version: "1.0.1" name: soul description: "Use when observing user patterns and synthesizing a persistent identity profile" effort: high

Iron Law

Observation is not judgment. Record patterns only.

Red Flags

• "They seem like a detail-oriented person" → STOP, because generic descriptions that could apply to any knowledge worker are rejected — every dimension needs 2+ evidence citations from actual observations.
• "They were direct last time and verbose this time, so they're inconsistent" → STOP, because contradictions become conditional rules ("direct in chat, verbose in reports"), not averages — context determines the pattern.
• "I noticed they mentioned their health situation" → STOP, because medical data, financial details, relationship status, and political/religious beliefs must never be recorded — log "sensitive signal omitted" without content.
• "5 observations should be enough to synthesize a profile" → STOP, because the minimum threshold is 10 sessions OR 30 observations — premature synthesis produces unreliable profiles.
• "The soul profile shifted significantly but the new version is clearly better" → STOP, because >30% drift in any dimension requires explicit user acknowledgment with "SIGNIFICANT DRIFT DETECTED" — never auto-apply large shifts.

Soul

Build and maintain a persistent user identity profile from behavioral signals observed across sessions. Every claim in SOUL.md must be evidence-backed. Contradictions become conditional rules, not averages.

When to Use

• User wants a persistent identity profile synthesized from session history
• User wants to understand their own patterns and behavioral tendencies
• Another skill needs user context that goes beyond project preferences
• User wants to track how their communication style or decision patterns shift over time

Subcommands

The observation store is written by the SessionStart/SubagentStart/Stop hooks and read through the soul_* MCP tools. Never append to it by hand. Two writers with two layouts is how this skill previously produced observations nothing ever synthesized.

Options

Mode Behavior

The mode lives in soul-active.json and the hooks read it — learning and hybrid are what turn automatic observation on.

• manual: nothing is recorded unless the user calls soul learn.
• learning: the hooks observe every session. Synthesis still requires soul propose.
• hybrid: learning, plus a synthesis prompt once proposal_due is set.

Workflow

init

1. If SOUL.md already exists, warn and require --force.
2. Load references/templates/{template}.md and write it as the SOUL.md stub.

The observation log needs no bootstrap. The hooks create it on first signal.

learn

For a signal the hooks did not catch. Call soul_record_observation with signal, category (correction | emotional | style), optional confidence, and raw_context trimmed to 200 characters. Report the count it returns.

show

1. soul_get_profile → the profile and its metadata. A null profile means synthesis has not run.
2. soul_get_observations for the citations behind each dimension.
3. Render the profile with its evidence inline, then the pool stats.

propose

1. soul_get_readiness. Below threshold (10 sessions or 30 observations), print the shortfall and stop.
2. soul_get_synthesis_context — recency-weighted observations, shipping entries, the current

profile, and a drift pre-check, already assembled. Do not re-read the files it summarizes.

1. Dispatch soul-keeper (Pikachu, opus) with that context. It applies

references/synthesis-algorithm.md.

1. Output the proposed SOUL.md with citations. Surface any dimension the drift pre-check flags above

30% with the words SIGNIFICANT DRIFT DETECTED.

1. Write nothing. propose ends in review.

apply

1. Require that propose ran in this session.
2. Ask: "Apply this soul update? This overwrites SOUL.md. [yes/no]"
3. On yes, write the proposed profile. On no, discard it and leave the profile untouched.

diff

Read the current profile via soul_get_profile, compare dimension by dimension against the proposed one, and flag semantic shifts above 30% per references/synthesis-algorithm.md.

retro

Call soul_retro with period and optional projects. It scans the git history, computes the metrics in references/retro-metrics.md, detects the trend against previous retros, and appends the shipping observation itself. Render its report; do not recompute what it returns and do not write the observation a second time.

reset

1. Require explicit confirmation.
2. Archive the current SOUL.md under archive/SOUL-{timestamp}.md.
3. Leave the observation log alone unless the user asks for it. The pool is the evidence; discarding

it silently destroys every citation the next profile would rest on.

Storage

Under CLAUDE_PLUGIN_DATA when set, otherwise <plugin>/.data. The hooks, the MCP handlers, and this skill must all use the same layout — they did not, and that is what made half the pipeline invisible to the other half.

Set CLAUDE_PLUGIN_DATA to keep this outside the plugin directory. Without it the store sits in the plugin install and does not survive a reinstall.

Subagents

soul-keeper: { model: opus, agent: pikachu, constraint: "every dimension needs 2+ evidence citations, contradictions become conditional rules, generic descriptions are rejected" }
analyst: { model: sonnet, agent: alakazam, constraint: "extract signals matching observation catalog, return structured JSON, no inferences beyond what the signal supports" }

Gotchas

Generic Soul Trap

The most common failure: producing dimensions that could describe any knowledge worker. soul-keeper must run the anti-generic filter on every dimension before output. If a dimension reads like a LinkedIn bio, it is rejected.

Token Budget

Observation logs can grow large. When calling soul-keeper, summarize observations older than 30 days to a single paragraph per session. Only the last 5 sessions are passed verbatim. Total input to soul-keeper must not exceed 500 tokens of observation data per session in the summary window.

Contradiction Handling

A user who is "direct in chat but verbose in reports" is NOT contradictory — those are conditional behaviors. Do not force a single characterization. The conditional rule format preserves predictive power; averaging destroys it.

Privacy

SOUL.md must never contain medical data, financial details, relationship status, or political/religious beliefs unless the user explicitly provides them as work-relevant context. If a sensitive signal is observed, note "sensitive signal omitted" in the observation log without recording the content.

Drift Approval

A >30% shift in any dimension is not automatically applied. It requires explicit user acknowledgment. Surface it visibly in propose output with the phrase "SIGNIFICANT DRIFT DETECTED" before applying.

Output

• SOUL.md at .data/soul/SOUL.md
• Template reference at references/templates/
• Observation signals reference at references/observation-signals.md
• Synthesis algorithm at references/synthesis-algorithm.md
• Retro metrics specification at references/retro-metrics.md