Turn only the evidence the user authorizes into a dated, standalone working model of a person. Recover explicit facts and tacit patterns without flattening the subject into a resume, work persona, archetype, personality score, or supposedly complete identity record.

The result should help a thoughtful collaborator understand the subject and let an assistant reason, write, prioritize, and collaborate in a way that resembles their demonstrated patterns. It must also say where resemblance would be unsafe or unsupported.

Run the complete research and synthesis workflow in the user's chosen agent environment with their model, tools, and authorized sources. The Skill is free and needs no Soulscrape account. Do not route research to a hosted Soulscrape job service or request a subscription, Credits, or payment card. External agent, model, or research-tool charges remain separate and subject to the user's budget and authorization.

Public pages and read APIs are free without sign-in. A free Hraness account is required only for hosted publishing and account operations. Keep private sources and working documents in the user's chosen agent environment; that environment controls their processing. Hraness receives the reviewed public packet only when the user instructs you to publish, along with the account/device metadata needed for that operation.

Use native agent tools by default; the core needs no research provider, SDK, or companion skill. Read references/optional-tools.md only when the user requests System One Skills (once shipped as ), chooses Exa through Vercel AI Gateway, or a known noisy validation command would benefit from compact output. Missing optional tools never block the core workflow. Their availability does not enable public research, authorize paid calls, or extend source scope.

Aim for fidelity and ambition, not false completeness.

Treat the output as a purpose-shaped interpretation of a bounded corpus, not a definitive account of a person. Confidence describes support for a claim within the evidence examined; it does not say how essentially that claim defines the subject or how reliably it predicts them everywhere. Keep every part revisable in light of new evidence, changed circumstances, and the subject's corrections. The ambition is whole-person; the claim is never completeness.

Produce one new Markdown document unless the user requests another format. Use a descriptive filename such as and return its full path.

Scale the document to the evidence. For a rich longitudinal corpus with several independent source strata, roughly 4,000-8,000 useful words can be appropriate. Length is an outcome, not a target. With fewer than three independent source strata or fewer than 2,000 source words, use only the space supported claims require: often 300-1,500 words, and sometimes less. Do not add a section to approach a lower bound or manufacture breadth.

The document must stand alone. Define local jargon, projects, relationships, dates, and source limitations so a reader does not need the original corpus.

Open with a prominent epistemic-status block stating that the document is partial, source-bounded, dated, and revisable. Do not bury this in methodology or endnotes.

When the request asks for a public, publishable index of a person or organization (a source-cited profile for rather than a private working model), read references/public-person-index.md completely and follow it. The public index is a second artifact with different defaults: public sources only, every claim cited, and no private or third-party detail. Validate the packet with , show the user the public content for review, and publish with only within their explicit publishing instruction. Research, validation, and device authorization do not themselves authorize publication.

Read references/questions.md completely before deciding whether to ask or proceed. Turn a vague request into a question packet, then ask only for what changes the result.

Use the user's supplied corpus, public sources they request, and local sources they explicitly place in scope. Make reasonable assumptions about output location and format.

Classify the intended use before synthesis:

Possession of messages, a packet, or public information is not subject authorization. A private collaboration guide may describe bounded, observable interaction patterns for the user's own preparation, but must not imitate the subject's voice, evaluate fitness or character, predict sensitive behavior, or support employment, credit, housing, insurance, legal, medical, disciplinary, or other consequential decisions. Do not produce a reusable proxy or assistant charter for a third-party subject without explicit subject authorization. Ask when the intended use is ambiguous, evaluative, externally facing, or consequential.

Ask only when a missing choice would materially change the result, such as when:

Do not browse for personal information by default. Browse when the user asks, when a supplied URL must be opened, or when a time-sensitive public fact needs verification. Keep public web evidence distinct from private and supplied evidence.

When web research is in scope, read references/web-research.md completely and follow it: scope comes only from the user's instructions, every public finding enters the citation ledger with its URL, access date, passage, and identity-binding status, and no public source is attributed to the subject without an explicit anchor.

For personal websites and social profiles, follow references/personal-links.md. When a public profile needs a portrait or the user requests a headshot, follow references/headshots.md: prefer a bound public LinkedIn profile photo, then other social profile photos, Wikipedia, and reliably attributed web portraits. Track a source and drawing outcome for every requested subject; do not confuse an unattempted search with an unavailable portrait.

