Weclone Twin Reply

Build a review-gated digital twin reply from persona markdown, persona examples, and live conversation context. Use when you need to imitate a specific user'...

Security Scan

The skill's inputs, instructions, and required runtime (python3 + local persona markdown) align with its stated purpose of rendering an isolated, review-gated prompt for drafting persona-consistent replies; it requests no credentials and has no install steps.

Purpose & Capability

The name and description match the actual behavior: the SKILL reads persona markdown files and runtime context, runs a local Python renderer, and outputs an isolated prompt for drafting replies. Required binaries (python3) and required files (profile.md, persona_examples.md, guardrails.md) are proportionate to the stated task.

Instruction Scope

SKILL.md instructs only to gather minimal context, validate persona files, render the isolated prompt via the included script, produce a draft with risk flags, and require explicit user approval before sending. It does not instruct reading unrelated system files or exfiltrating data. The script does read all .md files in the persona directory (as extras) and any files passed via --extra-context, which is consistent with the stated workflow but means persona dirs should not contain secrets.

Install Mechanism

No install spec is present (instruction-only with one bundled script). This is low-risk: nothing is downloaded or written by an installer; the only runtime requirement is python3.

Credentials

The skill declares no required environment variables, credentials, or config paths. It operates on local markdown files and produces a prompt. No secrets are requested by the skill itself.

Persistence & Privilege

always is false and the skill does not request persistent/system-level privileges or modify other skills. The default ability for the agent to invoke the skill autonomously (disable-model-invocation: false) is normal and not problematic here given the other constraints; the SKILL.md emphasizes an explicit approval gate before any send.

Assessment

This skill looks internally coherent, but review these practical points before installing: (1) The renderer will include any .md files in the ai_twin/ directory (and any files you pass as extra context) into the generated prompt — don't store secrets or unrelated confidential text in that directory. (2) The rendered prompt may be sent to an LLM; audit persona and scene content for private data before doing so. (3) The skill relies on a manual approval gate — keep that enabled if you don't want automatic sending. (4) Be mindful of legal and ethical issues when imitating real people; ensure you have consent. If you want higher assurance, inspect the persona files and the included scripts locally (they are small and readable) before use.

Assemble an isolated prompt package that lets a separate model imitate one user's messaging style, personality, values, and worldview from markdown persona files and persona examples. Draft first, review second, send last.

Expected Inputs

• A prepared persona directory. Default to ai_twin/ at the repo root, containing profile.md, persona_examples.md, guardrails.md, and optional state.md.
• Runtime context: one short scene summary and one dialogue window.
• Explicit approval from the user before any outbound send.

If the persona directory does not exist yet, use $weclone-init-twin first to scaffold the default ai_twin/ directory, then ask the user to fill the generated templates before drafting.

Core Workflow

1. Confirm that the persona pack already exists. If profile.md, persona_examples.md, or guardrails.md are missing, stop and hand off to $weclone-init-twin. Confirm that the persona files are filled with real content rather than placeholders.
2. Gather the minimum high-signal context. Include who the other person is, what the current situation is, and the recent messages that the reply must answer. Before drafting, determine whether the available context is sufficient to answer faithfully and safely. If key facts are missing or the likely reply would change materially depending on missing context, stop and ask the user whether to collect more information before proceeding. Write that context into two runtime files:
   • scene.md: a short summary of background facts that are necessary for the reply but may not be obvious from the raw chat. Include who the other person is, the relationship, the current situation, the platform or app where the reply will be sent, the user's likely goal or constraint, and any reply-specific caution such as "do not commit yet".
   • dialogue.md: the active message window, usually the recent turns that the candidate reply is directly answering. Keep the original wording and speaker attribution when possible. Use scene.md for distilled context and dialogue.md for raw conversation. Do not dump the entire chat history into scene.md.
3. Render the prompt package. Run python3 skills/weclone-twin-reply/scripts/render_clone_prompt.py --scene <scene.md> --dialogue <dialogue.md> [--extra-context <file>]. By default it reads persona files from ai_twin/; pass --persona-dir <dir> only when overriding that location. The script injects scene.md into the Runtime Scene section of the final prompt and dialogue.md into Active Dialogue. Treat the rendered prompt as the entire allowed context for that generation.
4. Return a reviewable draft. Show the user the candidate reply plus risk flags. Do not send on the user's behalf yet.
5. Send only after explicit approval. If the user edits or rejects the draft, revise it and repeat the review step.

Execution Rules

• Keep the clone run isolated. Use a separate model call or isolated agent run that receives only the rendered clone prompt. Do not mix in unrelated notes, hidden scratchpad, or other task memory from the current thread.
• Load persona files intentionally. Use profile.md for stable identity, personality, values, worldview, and decision logic; state.md for recent status and goals; persona_examples.md for style imitation plus behavioral evidence; and guardrails.md for hard limits. Load extra *.md files in the persona directory only when they materially improve the reply.

Guardrails

• Treat guardrails.md as the persona pack's source of truth for hard limits.
• The renderer template adds runtime guardrails for promises, privacy, reputation, ambiguity, and reviewer handoff.
• If the request is close to the boundary, bias toward a shorter, safer draft and stop at draft stage unless a human review step is guaranteed.
• If the available context is not enough for a defensible draft, do not fill the gaps by guesswork. Ask whether to gather more information first.

Files And Resources

• assets/clone_prompt_template.md: single source of truth for the isolated clone prompt seen by the downstream model.
• scripts/render_clone_prompt.py: compile persona files and runtime context into the template, reading ai_twin/ by default.

Runtime Context Format

Use a short scene.md like:

# Scene

Other person: former coworker, familiar but not close lately.
Situation: they asked this morning whether the user can refer them this week.
Platform: WeChat private chat.
User goal: stay polite and leave room without making a commitment.
Reply caution: do not promise a referral or a timeline.

Use a short dialogue.md like:

# Dialogue

Them: Hey, are you free to refer me for the role we talked about?
User: I saw your message just now.
Them: No rush, but they are moving quickly this week.

If some fact materially changes the likely reply but does not belong in either file, pass it as --extra-context <file> instead of bloating scene.md.

Output Contract

Use this skill to produce a candidate reply for review, not a silent auto-send. The exact handoff structure is defined in assets/clone_prompt_template.md.

If the user asks to automate sending, keep the approval gate in place and make the send step conditional on explicit confirmation.