DV-Gym | Stage 1: AI-Era Design Verification Foundations

Lecture 3: Skills: Reusable Playbooks for Agents

📖 What Is a Skill?

  • Packaged Knowledge: A structured directory (SKILL.md + scripts + examples).
  • Specialized Playbook: Teaches an agent how to execute a complex task repeatably.
  • Institutional Memory: Encodes senior engineer intuition so junior engineers and agents execute consistently.

🎯 Core Advantages

  • Version Controlled: Tracked in Git alongside testbenches and VIPs.
  • On-Demand Loading: Injected only when needed to prevent context window bloat.
  • Portability: Reusable across different projects, chat interfaces, and agent runners.
www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

Skills vs. Tools vs. Rules

🛠️ Tools vs. Skills

  • Tools (Capabilities): The raw API or CLI actions available to an agent (run_sim, grep_log, git_checkout).
  • Skills (Playbooks): The strategic procedure that determines which tool to run, when, and in what order.
  • Analogy: Tool = Oscilloscope probe; Skill = How to debug an eye diagram violation.

📋 Rules vs. Skills

  • Project Rules (AGENTS.md): Global, always-on constraints (e.g., active-low reset naming conventions).
  • Skills (SKILL.md): Task-specific playbooks triggered conditionally (e.g., only when an objection hang occurs).
  • Analogy: Rule = Company coding standard; Skill = Root-cause triage flowchart.
www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

DV Skills in Practice: Diagnostic & Formal

🔍 uvm-hang-triage

  • Objective: Systematically isolate simulation timeouts.
  • Audit Points: Tracks objection counts, detects missing item_done, flags unclocked forever loops, and checks drain time.
  • Constraint: Strictly forbids bumping the timeout threshold to artificially mask a deadlock.

📐 sva-from-spec

  • Objective: Convert spec requirements into synthesizable SVA.
  • Rules: Enforces explicit clock/reset declarations and flags vacuous antecedent pass risks.
  • Constraint: Forbids unbounded delays (##[1:$]) without explicit verification justification.
www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

DV Skills in Practice: Coverage & Code Quality

📊 coverage-review

  • Objective: Audit functional coverage closures.
  • Checklist: Distinguishes structural line coverage from functional intent; verifies cross-coverage bins.
  • Principle: Enforces the golden rule: "100% functional coverage does not prove bug-free design."

🛡️ ai-tb-audit

  • Objective: Gatekeeper review for LLM-generated code.
  • Checks: Inspects shallow vs. deep copy methods (do_copy), missing super.* phase calls, and 2-state compares on 4-state buses.
  • Safety: Catches non-synthesizable delays and zero-delay scheduler race hazards.
www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

Anatomy of a Skill

  • A skill is a folder, not a prompt: typically skills/uvm-hang-triage/SKILL.md plus optional scripts and examples.
  • SKILL.md has two layers:
    • Frontmatter (name, description) — tells the agent when to load this playbook.
    • Body — the procedure the agent must follow once loaded.
  • Body should include: inputs, ordered steps, definition of done, and explicit never rules.
  • Scope: load on demand for matching tasks; do not dump every skill into every chat.
  • Quality bar: if a senior would not sign the procedure, it is not ready to ship as a skill.

 

www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

Example SKILL.md: uvm-hang-triage

Front matter = when to use it. Body = how to work. YAML --- belongs inside the fence so it is not a slide break.

---
name: uvm-hang-triage
description: Debug a UVM sim that hangs or times out. Use when logs show
  undropped objections, forever loops, or UVM_FATAL timeout.
---
# UVM hang triage
## Inputs
- Failing test name, sim.log, timeout / plusarg settings
## Procedure (do not patch until step 4)
1. Search the log for objection traces (raised / dropped / count != 0).
2. Find forever, wait(, and raise_objection without a matching drop.
3. Check drain time, set_timeout, and sequences that never item_done.
4. Propose a minimal patch (usually one drop_objection or join/disable).
5. Re-run smoke. Stop if compile fails or a new hang appears.
## Definition of done
- TEST PASSED (or a new, different error you report).
- RCA note: root cause + files touched. No golden / timeout-limit hacks.
## Never
- Dump a full rewritten env. Do not raise timeouts to "make it pass."