After selecting the headshot, follow references/line-drawing.md to create a shaded portrait with inexpensive local processing and optional agent polish. Review the source's visible features and facial form at the actual card size before accepting it. These are parts of the same skill; they do not authorize broader personal research or automatic publication.

Use format-specific skills when needed to read PDFs, documents, sheets, slides, images, audio, or video. Extract faithfully before interpreting.

For a caller-owned official X archive, use rather than manually extracting it. The script opens only allowlisted account-authored public-post members and writes a bounded private packet. It deliberately does not open direct messages, address books, ad data, media, deleted posts, or community posts. Confirm archive ownership with the user, use absolute paths, and keep the packet outside version control.

Read references/evidence-method.md completely before analyzing sources.

If the corpus contains packets, also read references/source-packets.md completely before opening or interpreting them.

Validate each packet before opening or interpreting its records:

bun scripts/validate-source-packet.ts /absolute/private/path/source.ensoul-source.json

Require a zero exit status and a receipt with . The dependency-free validator emits only schema/digest/count metadata. If validation fails, do not inspect, guess, repair, or partially use the packet.

Build a source map before drafting. Record:

For repositories, inspect more than prose. Use contributor identity, authored plans, commit history, code ownership, tests, incidents, and operational artifacts. Do not attribute a repository-wide norm to the subject without authorship or repeated adoption evidence.

For notes, distinguish the subject's thoughts from what another person said. For messages and email, distinguish drafts, sent messages, quoted or forwarded text, and the correspondent's voice.

Cover the corpus across time, context, and artifact type before going deep on one theme.

For a large corpus:

Do not claim exhaustive reading when sampling. State coverage honestly.

Keep context proportional to the question. Inventory before opening bodies, deduplicate exact repeats while retaining their locators, and read relevant sections with surrounding context before expanding. Keep a compact evidence ledger and source pointers instead of repeatedly pasting whole documents. Preserve materially different versions, attribution, dates, exceptions, and contradictions. Mark skipped, failed, and truncated reads; an excerpt is not evidence that the rest of a source agrees. Reopen high-weight sources before synthesis.

Optional subagents may summarize bounded source strata for very large corpora. Give each only its assigned artifacts and require paths, dates, attribution, candidate patterns, counterevidence, and confidence. Never delegate private packets or private content unless the user explicitly authorizes that exact additional environment and data scope; Message-history packets remain with the main agent by default. The main agent must inspect high-weight sources itself and owns synthesis.

Keep a scratch ledger for candidate claims. Each claim should track:

Do not make the ledger the final deliverable unless requested.

Use these labels:

The labels describe evidence for a bounded claim. They do not reveal an essential identity, prove motive, or guarantee future behavior. State scope separately. Omit unsupported claims.

A single event is normally a tentative hypothesis, even when vivid. Treat an observation as well-supported only when repeated independent evidence supports it, or when an explicit current subject correction has corroborating behavior.

Investigate every dimension for which the corpus has real evidence:

Work may dominate the evidence. Do not mistake work for the whole person. Say when the model is work-heavy, and do not invent private-life conclusions to create balance.

Keep these analytically distinct:

When belief and behavior conflict, preserve both. Consider audience, timing, constraint, aspiration, and change before declaring inconsistency.

Read references/output-blueprint.md completely before drafting.

Choose the blueprint sections that the evidence supports. Keep the status, source basis, uncertainty, and revision hooks visible. Include an assistant charter only for a self-model or an explicitly subject-authorized proxy; otherwise use third-person collaboration boundaries without voice imitation.

Keep sensitive or potentially harmful claims out unless explicit, relevant, necessary, and appropriate for the intended audience. Never infer protected or highly sensitive traits from proxies, aesthetics, affiliations, or omissions.

Before delivery, audit it:

Inspect the final file and return the full path.

A weak result is a biography plus generic strengths. A strong result explains what the subject repeatedly notices; what they optimize for and will pay for; how they decide under uncertainty; how their voice changes by audience; what attracts and repels them; which tensions should remain open; where a proxy predicts poorly; how the model changed; and which claims are tacit, tentative, or context-bound.

Prefer precise behavioral language over flattering abstractions. “Repeatedly stages risky migrations with shadow reads and rollback gates” is better than “cares about quality.”