.cjs balance validate-freeze

4a. Post-Generation

After the agent completes:

Update balance status to in_progress:

node bin/gf-tools.cjs state patch balance_status in_progress

Display: "Balance documentation generated with 4 balance files and CSV value updates. Review the output, then run /gf:balance again to validate with the quality gate."

Show progress:

node bin/gf-tools.cjs progress full

Step 5: Review and Quality Gate

Display: "Running balance quality gate..."

5a. Summarize Generated Balance Files

Read and display a brief summary of what was generated:

Count balance files in .gf/stages/03b-balance/ (difficulty.md, economy.md, monetization.md, tuning.md)
List which files exist and their status from frontmatter

If AUTO_MODE: Skip the "review and approve" prompt. Proceed directly to Step 5b (extract 7B inputs for quality gate).

If NOT AUTO_MODE: Ask user: "Review the balance documentation and approve to run the quality gate, or request specific adjustments."

If user requests adjustments: note the feedback, re-run the generator (Step 4) with the feedback added to the agent prompt
If user approves (or says "continue", "approve", "ok", "yes"): proceed to Step 5b

5b. Extract 7B Inputs for Traceability

node bin/gf-tools.cjs balance extract-7b

5c. Spawn Quality Reviewer Agent

Spawn .claude/agents/gf-balance-quality-reviewer.md via the Task tool:

Run the balance quality gate.

**Balance files location:** .gf/stages/03b-balance/
  - difficulty.md (numerical target matrix + per-segment difficulty)
  - economy.md (resource flow tables + cross-resource dependencies)
  - monetization.md (protected parameters + experience protection rules)
  - tuning.md (tuning priority table + validation thresholds + rollback + self-check)

**Quality criteria:** .claude/skills/gf-balance/references/balance-quality-criteria.md
**Language:** {LANGUAGE} -- REVIEW.md content in this language

**7B Balance Inputs (for traceability check):**
{paste full JSON from balance extract-7b}

**CLI commands available:**
- node bin/gf-tools.cjs balance validate-economy --flows '{...}'  (validate economy sustainability)
- node bin/gf-tools.cjs balance validate-freeze  (verify schema integrity)

**Instructions:**
- Read all 4 balance files from .gf/stages/03b-balance/
- Read balance-quality-criteria.md for the 8 auto-fix + 4 must-ask classification rules
- Run 8 auto-fix checks (structural): fix issues directly, log all fixes
- Run 4 must-ask checks (creative): flag for user decision, NEVER auto-fill
- Run economy validation via balance validate-economy CLI command
- Run freeze integrity via balance validate-freeze CLI command
- Run self-check validation from tuning.md Section 4
- Write REVIEW.md to .gf/stages/03b-balance/REVIEW.md
- After auto-fixes: run balance validate-freeze to confirm schema integrity
{IF AUTO_MODE:}
**AUTO MODE:** For must-ask items, use AI best judgment instead of flagging for user decision. Apply your recommendation directly. Log what you decided in REVIEW.md under a new 'Auto-Resolved Decisions' section.
{END IF}

5d. After Quality Gate

If AUTO_MODE:

Skip presenting REVIEW.md to user. Skip collecting user decisions.
Proceed directly to Step 6 (mark complete).

If NOT AUTO_MODE: Read and display .gf/stages/03b-balance/REVIEW.md contents to the user.

If REVIEW.md has "Needs User Decision" items:

Present each decision to the user and collect their answers.
Apply their decisions to the relevant balance files.
Re-run the quality gate to verify all issues are resolved.

If all checks pass (no user decisions needed):

Display: "Balance validated. Proceeding to completion."
Go to Step 6.

Show progress:

node bin/gf-tools.cjs progress full

Step 6: Mark Complete and Verify

6a. Verify Freeze Integrity

Run freeze integrity check:

node bin/gf-tools.cjs balance validate-freeze

If intact is false: Display "WARNING: Schema structure has been modified during balance work. This must be resolved before completion." Stop.

6b. Update State

node bin/gf-tools.cjs state patch balance_status complete
node bin/gf-tools.cjs state update balance complete

6c. Commit Balance Files

node bin/gf-tools.cjs commit "feat(balance): complete Stage 3B numerical balance" --files .gf/stages/03b-balance/ .gf/stages/03a-data-schema/configs/

If AUTO_MODE: Display: "Balance stage complete (auto mode). Proceeding to next stage..."

If NOT AUTO_MODE: Display: "Stage 3B Numerical Balance complete. Run /gf:production for Stage 4."

Step 7: Display Progress

node bin/gf-tools.cjs progress full

Constraints

These rules are locked decisions and must never be violated:

Automated generation from 7B inputs: The balance generator reads all 7B inputs at once and produces the complete balance documentation. No per-category questioning -- 7B inputs provide sufficient input for automated generation.
4 output files by concern: difficulty.md (numerical target matrix + per-segment tables), economy.md (resource flow tables + cross-resource dependencies), monetization.md (protected parameters + experience protection), tuning.md (tuning priority + validation thresholds + rollback + self-check).
Frozen schema boundary: You may ONLY modify: sample data row values, default value fields, validation range fields. You may NOT add/remove/rename tables, fields, types, PKs, or FKs. This is the #1 constraint from the source specification.
Tables over narratives: All balance output uses structured tables. Prose is limited to brief notes beneath tables.
No ghost values: Every balanced value must trace to a 7B input category from a Stage 2 system. Values without upstream justification are rejected by the quality gate.
CSV updates via CLI: CSV config files are updated via node bin/gf-tools.cjs balance update-csv, which preserves BOM, field names, and field types rows. Never manually edit CSV files.
Quality gate before completion: The quality reviewer must validate all balance files before completion. 8 auto-fix categories (structural) + 4 must-ask categories (creative decisions) per balance-quality-criteria.md.
7B Category 8 exclusion: No-go zones from 7B Category 8 define problems that must NOT be solved by tuning numbers. The balance generator skips these entirely.
State-driven 3-way branching: Orchestrator uses balance status to determine flow: generate (not_started), review (in_progress), complete (complete). Each invocation picks up where the previous left off.
Freeze verification at completion: Before marking balance complete, verify freeze integrity via node bin/gf-tools.cjs balance validate-freeze. If schema structure was modified, completion is blocked.
Auto mode: When --auto flag is present, all user interaction is skipped. Session split prompts are bypassed (always single pass). Review approval prompt is skipped. Quality gate must-ask items are resolved by AI judgment. The structural quality (traceability, freeze integrity, economy validation) is identical to interactive mode.