Mode C (Policy Sensing and Skill Evolution) in Patent Disclosure Skill: Complete Architecture Guide

Mode C is the optional "skill-evolution bypass" that fetches recent patent policy trends from official sources and generates a human-reviewed checklist before any skill files are modified.

Mode C, also called Policy Sensing or Skill Evolution, is a specialized operational mode in the patent-disclosure-skill repository. Unlike the core disclosure workflow (Mode A) or plain-reading mode (Mode B), this user-triggered side-path provides a policy radar for detecting regulatory shifts—without automatically altering your skill configuration.

How Mode C Activation Works

Mode C is disabled by default and requires explicit user invocation. According to SKILL.md (lines 18-20), valid trigger commands include:

• /patent-evolve
• /技能进化 (Chinese command)

When triggered, the system executes a five-stage gated workflow that never interrupts normal patent drafting operations.

The Five-Stage Mode C Workflow

Stage 1: Intake and Guardrails Loading

The runner first reads two critical prompt files:

read_file("prompts/evolution/guardrails.md")  # Safety boundaries

read_file("prompts/evolution/intake.md")      # User intent gathering

The guardrails.md file (lines 12-13) establishes hard prohibitions against automatic edits, while intake.md structures how user objectives are captured.

Stage 2: Policy Research Execution

research_results = run_prompt("prompts/evolution/research.md")

This stage performs WebSearch plus official site scraping, targeting authoritative sources like the China National Intellectual Property Administration (CNIPA). The research.md prompt ensures only verified policy documents are retrieved—unverified media sources are explicitly excluded per guardrails.md (lines 15-16).

Stage 3: Evolution Backlog Generation

backlog_path = emit_backlog(research_results)

# Writes: outputs/evolution/EVOL-YYYYMMDD-HHMM.md

The emit_backlog.md prompt formats findings into a time-stamped evolution checklist containing viewpoint-to-source-URL mapping tables. Output files follow the schema defined in references/schemas/evolution_backlog.schema.yaml and are git-ignored by default.

Stage 4: Human Review Gate

show_to_user(backlog_path)
decision = await_user_decision()  # Blocks until human input

The guardrails.md human-review gate (lines 28-37) presents four decision options:

Decision	Action
全部采纳 (Adopt All)	Proceed to apply all suggested changes
采纳 E10 (Adopt Specific)	Apply only selected entries
全部搁置 (Defer All)	Archive without modifications
沉淀到 docs/evolution/	Move to documentation without skill changes

Stage 5: Conditional Application

if decision.startswith("全部采纳"):
    read_file("prompts/evolution/apply_after_confirm.md")
    apply_changes(decision)  # Minimal, targeted edits only

The apply_after_confirm.md prompt executes minimal prompt modifications only after explicit reviewer approval. No changes touch SKILL.md, core prompts, or reference schemas without this human gate.

Mode C Output Structure

Primary Output Location

outputs/evolution/
└── EVOL-20250115-1430.md   # Time-stamped checklist (git-ignored)

Post-Confirmation Archival

Confirmed checklists may be copied to docs/evolution/ per docs/evolution/README.md (line 5), creating an auditable policy history without polluting the active skill configuration.

Safety Architecture and Guardrails

Mode C implements defense-in-depth protections:

• No automatic skill mutation: guardrails.md (lines 12-13) mandates human approval before any SKILL.md, prompt, or schema modifications
• Source authority validation: Only official government and examination guidelines qualify; industry blogs and unverified media are rejected
• Workflow isolation: Mode C runs mutually exclusive with Modes A and B per SKILL.md (lines 77-89), ensuring draft integrity

Complete Mode C Invocation Example

# Command-line trigger detection

if user_input.strip().lower() in {"/patent-evolve", "/技能进化"}:
    # Stage 1: Load guardrails and intake

    read_file("prompts/evolution/guardrails.md")
    read_file("prompts/evolution/intake.md")
    
    # Stage 2: Execute policy research

    research_results = run_prompt("prompts/evolution/research.md")
    
    # Stage 3: Generate evolution backlog

    backlog_path = emit_backlog(research_results)
    
    # Stage 4: Human review gate (blocking)

    show_to_user(backlog_path)
    decision = await_user_decision()
    
    # Stage 5: Conditionally apply approved changes

    if decision.startswith("全部采纳"):
        read_file("prompts/evolution/apply_after_confirm.md")
        apply_changes(decision)

Helper functions (read_file, run_prompt, emit_backlog, show_to_user, await_user_decision, apply_changes) are implemented in tools/shared/ within the repository.

Key Files in Mode C Architecture

File Path	Purpose in Mode C
SKILL.md	Central routing table defining Mode C activation and isolation rules
prompts/evolution/guardrails.md	Global safety boundaries and human-review gate
prompts/evolution/intake.md	User intent capture for policy sensing sessions
prompts/evolution/research.md	WebSearch and official site scraping execution
prompts/evolution/emit_backlog.md	Checklist formatting and EVOL-*.md generation
prompts/evolution/apply_after_confirm.md	Minimal post-approval modifications
docs/evolution/README.md	Archival procedures for confirmed checklists
references/schemas/evolution_backlog.schema.yaml	Structural schema for evolution backlog tables

When to Use Mode C (Policy Sensing and Skill Evolution)

Trigger Mode C when you need to:

1. Detect regulatory shifts before they affect patent examination outcomes
2. Validate skill currency against latest CNIPA or WIPO guidance
3. Audit policy influence on your disclosure workflows with full traceability
4. Stage prompt updates through a reviewable, reversible process

Avoid Mode C for urgent drafting tasks—its gated workflow introduces deliberate latency for safety.

Summary

• Mode C is an optional, user-triggered bypass for policy sensing and skill evolution, disabled by default in the patent-disclosure-skill repository
• It operates through a five-stage gated workflow: guardrails loading → research → backlog emission → human review → conditional application
• Critical safety feature: No automatic edits to SKILL.md, prompts, or schemas; all changes require explicit human approval per guardrails.md
• Outputs are time-stamped checklists (EVOL-*.md) stored in outputs/evolution/, with optional archival to docs/evolution/
• Mode C is mutually exclusive with Modes A and B, ensuring zero impact on active patent disclosure workflows

Frequently Asked Questions

How do I activate Mode C in Patent Disclosure Skill?

Send the trigger command /patent-evolve or /技能进化 to the skill runner. Mode C does not activate automatically—you must explicitly request policy sensing. The trigger is recognized in SKILL.md (lines 18-20) and routes execution to the evolution prompt sequence.

Can Mode C modify my skill files without permission?

No. guardrails.md (lines 12-13) enforces a hard prohibition against automatic edits. The human-review gate (lines 28-37) blocks all file modifications until you explicitly select 全部采纳 (Adopt All), 采纳 E10 (Adopt Specific), or archive the results. This is architecturally guaranteed, not merely a preference setting.

What sources does Mode C use for policy research?

The research.md prompt restricts sources to official government and examination authorities—primarily CNIPA publications and WIPO guidance documents. guardrails.md (lines 15-16) explicitly rejects unverified media, industry blogs, and third-party interpretations. Only authoritative URLs populate the evolution backlog table.

Where are Mode C outputs stored and for how long?

Primary outputs write to outputs/evolution/EVOL-YYYYMMDD-HHMM.md and are git-ignored. After your approval, you may optionally move checklists to docs/evolution/ for permanent archival per docs/evolution/README.md (line 5). Unapproved outputs remain in the temporary outputs directory and may be purged per your retention policies.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}