aeonfun · aaronjmars · Jul 21, 2026 · Jul 21, 2026 · Jul 21, 2026 · Jul 21, 2026
diff --git a/.github/CONTRIBUTING.md b/.github/CONTRIBUTING.md
@@ -12,9 +12,11 @@ reading it could predict its subject's take on a new topic.**
   or [`STYLE.template.md`](../STYLE.template.md), improve a `_GUIDE.md`, fix a
   factual error or broken link.
 
-Build your soul first with the [soul-builder skill](../SKILL.md) (`/soul-builder`)
-or by hand from the templates. See [`examples/_GUIDE.md`](../examples/_GUIDE.md)
-for what goes in the calibration files.
+Build your soul first with the [soul-builder skill](../BUILD.md) (`/soul-builder`)
+or by hand from the templates — [`BUILD.md`](../BUILD.md) is the full spec for
+what a soul folder must contain. See [`examples/_GUIDE.md`](../examples/_GUIDE.md)
+for what goes in the calibration files, and [`data/_GUIDE.md`](../data/_GUIDE.md)
+for what belongs in `data/`.
 
 ## Before you start
 

diff --git a/.github/README.md b/.github/README.md
@@ -110,14 +110,14 @@ Soul files are plain markdown — if an agent can read files, it can embody you.
 | Framework | Language | Stars |
 |-----------|----------|-------|
 | [Aeon](https://github.com/aeonfun/aeon) | YAML/Markdown | — |
-| [OpenClaw](https://github.com/openclaw/openclaw) | TypeScript | 322k |
-| [Nanobot](https://github.com/HKUDS/nanobot) | Python | 34.6k |
-| [ZeroClaw](https://github.com/zeroclaw-labs/zeroclaw) | Rust | 27.8k |
-| [PicoClaw](https://github.com/sipeed/picoclaw) | Go | 25.3k |
-| [NanoClaw](https://github.com/qwibitai/nanoclaw) | TypeScript | 24k |
-| [OpenFang](https://github.com/RightNow-AI/openfang) | Rust | 14.9k |
-| [IronClaw](https://github.com/nearai/ironclaw) | Rust | 10.4k |
-| [Hermes Agent](https://github.com/NousResearch/hermes-agent) | Python | 8.7k |
+| [OpenClaw](https://github.com/openclaw/openclaw) | TypeScript | 383k |
+| [Hermes Agent](https://github.com/NousResearch/hermes-agent) | Python | 218k |
+| [Nanobot](https://github.com/HKUDS/nanobot) | Python | 46k |
+| [ZeroClaw](https://github.com/zeroclaw-labs/zeroclaw) | Rust | 32.3k |
+| [NanoClaw](https://github.com/qwibitai/nanoclaw) | TypeScript | 30.3k |
+| [PicoClaw](https://github.com/sipeed/picoclaw) | Go | 29.7k |
+| [OpenFang](https://github.com/RightNow-AI/openfang) | Rust | 18k |
+| [IronClaw](https://github.com/nearai/ironclaw) | Rust | 12.5k |
 | Claude Code · OpenCode · Codex · Goose | various | — |
 
 Also works with any model via system prompt — see [Using With Other Tools](#using-with-other-tools).

diff --git a/.github/SECURITY.md b/.github/SECURITY.md
@@ -24,7 +24,8 @@ that part isn't a secret and we'd rather act on it fast.
 
