Installs into .claude/skills of the current project.
Are you the author of Forge?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/nhan03bit-forge)
# /forge — Product Builder Pipeline (GSD Template)
> GSD template for product-building lifecycle. Forge = GSD with product-specific presets.
> `clarify → feasibility → plan = gsd-plan | build = gsd-execute | validate = gsd-verify | improve = gsd-execute (fix wave) | ship = gsd post-verify`
## Relationship to GSD
Forge is **not a separate workflow** — it's a GSD template/preset with product-building phases.
All execution flows through GSD primitives:
| Forge Phase | GSD Equivalent |
|-------------|----------------|
| clarify → feasibility → plan | gsd-plan |
| build | gsd-execute |
| validate | gsd-verify |
| improve | gsd-execute (fix wave) |
| ship | gsd post-verify |
## Trigger
- Keywords: "forge", "build product", "build app", "create app", "ship", "from idea to product"
- When user has a product idea and wants structured execution
## Modes
- `--guided` (DEFAULT): Step-by-step, plain language, explains each phase before executing
- `--builder`: Enforced phase gates, validation required, auto-decision extraction
- `--full`: No gates, experienced users, maximum speed
## State
Forge state persists at `~/.ultrathink/forge/projects/<hash>.json` where hash = first 8 chars of SHA256 of the project working directory path.
**Read state at session start.** If forge-state exists for current project, resume from last stage.
**State format (JSON — never Markdown):**
```json
{
"project": "descriptive-slug",
"project_path": "/absolute/path",
"stage": "clarify|feasibility|plan|build|validate|improve|ship|complete",
"mode": "guided|builder|full",
"created": "ISO-8601",
"updated": "ISO-8601",
"spec": {
"target_user": "",
"problem": "",
"value_prop": "",
"stack": "",
"feasibility_score": null,
"complexity_score": null
},
"phases": [
{
"id": 1,
"name": "Phase name",
"features": [
{ "id": "1.1", "name": "Feature name", "passes": false }
]
}
],
"current_phase": 1,
"current_feature": "1.1",
"evaluation": {
"functionality": null,
"design": null,
"craft": null,
"originality": null,
"overall": null
},
"decisions": [],
"history": []
}
```
## Phase 1: CLARIFY
Ask the user 3-5 targeted questions to understand the product:
1. **Who** is the target user?
2. **What** specific problem does this solve?
3. **Why** is this better than existing solutions?
4. **What** is the MVP scope? (what's the ONE thing it must do?)
5. **What** stack preference? (or let forge decide)
In `--guided` mode: Explain why each question matters. Use plain language.
Output: Update `spec` in forge-state.json.
## Phase 2: FEASIBILITY
Score the idea:
- **Tech complexity** (1-5): How hard to build?
- **Novelty** (1-5): Does this exist already? What's new?
- **Time estimate**: Realistic timeline with the chosen stack
- **Risk factors**: What could go wrong?
In `--builder` mode: If `feasibility_score < 3`, STOP. Ask user to refine the idea or adjust scope.
Output: Update `spec.feasibility_score` and `spec.complexity_score`.
## Phase 3: PLAN
Generate a phased roadmap:
- 3-5 phases, each with 5-10 atomic features
- Each feature has a clear pass/fail test
- Order by dependency (build foundations first)
- Assign features to sprints
Use existing skills: invoke `/plan` for task decomposition, `/scout` for tech research.
In `--guided` mode: Show one phase at a time. Explain what each feature does in plain language. Ask "Ready to proceed?" before continuing.
Output: Update `phases` array in forge-state.json. Each feature starts with `"passes": false`.
## Phase 4: BUILD
For each feature in the current phase:
1. Read the feature spec from forge-state
2. Build it (write code, create files, install deps)
3. Run the project's test command
4. If tests pass → set `feature.passes = true`, commit, move to next
5. If tests fail → fix and retry (max 3 attempts per feature)
Chain existing skills as needed: `/react`, `/nextjs`, `/tailwindcss`, `/drizzle`, `/api-designer`, etc.
**One feature at a time.** Never build multiple features simultaneously.
**Commit after each feature.** Message: `feat(forge): <phase>.<feature> — <description>`
In `--guided` mode: Before building each feature, explain what you're about to do and why.
Output: Update `phases[n].features[m].passes` after each feature.
## Phase 5: VALIDATE
Run the evaluator against the current phase:
1. `npm run build` — must succeed
2. `npm run test` — must pass
3. Structural checks: files exist, routes respond, no console errors
4. If user has configured Playwright in `~/.ultrathink/config.json` → run click-through tests
Scoring (0-1 each):
- **Functionality**: Does it work as specified?
- **Design**: Is the UI/UX acceptable?
- **Craft**: Code quality, no console.log, proper error handling
- **Originality**: Does it deliver the unique value prop?
In `--builder` mode: Must score above `pass_threshold` (default 0.7) to proceed. If below → auto-enter IMPROVE phase.
Output: Update `evaluation` in forge-state.json.
## Phase 6: IMPROVE
Fix validation failures:
1. Read evaluation scores and failure details
2. Prioritize by impact (functionality > design > craft > originality)
3. Fix issues one at a time
4. Re-validate after fixes
Loop: IMPROVE → VALIDATE until passing or max 3 improvement cycles.
## Phase 7: SHIP
Prepare for deployment:
1. Generate/update README.md with setup instructions
2. Create PR if on a feature branch
3. List deployment steps for the chosen stack
4. Generate launch checklist
In `--guided` mode: Walk through each step. Explain what deployment means. Help with any unfamiliar steps.
Output: Set `stage: "complete"` in forge-state.json.
## Phase Gates (Builder mode only)
```
CLARIFY ──[spec complete]──→ FEASIBILITY ──[score ≥ 3]──→ PLAN
↑ │
└──[refine idea] │
▼
SHIP ←──[eval ≥ threshold]──── VALIDATE ←──── BUILD ←── [plan approved]
│ ↑
└──→ IMPROVE ─────────────┘
```
Gates are enforced by checking forge-state before entering each phase.
Cannot skip phases. Cannot enter BUILD without approved PLAN.
Cannot enter SHIP without passing VALIDATE.
## Decision Extraction (Builder mode)
When the user corrects forge's output during any phase:
- Extract the correction as a decision rule
- Save to `~/.ultrathink/decisions/projects/<hash>.json`
- Apply in future forge sessions for this project
Example:
- User: "No, use Drizzle not Prisma for this project"
- Extracted decision: `{ "rule": "Use Drizzle ORM, not Prisma", "scope": "<project-path>", "source": "user" }`
## Resuming
When starting a new session in a project that has forge-state:
1. Read `~/.ultrathink/forge/projects/<hash>.json`
2. Display current stage and progress
3. Ask: "Continue from [stage]?" or "Start fresh?"
The `forge-hydrate.sh` hook handles this automatically at session start.
## Forge Utility Functions
To compute project hash:
```bash
echo -n "$PWD" | shasum -a 256 | cut -c1-8
```
To read forge state:
```bash
HASH=$(echo -n "$PWD" | shasum -a 256 | cut -c1-8)
STATE="$HOME/.ultrathink/forge/projects/${HASH}.json"
```