 

www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

Workflows: Multi-Step Procedures with Gates

🔄 What Is a Workflow?

  • Structured Pipeline: A deterministic sequence of tasks connecting multiple skills, tools, and reviews.
  • Quality Gates: Transition between steps requires objective validation (e.g., compile clean, zero regressions).
  • Predictable Execution: Replaces ad-hoc trial and error with repeatable engineering standard operating procedures.

🚪 Key Verification Gates

  • Gate 1 (Syntax): Clean compile with zero errors and zero linter warnings.
  • Gate 2 (Smoke): Single-seed test pass with no assertion failures or UVM errors.
  • Gate 3 (Integrity): Known-bug mutant injection to verify checker vitality.
  • Gate 4 (Human Gate): Peer review sign-off before merging into main.
www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

DV Automated Workflows: Bug RCA & Feature Bringup

🐛 Bug Triage to RCA Workflow

  1. Ingest failing log and record seed.
  2. Classify failure category (RTL, TB, or Test).
  3. Extract minimal reproducible slice.
  4. Identify root cause on waveform.
  5. Propose minimal surgical fix.
  6. Trigger localized regression suite.

✨ Spec-to-Smoke-Test Workflow

  1. Ingest chapter from architectural PRD.
  2. Generate feature & corner-case checklist.
  3. Propose sequence items & concurrent SVA.
  4. Scaffold UVM driver/monitor skeletons.
  5. Execute smoke simulation in sandbox.
  6. Human engineer reviews code diff.
www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

Model Context Protocol (MCP): The Universal Bridge

🔌 The Integration Problem

  • Verification involves fragmented silos: EDA tools, git repos, Jira tickets, spec wikis, and UCDB databases.
  • Without a standard, every AI tool requires custom connectors for every internal tool ( complexity).
  • Custom plugins break with every tool version update.

⚡ The Solution: Open MCP Standard

  • The "USB-C" of AI Integrations: One uniform protocol connecting LLM agents to external platforms.
  • Standardized Server Primitive: Connect an agent to any simulator, regression farm, or spec repository.
  • Dynamic Context Ingestion: Fetches real-time project state without manual pasting.
www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

The Three Primitives of an MCP Server

Tools and resources

  • Tools: actions the agent can call (may change state):
    • trigger_regression(seed, suite)
    • jira_create_issue(...)
  • Resources: read-only data the agent can load into context:
    • Spec chapters, register maps, Git diffs, Jira ticket bodies.

Prompts (the third primitive)

  • Reusable prompt templates the MCP server exposes — not the chat you typed.
  • Example: a “UVM hang triage” template that already asks for test name + log path.
  • Example: a “file a jira from this fail log” - takes log path, signature, RCA analysis.
  • The agent (or you) selects the template; it is filled with live resources/tools.

 

www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

Real-World DV Architecture: Connected MCP Servers

Bug database (Jira) MCP

  • Search open/closed tickets by error signature, block, or test name.
  • Read history: prior RCA, owners, workarounds, related PRs.
  • Write (gated): file a ticket after debug with log excerpt + suspected file.

Source control (Git) MCP

  • Blame/log around the failing module; list recent commits on the branch.
  • Fetch the patch that closed a related Jira (commit message ↔ ticket ID).
  • Open a review branch / PR after human approval.
  • Also typical: Docs MCP (spec chapters) and EDA/CI MCP (logs, smoke runs).
  • Skills still choose when to call each server; MCP only provides access.

 

www.verificationexcellence.in
DV-Gym | Stage 1: AI-Era Design Verification Foundations

Example: RCA Agent with Jira + Git

  1. Triage: EDA MCP pulls the failing log; skill classifies “AXI timeout.”
  2. Look back: Jira MCP searches past tickets with the same signature / block.
  3. Look at code: Git MCP maps those tickets to commits and shows what fixed (or failed to fix) it last time.
  4. Debug: Agent proposes a hypothesis using that history + current log (does not merge).
  5. File the bug: If it is new, Jira MCP drafts a ticket: failing test, seed, log snippet, suspected RTL/TB file, links to similar issues.
  6. HITL: Engineer confirms RCA and decides: attach to existing ticket, file new, or drop.

Principle: MCP connects Jira and Git; the RCA skill decides the lookup and filing order; the engineer still owns the ticket.

 

www.verificationexcellence.in