 Please include what you can:
 
-- The file or component affected — `SKILL.md` (the builder), a `scripts/*.mjs`
+- The file or component affected — `BUILD.md` (the builder), `SKILL.md` (the
+  embodier), a `scripts/*.mjs`
   helper, a template, or a specific `examples/<name>/` soul.
 - A minimal reproduction (for injection: the source text that triggers it).
 - The impact you can demonstrate.

diff --git a/.gitignore b/.gitignore
@@ -1 +1,3 @@
 .DS_Store
+__pycache__/
+*.pyc
diff --git a/.markdownlint-cli2.jsonc b/.markdownlint-cli2.jsonc
@@ -14,6 +14,5 @@
     "MD053": true, // every link/image reference definition must be used
     "MD054": true  // disallowed malformed link/image styles
   },
-  "globs": ["**/*.md"],
-  "ignores": ["node_modules"]
+  "globs": ["**/*.md"]
 }
diff --git a/BUILD.md b/BUILD.md
@@ -189,6 +189,7 @@ When done, you should have created:
 - `STYLE.md` — Their voice
 - `MEMORY.md` — Empty memory log
 - `examples/good-outputs.md` — Calibration examples
+- `examples/bad-outputs.md` — Negative calibration examples
 - Optionally: `data/influences.md` if built from interview
 
 The user can then invoke `/soul` to embody their digital identity.
diff --git a/examples/garry-tan/README.md b/examples/garry-tan/README.md
@@ -2,20 +2,17 @@
 
 A `soul.md` identity distillation for Garry Tan — co-founder of Posterous, founding partner of Initialized Capital, president and CEO of Y Combinator.
 
-This folder is structured to be dropped into `aeonfun/soul.md` at `examples/garry-tan/`.
 
 ## What's in here
 
 ```
-garrytan/
+garry-tan/
 ├── SOUL.md              ← Identity, worldview, opinions
 ├── STYLE.md             ← Voice mechanics
 ├── MEMORY.md            ← Empty session log
 ├── data/
-│   ├── influences.md
-│   ├── x/               ← Twitter archive target (see scripts/)
-│   ├── writing/         ← YC blog + personal essays
-│   └── yt/              ← YouTube transcripts
+│   └── influences.md
+│                        ← x/, writing/, yt/ are created by scripts/ on first run
 ├── examples/
 │   ├── good-outputs.md  ← 12 calibration samples
 │   └── bad-outputs.md   ← 10 anti-patterns
@@ -24,8 +21,12 @@ garrytan/
 │   ├── fetch_writing.py ← YC blog + personal posts → data/writing/
 │   └── fetch_yt.py      ← YouTube transcripts → data/yt/*.md
 └── tests/
-    ├── prediction_test.md   ← 10 unseen prompts + expected takes
-    └── weak_model_test.md   ← gpt-4o-mini run + verdict
+    ├── prediction_test.md      ← 10 unseen prompts + expected takes
+    ├── platform_tests.md       ← 3 platform-shape prompts
+    ├── weak_model_test.md      ← protocol + verdict
+    ├── run_weak_model.py       ← the harness
+    ├── results_gpt-4o-mini.md  ← full transcript
+    └── scores_gpt-4o-mini.md   ← per-prompt grades (38.5/48 PASS)
 ```
 
 ## How to rebuild the corpus
@@ -46,7 +47,7 @@ python scripts/fetch_yt.py
 
 ```bash
 # From a project root, copy into a /soul directory and invoke
-claude --skill soul "write a tweet reacting to the latest OpenAI announcement"
+/soul write a tweet reacting to the latest OpenAI announcement
 ```
 
 The SKILL.md from the parent `soul.md` repo handles the embodiment logic.

diff --git a/examples/garry-tan/scripts/fetch_x.py b/examples/garry-tan/scripts/fetch_x.py
@@ -27,22 +27,16 @@
 import json
 import os
 import pathlib
-import re
 import sys
 import time
 import urllib.parse
 import urllib.request
 
-HANDLE = "garrytan"
 USER_ID = "19301828"  # @garrytan, stable numeric id
 OUT = pathlib.Path(__file__).parent.parent / "data" / "x"
 OUT.mkdir(parents=True, exist_ok=True)
 
 
-def _slug(s: str, n: int = 40) -> str:
-    return re.sub(r"[^a-z0-9]+", "-", s.lower()).strip("-")[:n]
-
-
 def _write(tweet_id: str, date: str, text: str, source: str, reply_to: str | None) -> None:
     safe_date = date.split("T", 1)[0]
     path = OUT / f"{safe_date}_{tweet_id}.md"

diff --git a/examples/garry-tan/tests/run_weak_model.py b/examples/garry-tan/tests/run_weak_model.py
@@ -2,9 +2,11 @@
 """
 Run the soul file against a weak model via OpenRouter.
 
