Remove all hardcoded paths, make skill portable

- SKILL.md: replace ~/Private_Projects/... with $SGO_DIR pattern,
add auto-detection of install location
- Scripts: data dir now defaults to <project>/data/nemotron/
instead of ~/Data/nvidia/...
- AGENT.md: use relative paths, fix domain-specific language
- Synced installed skill at ~/.claude/skills/sgo/

AGENT.md

@@ -35,7 +35,7 @@ If `.env` doesn't exist, copy `.env.example` and ask the user to fill it in. Do
 
 ### Check data
 
-If `~/Data/nvidia/Nemotron-Personas-USA/dataset_info.json` exists, the persona dataset is ready. If not, run:
+If `data/nemotron/dataset_info.json` exists (relative to the project root), the persona dataset is ready. If not, run:
 
 ```bash
 uv run python scripts/setup_data.py
@@ -116,7 +116,7 @@ uv run python scripts/evaluate.py \
 
 **Present results to the user**:
 
-1. Overall score distribution (avg, swipe-right %, swipe-left %)
+1. Overall score distribution (avg, positive %, negative %)
 2. Breakdown by each stratification dimension
 3. Top 5 attractions (aggregated)
 4. Top 5 concerns (aggregated)

SKILL.md

@@ -8,53 +8,63 @@ argument-hint: "[entity-path or description]"
 
 You are executing the SGO pipeline. This optimizes an entity the user controls (θ) against a population of evaluators (x), using LLM-based scoring and counterfactual probes to estimate a semantic gradient.
 
-**Repo**: https://github.com/xuy/semantic-gradient-optimization
-**Local**: `~/Private_Projects/semantic-gradient-optimization/`
+**Repo**: https://github.com/xuy/sgo
 
-Read the full framework: `~/Private_Projects/semantic-gradient-optimization/README.md`
-Read the agent guide: `~/Private_Projects/semantic-gradient-optimization/AGENT.md`
+## First: find the install location
+
+The SGO repo could be installed at:
+- `~/.claude/skills/sgo/` (if installed as a skill)
+- Wherever the user cloned it
+
+Run `ls ~/.claude/skills/sgo/scripts/evaluate.py 2>/dev/null` to check. If not found, ask the user where they cloned SGO, or offer to clone it:
+
+```bash
+git clone https://github.com/xuy/sgo.git ~/.claude/skills/sgo
+```
+
+For the rest of this file, `$SGO_DIR` refers to the SGO install location. All paths are relative to it.
 
 ## Quick Reference
 
 ```
 Pipeline:  Build Entity → Build Cohort → Score f(θ,x) → Probe Counterfactuals → Act & Re-evaluate
-Scripts:   ~/Private_Projects/semantic-gradient-optimization/scripts/
-Templates: ~/Private_Projects/semantic-gradient-optimization/templates/
+Scripts:   $SGO_DIR/scripts/
+Templates: $SGO_DIR/templates/
+Data:      $SGO_DIR/data/
+Results:   $SGO_DIR/results/
 ```
 
 ---
 
-## Interactive Flow
+## Phase 0 — Setup
 
-Follow the AGENT.md guide, but here's the condensed version:
+1. **Dependencies**: `cd $SGO_DIR && uv sync` (or `pip install datasets huggingface_hub openai python-dotenv`)
+2. **API key**: Check `$SGO_DIR/.env` exists. If not, `cp $SGO_DIR/.env.example $SGO_DIR/.env` and ask the user to fill in their LLM API key. Do NOT read `.env` — just confirm it's configured.
+3. **Persona data**: Run `uv run python $SGO_DIR/scripts/setup_data.py`. This downloads NVIDIA Nemotron-Personas-USA (~2GB) to `$SGO_DIR/data/nemotron/` on first run and skips if already cached.
 
-### Phase 0 — Setup
-
-1. Check `~/Private_Projects/semantic-gradient-optimization/.env` exists with LLM API keys
-2. Check `~/Data/nvidia/Nemotron-Personas-USA/dataset_info.json` exists (if not, run `setup_data.py`)
-3. Ensure deps installed: `cd ~/Private_Projects/semantic-gradient-optimization && uv sync`
+---
 
-### Phase 1 — Entity (θ)
+## Phase 1 — Entity (θ)
 
 Ask the user: **"What are you optimizing?"**
 
 Options: product, resume, pitch, policy, profile, or custom.
 
-- If they have a document, save it to `entities/<name>.md`
-- If not, use a template from `templates/` and fill it in together
-- Confirm: "Here's what evaluators will see. Anything to add or remove?"
+- If they have a document, save it to `$SGO_DIR/entities/<name>.md`
+- If not, use a template from `$SGO_DIR/templates/` and fill it in together
+- Confirm: *"Here's what evaluators will see. Anything to add or remove?"*
 
-### Phase 2 — Cohort ({xᵢ})
+---
 
-Ask: **"Who is your target audience?"**
+## Phase 2 — Cohort ({xᵢ})
 
-Then ask: **"What dimensions matter for segmentation?"**
+Ask: **"Who is your target audience?"** and **"What dimensions matter for segmentation?"**
 
 Two paths:
 
 **A) Nemotron dataset** (preferred — census-grounded, 1M US personas):
 ```bash
-cd ~/Private_Projects/semantic-gradient-optimization
+cd $SGO_DIR
 
