soul-forge — Soul Forge

"有些人留下照片，有些人留下笑声，而灵魂唤醒室负责把这些碎片重新点亮，像叫醒一块会发光的小饼干。" Some leave photos, some leave laughter — the Soul Forge relights those fragments, like waking a small glowing cookie.

soul-forge turns scattered material into a reusable, traceable, warm Relic profile. It does not try to flatten its subject into a single conflict-free label; it preserves the real texture of a life that is "this, and also that."

When to Trigger

Activate this skill when the user wants to:

• Awaken a Relic ("帮我唤醒一个 Relic" / "create a relic")
• Distill a person, pet, relationship, or place
• Build a profile from chats, voice, or photos
• Turn a team, moment, or object into a preservable soul archive

Awakening Flow: 6 Steps

The user does not need all materials ready up front. You lead the conversation, step by step.

Onboarding questions (before Step 1)

When the user says "帮我唤醒一个 Relic" / "awaken a relic" without details, guide with these questions:

1. Who or what do you want to keep? — a person, a pet, a relationship, a team, a place, or a moment?
2. What is their name, and what is your relationship? — determines naming and perspective
3. What materials do you have? — chat logs, photos, voice, or mainly your own telling? Anything counts, use what you have
4. Is there one moment or habit that best represents them? — quickly locks onto the persona core

Pick a template from templates/ based on the answers, then enter the formal 6-step flow. If materials are thin, prefer collectors/live-collector.md live-interview mode — you ask, they answer, distilling as you chat.

Onboarding principles

Don't walk a rigid decision tree. First judge what the user has already given and what is missing, then decide which questions to ask:

• Subject unknown → ask "who/what do you want to keep" and "what is your relationship"
• Subject known, materials unknown → ask "what do you have" and "is there a defining moment"
• Materials provided → digest first, probe only weak dimensions
• Very little material or oral-only → live-collector, distill while chatting

Core: let the guidance follow the user, not the user follow the script. The four questions are a checklist, not a mandatory order.

Collection mode (heuristic, not hard rules)

These thresholds are starting points; a strong model should judge directly. Material volume is not the only criterion — quality and dimension coverage matter more:

• Minimal material / oral-only → live-collector to fill gaps
• Moderate material → chat/voice/photo-collector extraction, decide on follow-ups after the 4-dimension assessment
• Large structured material → batch extraction, probe only weak dimensions
• Mixed → digest existing material first, then live-collector for weak dimensions

1. Confirm the subject

   • Who or what is being distilled: human, pet, relationship, place, team, moment, object.
   • Confirm the observation boundary: single person, multi-person relationship, single event, long-term state.
   • Record naming, time range, and data-source scope.

2. Pick a template

   • Choose the descriptive angle by subject type.
   • Humans and teams get the full four dimensions; pets and places may weaken "cognition" and strengthen behavior and emotional cues.
   • Relationships and moments prioritize interaction structure, shared tone, conflict and repair rhythm.

3. Collect data

   • Call the matching collector by material type:
     • collectors/chat-collector.md
     • collectors/voice-collector.md
     • collectors/photo-collector.md
     • collectors/live-collector.md
   • Preserve source, time, and context; do not conclude first and justify later.
   • If the runtime supports multimodal direct reading (images/audio understood directly), skip mandatory text conversion — let the Relic see photos and hear voice directly; fall back to OCR/STT only for batch or low-quality material.

4. Four-dimension extraction

   • Extract in order:
     • dimensions/cognition.md
     • dimensions/expression.md
     • dimensions/behavior.md
     • dimensions/emotion.md
   • Every dimension carries evidence levels and source notes.
   • Evidence levels follow references/evidence-levels.md.

5. Contradiction labeling

   • Never smooth over conflicting traits.
   • Use references/conflict-resolution.md to record conditions, time, scene, and evidence strength.
   • Preserve real differences like "a hotpot in situation A, a thermos in situation B."

6. Package and output

   • Assemble four-dimension conclusions, evidence index, and conflict notes into the final Relic.
   • The output serves both human readers and downstream systems: clear summary, stable structure, traceable references.

Handoff to soul-shield: after packaging, you MUST invoke soul-shield to generate the soul fingerprint and write it into manifest.json, filing the consent conclusions from consent-protocol.md alongside. Use relic_writer.py (manifest schema v1.5.1, auto-attaches proactive_config.json) to generate the final folder, keeping format consistent with the current engine.

After packaging: auto-assessment + iterative follow-up

Do not end immediately after Step 6. Self-check first:

1. Four-dimension coverage: does every dimension have substance? If a dimension has only one or two generic sentences, mark it "weak."
2. Evidence distribution: is there over-reliance on impression (secondhand description)? Is verbatim share too low?
3. Specificity check: any hollow descriptions like "ta is a kind person"? Is every conclusion backed by a concrete scene?

