soul
FlaggedEmbody this digital identity. Read SOUL.md first, then STYLE.md, then examples/. Become the personβopinions, voice, worldview.
examples/steipete/README.md
View on GitHub# steipete soul file
Drop-in identity stack for **Peter Steinberger** (`@steipete`) β Austrian iOS-dev-turned-agentic-engineering-builder, founder of PSPDFKit (acquired), creator of OpenClaw π¦, and (as of Feb 2026) part of OpenAI's mission to bring agents to everyone.
Loaded as a system prompt, this stack lets any capable LLM (including weak ones like `gpt-4o-mini`) write in Peter's voice without fine-tuning or a GPU.
## Quickstart
See `QUICKSTART.md` for the full system-prompt template and a minimal one-shot version. The 30-second version:
```text
Load SOUL.md + STYLE.md + examples/good-outputs.md + examples/bad-outputs.md
as the system prompt. Then write as Peter β short sentences, tl;dr opener,
named tools (codex, gpt-5-codex on mid, OpenClaw, claude-code-mcp),
partisan stance with receipts, no banned vocab, no moralizing closer.
```
## What's in this folder
```
steipete-soul/
βββ README.md β you are here
βββ QUICKSTART.md β drop-in system prompt template (full + minimal)
βββ SOUL.md β identity: 12 worldview items, 5 modes, 8 tensions, pet peeves
βββ STYLE.md β voice: vocab tables, sentence rules, 10 rhetorical moves, platform register, 10-q grader
βββ examples/
β βββ good-outputs.md β 12 calibration samples + 12 verbatim quote anchors with source URLs
β βββ bad-outputs.md β 10 anti-patterns with β/π/β
structure + 7-q grader checklist
βββ tests/
β βββ prediction-test.md β 12 prompts on topics he hasn't taken on, predicted take + 0-2 rubric (pass β₯ 18/24)
β βββ weak-model-results.md β latest run: 39/48 = 3.25/4 PASS on gpt-4o-mini via OpenRouter (2026-04-28)
βββ scripts/
β βββ fetch-data.sh β reproducible pipeline: blog + GitHub + transcripts (idempotent)
β βββ pull-x.mjs β surf-ai MCP pull for @steipete + @openclaw posts/replies/profile
β βββ weak-model-test.mjs β runs the 12 prompts against any OpenAI/OpenRouter model
βββ data/
βββ influences.md β people, coinages, tools, concepts that shape his voice
βββ writing/ β 47 steipete.me HTML posts (2012-2014 iOS classics + 2025-2026 agent era)
βββ github/ β 16 README/VISION/AGENTS files (openclaw, poltergeist, peekaboo, codexbar, claude-code-mcp, mcporter, tachikoma, oracle, vibemeter, birdclaw, agent-rules, agent-scripts, interposekit, aspects)
βββ transcripts/ β 8 podcast/talk transcripts (Lex Fridman 491, Pragmatic Engineer "I ship code I don't read", Builders Unscripted, Maximize Claude Code, etc.)
βββ x/
βββ steipete.profile.json β @steipete profile (496k followers)
βββ steipete.posts-original.json β 100 timeline posts (filtered to 36 self-authored)
βββ steipete.replies.json β EMPTY ([]) β this pull failed silently and
β needs a re-run: SURF_API_KEY=... node scripts/pull-x.mjs
βββ openclaw.profile.json β @openclaw profile
βββ openclaw.posts-original.json β 100 posts (59 openclaw-authored)
βββ openclaw.replies.json β 100 replies
βββ curated-tweets.md β 129 themed entries (lobster way 63, models & tools 16, agents/harness 3, releases & receipts 12, industry takes 2, Apple ecosystem 2, other 31)
```
## Voice validation
The soul stack is validated by an automated weak-model test on `gpt-4o-mini`:
```bash
OPENROUTER_API_KEY=sk-or-... MODEL=openai/gpt-4o-mini node scripts/weak-model-test.mjs
```
Each of 12 prompts (mirroring `tests/prediction-test.md`) is scored on:
- **Voice (0-2)** β signal density (named tools, lobster register, tl;dr opener, partisan beat) + steipete-distinctive vocabulary markers
- **Stance (0-2)** β not-hedging + has-opinion + has-anchor (number / version / named tool)
- **Anti-pattern penalty (up to -2)** β banned vocab (leverage, unlock, revolutionize, "thrilled," "both have strengths," etc.)
Latest result: **39/48 = 3.25/4 average, PASS**. See `tests/weak-model-results.md` for per-prompt breakdown.
## Why this voice is hard to fake
Peter is one of the most online builders alive in 2026 and his voice has a few hard-to-imitate features:
1. **Receipts everywhere.** Every claim is anchored to a tool name, a number, or a date. "vibetunnel b12 is out, 156k LOC, 3,144 commits, 38 contributors" β not "we shipped a major update."
2. **Partisan, with named comparisons.** "I switched from Claude Code to codex in August. ~230k context vs 156k seals it." Hedging both-sides is fatal.
3. **The lobster register** (π¦ / "the lobster way" / "the claw is the law" / "EXFOLIATE!") is real and specific to OpenClaw / single-user / local-first contexts. Forced lobster outside that context is a tell.
4. **Honest about bugs and burnout.** Black Eye Club, slot machine, the year he doesn't remember. Sanitized output sounds wrong.
5. **Tool casing matters.** Lowercase `codex`, `claude code`, `gpt-5-codex`, `vercel`. Capitalize `Cursor`, `OpenClaw`, `OpenAI`, `CodexBar`. Casing tells you whether the writer actually uses the tool or is reading press releases.
The good/bad-outputs files codify these signals so the model has concrete calibration in addition to abstract rules.
## Reproducing the data pipeline
```bash
# Blog posts (47 files), GitHub READMEs (16 files), YouTube transcripts (8 files)
bash scripts/fetch-data.sh
# X corpus (requires SURF_API_KEY for surf-ai MCP)
SURF_API_KEY=... HANDLES="steipete,openclaw" node scripts/pull-x.mjs
# Voice-holding test (requires OpenAI or OpenRouter key)
OPENROUTER_API_KEY=sk-or-... MODEL=openai/gpt-4o-mini node scripts/weak-model-test.mjs
```
All scripts are idempotent β re-running skips files already cached.
## Influences (cited in SOUL.md)
People, coinages, tools, and concepts that shape his voice β see `data/influences.md` for the full list. Highlights:
- **Coinages**: tl;dr opener, the lobster way, blast radius, slot machine, Black Eye Club, "you can just do things," "three angles," Claudoholic, "full-breadth developer."
- **Daily stack**: codex (`gpt-5-codex on mid`), OpenClaw, CodexBar, Peekaboo, claude-code-mcp, mcporter, Poltergeist, VibeMeter, VibeTunnel, Tachikoma.
- **People**: Sam Altman / OpenAI culture (joined Feb 2026), Geoffrey Huntley (parallel-agent loops), Andrej Karpathy (vibe coding, with respectful distance), Sebastian Bauer (clawsweeper / lobster register), the iOS-dev old guard.
- **What's NOT an influence**: VC announcement style, AI-doomer register, Karpathy em-dash flourish, prompt-engineering listicles, both-sides "it depends" hedging.
## License
Public corpus is reproduced for educational / soul-file purposes under fair use. Original writings remain copyright Peter Steinberger and the respective platforms. See `aeonfun/soul.md` repo for the umbrella project license.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |