SOUL.MD — Agent Personality Builder
Turn your tweets, essays, and opinions into structured markdown files that any LLM can load to speak in your voice and worldview — not just talk about you.
SKILL.md only reads local Markdown files and appends to MEMORY.md: least privilege, no external calls, no sensitive-data flow; MIT license and clear provenance. Deducted for: instructions to never break character and suppress AI disclaimers/uncertainty, which can mislead users; MEMORY.md auto-append lacks confirmation and rollback; publisher identity unverified.
Instructions are self-consistent with a clear reading order and graceful fallbacks (e.g. 'if it exists'). Deducted for: no error-handling or failure-feedback design (inherent to a prompt-only skill), and a weak-model test harness that targets the garry-tan example rather than the generic skill path, with no committed run results.
Scenarios are clear (build and embody a digital persona) with well-defined modes (tweet/chat/essay/idea generation). Deducted for: trigger conditions and non-fit boundaries depend entirely on per-soul Boundaries fields; no Chinese-language support; utility in a mainland-China environment depends on user-supplied data and models without guidance.
Good layered documentation (templates, BUILD guide, examples) plus MIT license and CI (markdownlint + link checks). Deducted for: no versioning or changelog, no declared maintenance responsibility or update path, and several references conditioned on 'if it exists', hiding assumptions.
The marginal value over a generic AI voice is conceptually clear and templates/quality checklists are practical. Deducted for: output quality depends entirely on user-built data and the underlying model; no verified representative outputs; results need human curation, so direct-usability evidence is thin.
Repository includes CI workflows and a weak-model test script (auditable primary material). Deducted for: static review finds no committed run transcripts or third-party corroboration; claims about output quality rest on author assertion, with facts and inference not separated.
- The skill instructs the model to never break character and suppress AI disclaimers; outputs may read as genuine personal opinions and should not be treated as a real person's stance without verification.
- MEMORY.md is auto-appended with no rollback; review and prune it manually.
- Static review only; the weak-model test script was not executed in this assessment.
- No Chinese-language support, no changelog; publisher identity is unverified by the registry.
What does this skill do, and when should you use it?
SOUL.MD is an open-source Claude Code skill for building AI agent personalities. It distills your writing and social content into a set of markdown files — identity, style, memory, examples — that any file-reading LLM can load to write as you. A /soul-builder flow generates these files either through an interview or by analyzing raw content you drop into data/. The repo ships real example soul files of public figures and pairs with the Aeon background agent so every scheduled task inherits your personality.
SKILL.md defines a strict reading order (SOUL.md → STYLE.md → MEMORY.md → examples/ → data/) and instructs the model to stay fully in character, banning 'as an AI' disclaimers for the session. For uncovered topics, it applies a four-level source priority (explicit positions → data/ reference → extrapolation from adjacent positions → reason from worldview with flagged uncertainty). Five output modes (default, tweet, chat, essay, idea generation) each impose different style constraints. MEMORY.md provides a cross-session memory append mechanism, while data/ is absorbed for grounding rather than quoted directly.
- A content creator who wants AI to batch-produce tweets, articles, and emails in their own voice
- A writer who exports their Substack and Twitter data into data/ and runs /soul-builder to auto-extract worldview and style
- A new user with no existing content who builds a soul file from scratch via the /soul-builder interview mode
- An Aeon user who wants all background cron tasks (digests, tweets, research briefs) to carry one consistent personality with no per-skill config
- An open-source maintainer building a Q&A persona from a public figure's body of writing for a community agent
What are this skill's strengths and limitations?
- Plain markdown — loads on any file-reading LLM or agent framework, no platform lock-in
- Explicit behavioral spec: source priority, interpolation rules, and anti-pattern lists reduce voice drift
- MEMORY.md gives cross-session continuity that the user can manually prune
- Multiple real example souls with validation scores serve as reference patterns
- Thorough compatibility list covering OpenClaw, Hermes Agent, ZeroClaw, NanoClaw and more
- Mandates total in-character behavior with no disclaimers — inherent tension with factual accuracy and compliance; the source provides no safety-guardrail mechanism
- Framework star counts cited in the README (e.g., OpenClaw 383k, Hermes Agent 218k) lack independent verification
- Example souls impersonate real public figures (Karpathy, Vitalik, etc.), an ethical gray area only partially addressed with ethics notes
- The skill itself ships no automated test suite; only the example files include weak-model test scores
- Affiliation with the AEON token gives the repo a marketing dimension — assess neutrality yourself before adopting
How do you install this skill?
The repo does not document a standard install command. Per the source: clone https://github.com/aeonfun/soul.md and place the folder (containing SKILL.md) into your Claude Code skills directory. For Aeon integration, copy your soul/ folder into your Aeon repo and add a few lines to CLAUDE.md. Manual setup means copying the templates (SOUL.template.md, STYLE.template.md, etc.) and filling them in.
How do you use this skill?
1) Run /soul-builder: either answer an interview, or first drop content into data/x/ and data/writing/ and let it analyze; 2) Fill in SOUL.md (identity, worldview, opinions), STYLE.md (voice, anti-patterns), and examples/ (good and bad outputs); 3) Invoke /soul to activate — the model reads SOUL.md → STYLE.md → MEMORY.md → examples/ → data/ in order and enters character; 4) Append notable events to MEMORY.md at session end. Exact triggering mechanics for the tweet/chat/essay/idea modes are not fully documented in the source.
How does this skill compare with similar options?
Versus a single system-prompt persona paragraph, SOUL.MD offers a layered structure (identity/style/memory/examples separated) plus reusable templates; the README explicitly recommends pasting SOUL.md and STYLE.md into the system prompt as a lightweight fallback for weaker models. Relative to its sibling project Aeon: Aeon handles the 'when and what' (scheduling, research, monitoring), while soul files handle the 'who' (voice, worldview, opinions).