-Reads OPENROUTER_API_KEY from env (or from ~/.hermes/.env as a fallback).
+Reads OPENROUTER_API_KEY from env.
 Writes a markdown transcript with one section per prompt.
 
+Requires `certifi` (pip install certifi).
+
 Usage:
     python tests/run_weak_model.py \
         --model openai/gpt-4o-mini \
@@ -21,7 +23,6 @@
 import pathlib
 import sys
 import ssl
-import urllib.error
 import urllib.request
 
 import certifi
@@ -89,11 +90,6 @@ def load_api_key() -> str:
     key = os.environ.get("OPENROUTER_API_KEY")
     if key:
         return key
-    fallback = pathlib.Path.home() / ".hermes" / ".env"
-    if fallback.exists():
-        for line in fallback.read_text().splitlines():
-            if line.startswith("OPENROUTER_API_KEY="):
-                return line.split("=", 1)[1].strip()
     sys.exit("OPENROUTER_API_KEY not found")
 
 
@@ -119,11 +115,8 @@ def call_model(api_key: str, model: str, system: str, user: str) -> str:
             "X-Title": "soul-garrytan weak-model test",
         },
     )
-    try:
-        with urllib.request.urlopen(req, timeout=90, context=_SSL_CTX) as resp:
-            data = json.loads(resp.read())
-    except urllib.error.HTTPError as e:
-        return f"[ERROR {e.code}: {e.read().decode(errors='replace')}]"
+    with urllib.request.urlopen(req, timeout=90, context=_SSL_CTX) as resp:
+        data = json.loads(resp.read())
     return data["choices"][0]["message"]["content"]
 
 

diff --git a/examples/garry-tan/tests/weak_model_test.md b/examples/garry-tan/tests/weak_model_test.md
@@ -48,20 +48,20 @@ For each response, grade:
 - **Stance match**: 0–2 (opinion aligned with SOUL.md)
 - **Anti-pattern hits**: subtract 1 per `bad-outputs.md` tell that appears (e.g., "that's a great question," emoji spam, hedging stack).
 
-Formula: `score = (voice + stance) - anti_pattern_hits`. Max 4 per prompt. Pass threshold: **average ≥ 3.0 across 13 prompts**.
+Formula: `score = (voice + stance) - anti_pattern_hits`. Max 4 per prompt. Pass threshold: **average ≥ 3.0 across 12 prompts**.
 
 ### 5. Run script
 
 ```bash
-# OPENAI_API_KEY required
-python tests/run_weak_model.py \
-  --model gpt-4o-mini \
-  --prompts tests/prediction_test.md,tests/platform_tests.md \
-  --soul /tmp/garry_prompt.md \
+OPENROUTER_API_KEY=... python3 tests/run_weak_model.py \
+  --model openai/gpt-4o-mini \
+  --soul SOUL.md --style STYLE.md \
+  --good examples/good-outputs.md --bad examples/bad-outputs.md \
   --out tests/results_gpt-4o-mini.md
 ```
 
-`run_weak_model.py` is a 30-line stub below.
+The prompts are defined in `run_weak_model.py` itself; `--style`, `--good`, and
+`--bad` are required.
 