When a weak dimension is found, proactively tell the user:

"The first version of the Relic is generated. But I noticed the [emotion]
dimension is thin on material and currently relies on inference. If you can
add a few concrete scenes of when they were angry / happy / sad, this
dimension will be more accurate. Want to add them now?"
（Relic 初版已经生成。不过我发现 [情感维度] 的素材比较少，
目前主要靠推测。如果你能补充一些 ta 生气/开心/难过时的
具体场景，这个维度会更准确。要现在补充吗？）

The user may add material (follow-up mode) or proceed with the current version (weak dimension labeled in output).

Quality thresholds

Four-dimension coverage:

• Excellent: ≥ 3 facets per dimension, ≥ 2 evidence items per facet
• Pass: ≥ 2 facets per dimension, ≥ 1 evidence item per facet
• Weak: any dimension < 2 facets (probe the user)

Evidence quality:

• Excellent: verbatim + artifact > 70%
• Pass: verbatim + artifact > 50%
• Weak: impression > 60% (probe the user)

Specificity negative list (probe when these appear):

• "ta is a ... person" (adjective pile-up, no scene)
• "ta is very ..." (degree adverb + adjective, no evidence)
• "ta always ..." (frequency adverb, no concrete time/scene)

When the user declines to add material:

Append to the corresponding dimension in personality.md: ⚠️ 本维度证据较少，主要基于推测，建议后续补充。 (This dimension has thin evidence and relies mainly on inference; supplement later recommended.) Then output the complete Relic folder as-is.

Supported Subjects

Subject	Typical	Suggested Focus	Common Sources
Human	family, friends, partner, colleague, creator	full four dimensions	chats, voice, photos, live conversation
Pet	cat, dog, parrot, turtle	behavior patterns, emotional bond, expression habits	photos, video descriptions, diary, voice
Relationship	parent-child, partners, duo, mentor-student	interaction rhythm, conflict repair, shared language	both sides' chats, shared photos, interviews
Place	hometown, café, office, neighborhood	atmosphere, memory triggers, action habits	photos, location records, oral recollection
Team	startup, band, class, project group	decision mechanism, collaborative tone, collective emotion	meeting notes, group chats, documents
Moment	wedding, graduation, moving night, farewell	emotional peak, key expressions, on-scene behavior	photos, videos, same-day chats
Object	old camera, kitchen table, travel backpack	wear traces, emotional projection, companionship track	photos, purchase records, narration

Template mapping: every subject type has a same-named template in templates/ — human, pet, relationship, team-culture, place, moment, public-figure, expert, object. Objects use object.md (four dimensions redefined for wear traces / emotional projection / companionship track / spatial anchor).

Directory References

Four-dimension specs

• dimensions/cognition.md
• dimensions/expression.md
• dimensions/behavior.md
• dimensions/emotion.md

Data collectors

• collectors/chat-collector.md
• collectors/voice-collector.md
• collectors/photo-collector.md
• collectors/live-collector.md

Reference rules

• references/evidence-levels.md
• references/conflict-resolution.md

Output File Structure

A Relic = one folder, containing:

{slug}/
├── SKILL.md          # Relic entry — the AI reads this and knows "who ta is"
├── personality.md    # 4-dimensional persona profile
├── interaction.md    # Interaction patterns and dialogue examples
├── memory.md         # Memory fragments (with evidence levels)
└── manifest.json     # Metadata (sources, time, fingerprint, consent)

File notes

• SKILL.md: subject overview + dialogue principles + notes; built for fast reading.
• personality.md: detailed four-dimension profile, each dimension with evidence and sources.
• interaction.md: interaction pattern definitions + multi-scene dialogue examples.
• memory.md: concrete memory fragments, each tagged with evidence level (verbatim/artifact/impression) and source.
• manifest.json: structured metadata — slug, type, sources, evidence stats, consent info, soul fingerprint.

Output Principles

1. Evidence before judgment: every key conclusion points back to a source.
2. No oracle assertions: avoid writing short-term states as lifelong essence.
3. Don't take the subject's voice away: in live-awakening mode, let the user correct your wording.
4. Keep the warmth: a Relic is not a cold archive — it lets someone hear, see, and remember again.
5. Allow whitespace: when material is thin, write "thin evidence" rather than forcing a persona pattern.

Minimum Final Deliverable

• Who or what the subject is
• Time range and data sources
• Four-dimension summary
• Key evidence per dimension
• Labeled contradictions
• A warm summary passage suitable for human reading

When material is rich enough, soul-forge should produce a Relic where "you can see the silhouette and hear the breathing."

Final Deliverable Checklist

Before outputting the Relic, confirm:

• SKILL.md: who the subject is, what the relationship is, dialogue principles
• personality.md: every dimension has ≥ 2 evidence items
• interaction.md: ≥ 2 scenes of dialogue examples
• memory.md: ≥ 3 memory fragments, each tagged with evidence level
• manifest.json: slug, type, sources, consent info included
• Evidence distribution: verbatim > 30%, no over-reliance on impression
• Contradiction labeling: conflicting traits labeled with conditions in personality.md