 # Filter
 uv run python scripts/persona_loader.py \
@@ -68,7 +78,7 @@ uv run python scripts/stratified_sampler.py \
   --output data/cohort.json
 ```
 
-**B) LLM-generated** (fallback — warn about mode collapse):
+**B) LLM-generated** (fallback — warn user about mode collapse risk):
 ```bash
 uv run python scripts/generate_cohort.py \
   --description "..." \
@@ -78,9 +88,12 @@ uv run python scripts/generate_cohort.py \
 
 Show the cohort distribution table. Confirm with user.
 
-### Phase 3 — Evaluate
+---
+
+## Phase 3 — Evaluate
 
 ```bash
+cd $SGO_DIR
 uv run python scripts/evaluate.py \
   --entity entities/<name>.md \
   --cohort data/cohort.json \
@@ -88,17 +101,17 @@ uv run python scripts/evaluate.py \
   --parallel 5
 ```
 
-Present: avg score, breakdown by segment, top attractions, top concerns, dealbreakers, most/least receptive.
+Present: avg score, breakdown by segment, top attractions, top concerns, dealbreakers, most/least receptive evaluators with quotes.
 
 Ask: **"Anything surprising? Want to dig into a segment?"**
 
-### Phase 4 — Counterfactual Probe
+---
+
+## Phase 4 — Counterfactual Probe
 
 Ask: **"What changes are you considering?"** and **"What won't you change?"**
 
-If unsure, propose changes based on top concerns from Phase 3.
-
-Save changes to `data/changes.json` (see `templates/changes.json` for format).
+If unsure, propose changes based on top concerns from Phase 3. Save to `data/changes.json` (see `templates/changes.json` for format).
 
 ```bash
 uv run python scripts/counterfactual.py \
@@ -111,10 +124,12 @@ Present the semantic gradient: ranked changes by avg Δ, % helped, % hurt, demog
 
 Ask: **"Which change do you want to make first?"**
 
-### Phase 5 — Iterate
+---
+
+## Phase 5 — Iterate
 
 1. User updates entity → save as `entities/<name>_v2.md`
-2. Re-evaluate: `--tag <new_tag>`
+2. Re-evaluate with same cohort: `--tag <new_tag>`
 3. Compare: `uv run python scripts/compare.py --runs <old> <new>`
 4. Show delta, new attractions, resolved concerns
 5. Ask: **"Another round, or are we good?"**
@@ -124,9 +139,9 @@ Ask: **"Which change do you want to make first?"**
 ## Key Principles
 
 - **Cohort is the control group** — keep it fixed across runs
-- **Census-grounded > LLM-generated** personas (see README § The Seeding Problem)
+- **Census-grounded > LLM-generated** personas (see README)
 - **Only probe actionable changes** — don't waste tokens on things they can't or won't change
-- **The gradient is semantic** — it's natural language, not numbers. Present the reasoning, not just the deltas.
+- **The gradient is semantic** — present the reasoning, not just the deltas
 - **Each run is tagged** — results are comparable longitudinally
 
 ## Arguments

scripts/persona_loader.py

@@ -21,7 +21,8 @@ import argparse
 from pathlib import Path
 from datasets import load_from_disk
 
-DEFAULT_DATA_DIR = Path.home() / "Data" / "nvidia" / "Nemotron-Personas-USA"
+PROJECT_ROOT = Path(__file__).resolve().parent.parent
+DEFAULT_DATA_DIR = PROJECT_ROOT / "data" / "nemotron"
 
 MBTI_TYPES = [
     "INTJ", "INTP", "ENTJ", "ENTP", "INFJ", "INFP", "ENFJ", "ENFP",

scripts/setup_data.py

@@ -1,7 +1,8 @@
 """
 Download and cache the Nemotron-Personas-USA dataset.
 
-Downloads 1M synthetic US personas (~2GB) from HuggingFace to ~/Data/nvidia/Nemotron-Personas-USA/.
+Downloads 1M synthetic US personas (~2GB) from HuggingFace.
+Default location: <project_root>/data/nemotron/
 Only runs once — subsequent calls detect the cached dataset and skip.
 
 Usage:
@@ -13,7 +14,8 @@ import argparse
 from pathlib import Path
 from datasets import load_dataset, load_from_disk
 
-DEFAULT_DATA_DIR = Path.home() / "Data" / "nvidia" / "Nemotron-Personas-USA"
+PROJECT_ROOT = Path(__file__).resolve().parent.parent
+DEFAULT_DATA_DIR = PROJECT_ROOT / "data" / "nemotron"
 
 
 def setup(data_dir: Path = DEFAULT_DATA_DIR):