 ---
 
@@ -107,55 +107,20 @@ The three things that usually break weak-model adherence:
 
 ---
 
-## Run script (reference implementation)
-
-```python
-# tests/run_weak_model.py
-import argparse, json, os, re, sys
-from openai import OpenAI
-
-ap = argparse.ArgumentParser()
-ap.add_argument("--model", default="gpt-4o-mini")
-ap.add_argument("--soul", required=True)
-ap.add_argument("--prompts", required=True)  # comma-sep .md files
-ap.add_argument("--out", required=True)
-args = ap.parse_args()
-
-client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])
-soul = open(args.soul).read()
-prompts = []
-for f in args.prompts.split(","):
-    prompts += re.findall(r"^## \d+\..*?\n(.*?)(?=\n---|\Z)", open(f).read(), re.S|re.M)
-
-system = (
-    "You are Garry Tan. Embody the identity below. Never break character. "
-    "Match the voice in good-outputs.md, avoid bad-outputs.md.\n\n"
-    f"<IDENTITY>\n{soul}\n</IDENTITY>"
-)
-
-with open(args.out, "w") as out:
-    for i, p in enumerate(prompts, 1):
-        resp = client.chat.completions.create(
-            model=args.model,
-            messages=[
-                {"role": "system", "content": system},
-                {"role": "user", "content": p.strip()},
-            ],
-            temperature=0.7,
-        )
-        out.write(f"## Prompt {i}\n\n{p.strip()}\n\n**Response:**\n\n{resp.choices[0].message.content}\n\n---\n\n")
-print(f"wrote → {args.out}", file=sys.stderr)
-```
+## Run script
 
-Run it, grade it, commit the result. A passing run is part of the deliverable.
+The harness is [`run_weak_model.py`](run_weak_model.py) — argparse CLI, OpenRouter
+via `urllib`, prompts defined in the script. Run it, grade it, commit the result.
 
 ---
 
 ## Status
 
 - ✅ Protocol defined
 - ✅ Prompts authored (prediction_test.md + platform_tests.md)
-- ✅ Run script reference implementation (above — saveable as-is)
-- ⏳ Live run pending: requires `OPENAI_API_KEY`. Grading is manual until we add an LLM-judge step.
+- ✅ Run script shipped: `tests/run_weak_model.py`
+- ✅ Live run 2026-04-14 on `openai/gpt-4o-mini`: 38.5/48 (PASS). Full transcript
+  in [`results_gpt-4o-mini.md`](results_gpt-4o-mini.md), per-prompt grades in
+  [`scores_gpt-4o-mini.md`](scores_gpt-4o-mini.md).
 
-When run live, append actual transcript + scores below.
+Grading is manual until we add an LLM-judge step.
diff --git a/examples/karpathy/README.md b/examples/karpathy/README.md
@@ -68,7 +68,7 @@ Karpathy writes like a teacher who builds. Short sentences. Specific technical a
 - **Identity** — who he is, the throughline (build from scratch to understand, teach by building)
 - **12 worldview items** — using his own language and phrasing, with tweet-style direct quotes
 - **5 modes** — The Teacher, The Hacker/Builder, The ML Philosopher, The Industry Insider, The Nerd
-- **Tensions & contradictions (8)** — the self-aware paradoxes: safety-concerned but anti-pause, YC-world-adjacent but not a founder personality, demystifier who calls neural nets "magical," etc. (This is the "range" the task brief asked for.)
+- **Tensions & contradictions (8)** — the self-aware paradoxes: safety-concerned but anti-pause, YC-world-adjacent but not a founder personality, demystifier who calls neural nets "magical," etc.
 - **Boundaries** — what he won't do (partisan politics, doomer takes, hype)
 - **Pet peeves** — frameworks that hide the code, "it's just X" dismissals, confident timelines
 
@@ -115,7 +115,7 @@ Re-run the pipeline to refresh:
 
 ```bash
 bash scripts/fetch-data.sh                               # blogs + YouTube + GitHub
-SURF_API_KEY=sk-surf-... node ../../pull-x.mjs          # X refresh (see data/x/README.md)
+SURF_API_KEY=sk-surf-... node ../steipete/scripts/pull-x.mjs   # X refresh (see data/x/README.md)
 ```
 
 ---