me-distill

name

me-distill

description

Distill stack-agnostic personal profile (~/Documents/me-profile/) from raw personal resource (Gmail md dumps, Calendar export, Drive notes, etc.) into a dual-layer always-on inject system - about_me.md (Layer A, 60-line cap, Tier 1, always loaded) plus on-demand facet detail files split across three axes (life Tier 1, work Tier 1, sensitive Tier 2). Use when user invokes "蒸餾我這個人" / "me distill" / "建 me profile" / "更新 me profile" / "重蒸 X facet" / "refresh my profile" / "re-extract <axis>" / "incremental run" / "daily delta" / "add new source" / "把 me 接進 OpenClaw" / "persona distillation" / "個人蒸餾" / similar Chinese-English triggers, or asks to refresh a stale profile, add a new facet, re-extract a single axis, append a new source (Calendar, work Gmail, Drive notes), trigger a daily delta run, or wire about_me into Claude Code memory / OpenClaw workspace / future life assistant. Supports four modes - from-scratch (initial extraction), incremental (merge delta into existing baseline without re-extracting unchanged facets), refresh (force re-extract specific facet), backfill (one-shot build v2.0 manifest from existing R1 state). Runs five phases - Phase A unsupervised discovery to inventory facets, Phase B user-confirmed facet map with axis classification, Phase C parallel per-axis extraction (life/work/sensitive subagents), Phase D meta-distill of about_me.md, Phase E confirm-and-optional-wire. Enforces evidence-tier tagging (verbatim/artifact/impression), kill-list filter, swap test, contradiction preservation, read-only on raw source, and human review gate between every phase. Incremental mode adds append-only on facet files (existing 7-section content immutable), Phase D quality gate auto-rollback, and snapshot-before-synthesize discipline via scripts/manifest_tool.py + scripts/quality_check.py.

me-distill

Distill a stack-agnostic personal profile from raw resource into a dual-layer inject system. Replaces ad-hoc note-taking with a disciplined five-phase pipeline.

Trigger

Invoke this skill for any of the following user signals:

• "蒸餾我這個人" / "me distill" / "建 me profile" / "個人蒸餾"
• "更新 me profile" / "重蒸 X" / "refresh my profile" / "incremental run" / "daily delta"
• "add new source" / "merge in <Calendar / Drive / 工作 Gmail>"
• "把 me 接進 OpenClaw" / "wire about_me into CC" / "接到 life assistant"
• "persona distillation" / "personal context extract"
• Any request to (re-)extract facets from raw personal sources for use as AI inject context

Default mode is conversational - no CLI flags. Expert mode (--discover / --wire openclaw / --axis sensitive / --incremental / --from-scratch / --force-synthesize) only when user explicitly types flags.

Modes

The skill runs in one of four modes. Mode selection is automatic based on the state of ~/Documents/me-profile/:

Mode	When	Behaviour
from-scratch	~/Documents/me-profile/ is empty / missing OR user passes --from-scratch	Run full Phase A-E pipeline against all source. Overwrites everything. Use for R1 (initial distill) or full rebuild.
incremental	_manifest.json (v2.0) exists AND new source added OR existing source cursor advanced OR user invokes --incremental	Merge delta. Facet files: append-only (existing 7 sections immutable, new evidence goes to ## Evidence Log section). about_me: re-synthesize only if delta_evidence ≥ 20 OR conflicts_pending > 0 OR age > 7 days. See references/incremental.md.
refresh	User explicitly asks "re-extract " or "重蒸 "	Forces single facet through Phase C even if needs_reextract == false. Other facets untouched.
backfill	~/Documents/me-profile/ has R1 facet files but NO _manifest.json (or older version). Run python3 scripts/r1_backfill.py once.	One-shot helper, builds v2.0 manifest reflecting current state. No re-extraction, no source modification. Prerequisite before first incremental run on a legacy R1 profile.

Output structure

~/.claude/skills/me-distill/
├── SKILL.md
├── scripts/
│   ├── manifest_tool.py    # Deterministic manifest helper (init/read/write/fingerprint/snapshot/rollback/enqueue/dequeue/mark_done/add_conflict/dedup_check)
│   ├── quality_check.py    # 6-dimension PASS/FAIL gate for about_me.md (used by Phase D auto-rollback)
│   └── r1_backfill.py      # One-shot helper: builds _manifest.json v2.0 from existing R1 me-profile/ state
└── references/             # Per-phase + incremental subagent prompts (see ## Detailed reference files)

~/Documents/me-profile/
├── about_me.md             # Layer A, 60-line hard cap, always-on inject, Tier 1
├── _manifest.json          # v2.0 state: last_run, source_paths (with cursor/fingerprint), facet_list, axis_classification, queue, snapshots, conflicts_pending, wired_targets — see PLAN-incremental.md § 2.2 for full schema
├── _snapshots/             # about_me.md history (kept per snapshot_retention cap, default 10)
├── discover-out/
│   ├── facet-inventory.md           # Phase A from-scratch output (R1 baseline); incremental runs APPEND ## Incremental run sections
│   ├── facet-map-approved.md        # Phase B confirmations; incremental runs APPEND ## R<N> Addendum sections
│   └── inventory-addendum-R<N>.md   # Per-round delta inventory (incremental only)
├── life/                   # Layer B detail, Tier 1, ~10-12 facets — 7 sections IMMUTABLE in incremental, new evidence in ## Evidence Log section
├── work/                   # Layer B detail, Tier 1, ~10-12 facets — same append-only contract
└── sensitive/              # Layer B detail, Tier 2 (life assistant only, NOT work assistant), ~5-8 facets — same append-only + sensitive masking

See PLAN-incremental.md § 2.2 for the full _manifest.json v2.0 schema.

Dual-layer architecture

• Layer A about_me.md = always-on inject, hard cap 60 lines, 7 sections, high-level "who they are". Read by every CC / OpenClaw / assistant session. Optimised for density per line, not completeness.
• Layer B {life,work,sensitive}/*.md = on-demand detail. Each facet 50-100 lines. AI only reads when conversation actually surfaces that topic. Source of truth for any drill-down.

Three-axis routing

• life (Tier 1) - Consumer / entertainment / social / food / gaming / travel facets. Safe for any assistant.
• work (Tier 1) - Job / career / dev tool / recruitment / company / professional facets. Safe for any assistant.
• sensitive (Tier 2) - Finance (banks, cards), mental health (counseling), age-restricted preferences (adult ACG platforms), security/identity. Only loaded by life-assistant context. Excluded from work-assistant context.

Ambiguous facets (e.g. design tool subscription = work resource or life hobby?) are resolved with the user in Phase B Q3.

Phase A - Unsupervised Discovery

Goal: Without any user-supplied hypothesis, enumerate every distinct facet present in the raw resource, with evidence count and Tier classification.

Input: A raw resource directory (default ~/.openclaw/workspace/memory/personal/ for Gmail md dumps; user may specify other path).

Output: _inventory.md at ~/Documents/me-profile/ (or a versioned subdir like v3-discover-sonnet/_inventory.md if iterating).

Subagent dispatch (parallel allowed for cross-validation): See references/phase-a-discover.md for the full subagent prompt covering three-stage discovery (explore → quantify → write inventory), read-only enforcement, and inventory output format.

Incremental mode: When _manifest.json exists, Phase A only scans the delta range past each source's cursor. New facets are surfaced; existing facets get evidence_count increments + needs_reextract = true. Output appends to discover-out/facet-inventory.md (not overwrite) + writes discover-out/inventory-addendum-R<N>.md. See references/incremental.md for the full incremental contract.

Quality bar:

• Must reach evidence-count cardinality (not just topic list)
• Must include facets with evidence < 10 if distinct (don't suppress long-tail signal)
• Must surface emergent cross-facet pattern (Section "跨 facet pattern / Emergent insight")
• Must declare limits / blind spots (Section "盲點 / Limit")
• Read-only on raw resource - never modify, move, delete source files

After Phase A completes, present _inventory.md to user and gate on review.

Phase B - User-Confirmed Facet Map

Goal: Convert the discovered inventory into a per-facet axis assignment that the user has actually confirmed.

Questions (ask in sequence, do not batch):

• Q1. Which facets to distill? (default: all; user may exclude noise like "EDM 促銷 facet")
• Q2. Confirm axis for each facet: life / work / sensitive
• Q3. Ambiguous cases - present each one with reasoning, let user decide:
  • Design tool subscriptions (a design-stock / code-sandbox SaaS): work tool or life hobby?
  • Job-search resources (a professional network / a job board): work (career progression) or sensitive (income / negotiation context)?
  • Gaming purchases: life (entertainment) or sensitive (spending pattern)?
• Q4. Confirm Tier 2 (sensitive) handling rules - see references/sensitive-handling.md
• Q5. Naming convention for facet files (default: short kebab-case = sender domain or brand)
• Q6. Output destination - confirm ~/Documents/me-profile/ or override

Write resolved axis map into _manifest.json before Phase C.

Incremental mode: Only NEW facets + ambiguous re-classify candidates are surfaced (R1-confirmed assignments are not re-asked). Q4-Q6 are skipped. If new_facet_count == 0 AND ambiguous_count == 0, auto-take subagent recommendation without PENDING_USER_ACTION. Confirmation appends to discover-out/facet-map-approved.md under ## R<N> Addendum section.

Phase C - Per-Axis Parallel Extraction

Goal: For each axis, dispatch a subagent that reads only the relevant facets' raw evidence and writes one .md per facet under {axis}/.

Dispatch: Three subagents in parallel - one per axis. Each owns its own facet list from _manifest.json.

Subagent prompts:

• references/phase-c-life.md - life axis specific discipline
• references/phase-c-work.md - work axis specific discipline
• references/phase-c-sensitive.md - sensitive axis specific discipline + references/sensitive-handling.md

Shared discipline (all three subagents):

• Apply 7-section facet template (TL;DR / 核心 fact / Pattern / 時間性 / 矛盾 / 數據點 / 不足) - see references/about-me-template.md
• Tag every claim with evidence-tier: verbatim / artifact / impression
• Apply kill list (see references/about-me-template.md)
• Apply swap test (see references/about-me-template.md)
• Preserve contradiction (do not paper over time-based / domain-based / essential conflicts)
• Read-only on raw resource

After Phase C, present sample 2-3 facet files (one per axis) to user and gate on review.

Incremental mode: Axis-level subagent dispatch (per SD3 hybrid) — if an axis has 0 facets needing re-extract, the entire axis subagent is skipped. For each facet that does run, existing 7 sections are IMMUTABLE; new evidence is appended to ## Evidence Log > ### R<N> Addendum section at file end. Conflicts (new evidence contradicts existing claim) write to ## Conflicts to Review sub-section AND register via manifest_tool.py add_conflict. A special reframe conflict class applies when a new source overturns an existing facet's TL;DR / core thesis (not a local data-point clash) — Phase C prepends a controlled ## ⚠️ R<N> Reframe banner above TL;DR as the sole write outside the Evidence Log. See references/incremental.md § "Reframe banner".

Phase D - Synthesize about_me.md

Goal: Meta-distill all Layer B facet files into Layer A about_me.md with 60-line hard cap and 7 sections.

Subagent prompt: references/phase-d-synthesize.md

Sections (in order):

1. Who I Am - identity, primary role, current setup
2. Domain Expertise - what they actually do well
3. Communication Style - how they show up across channels
4. Opinions / 行為傾向 - durable behavioural patterns
5. Stories & Lessons - 2-4 concrete anecdotes with date / amount / outcome
6. Presence & Distribution - where they exist online + identity scheme
7. Strengths & Gaps - honest both sides

Hard rules:

• 60-line cap including frontmatter
• Every section must be high-density (no filler sentences)
• Apply kill list aggressively (a sentence with "passionate / authentic / 重視 / 熱愛 / 深度" = rewrite)
• Apply swap test (a sentence that could describe a different person = rewrite)
• Stories must have specific date + specific amount + specific outcome (no vague "often" / "sometimes")
• Preserve contradiction visibly (e.g. "necessary expense miserly + impulse buy lavish")

After Phase D, present about_me.md to user and gate on review.

Incremental mode: Phase D is gated by a hybrid threshold — runs only when delta_evidence ≥ 20 OR conflicts_pending > 0 OR last_about_me_age > 7 days OR --force-synthesize. Snapshot-before-synthesize mandatory; after synthesize, scripts/quality_check.py runs a 6-dimension gate, but only the 5 hard-fail dimensions (line_count / kill_list / swap_test_density / section_count / stories_anchored) trigger auto-rollback on FAIL (via manifest_tool.py rollback about_me + PENDING_USER_ACTION); the 6th, sensitive_leak, is warn-not-fail — it emits WARN, is excluded from all_pass / exit code, and never rolls back (三軸全開 architectural decision). See references/incremental.md § "Snapshot-then-synthesize flow".

Phase E - Confirm and Optional Wire

Goal: Show final output, ask wire target, optionally create symlink / source-path config.

Questions:

• "Connect this to OpenClaw workspace? (writes symlink in workspace config)"
• "Connect this to CC memory dir? (registers in CC project memory)"
• "Both? Neither (just leave at ~/Documents/me-profile/)?"

If wire:

• Create symlink at the target location pointing at ~/Documents/me-profile/about_me.md
• Append target to _manifest.json under wired_targets

Always:

• Write final _manifest.json with: last_run, source_path, source_hash, facet_list, axis_classification, wired_targets, agent_used per phase

Quality discipline (PoC-validated, cross-phase)

These rules apply throughout. Detailed templates in references/about-me-template.md.

• Evidence-tier tagging mandatory - every fact-bearing sentence in every facet file ends with verbatim (direct quote / exact string from source), artifact (count / ID / structural pattern derivable from source), or impression (inference - lowest weight). Sentences without a tier marker fail review.
• Kill list self-check - the following words / phrases are forbidden in output unless lifted as a direct quote: passionate / authentic / 重視 / 熱愛 / 深度 / 全心投入 / 用心 / 致力於 / 注重 / 在乎. Catch yourself writing them = rewrite that sentence.
• Swap test - read each output sentence and ask "could this sentence describe a different person?" If yes, the sentence is generic = rewrite with specifics (name, date, amount, count).
• Preserve contradiction - three contradiction types are kept, not resolved:
  • Temporal (3-month-ago behaviour vs current behaviour)
  • Domain (work-self vs life-self - different communication styles)
  • Essential (genuinely conflicting traits - e.g. miserly on commute + lavish on streamer gift)
• Read-only on input - the raw resource directory is read-only. Never delete, move, modify, rename source files.
• Human review gate - between Phase A→B, B→C, C→D, D→E, present output and explicitly ask for confirmation before proceeding. Never run the full pipeline unattended unless user invokes --auto flag.

Conversational UX

• Default mode is conversational - ask questions one at a time, not in batches
• Each phase ends with a review gate
• Don't show CLI flags unless user explicitly types one
• Expert mode (--discover / --wire openclaw / --axis sensitive / --auto) for power users

Detailed reference files

For complete subagent prompts and discipline templates:

• references/phase-a-discover.md - Phase A unsupervised discovery prompt (with incremental mode section)
• references/phase-c-life.md - Phase C life axis prompt (with incremental mode section)
• references/phase-c-work.md - Phase C work axis prompt (with incremental mode section)
• references/phase-c-sensitive.md - Phase C sensitive axis prompt (with incremental mode section)
• references/phase-d-synthesize.md - Phase D about_me synthesis prompt (with incremental mode + quality gate section)
• references/incremental.md - Full incremental flow contract — when to use, 5-stage flow, manifest_tool.py cheatsheet, merge policy (dedup / facet collision / append-only), conflict surface format, completion report template, daily cron context
• references/about-me-template.md - shared 7-section template + kill list + swap test + evidence-tier spec + contradiction taxonomy
• references/sensitive-handling.md - sensitive axis specific rules (finance / mental health / ACG / identity masking)

Helper scripts (stdlib only):

• scripts/manifest_tool.py - manifest helper (run python3 scripts/manifest_tool.py --help)
• scripts/quality_check.py - about_me 6-dimension quality gate
• scripts/r1_backfill.py - one-shot R1 → v2.0 manifest backfill