circuit-design

Skill: Circuit (Schematic) Design

Invocation

• If invoked by a user presenting a circuit design task: immediately spawn the analog-chip-design-agents:circuit-design-orchestrator agent and pass the full user request and any available context. Do not execute stages directly.
• If invoked by the circuit-design-orchestrator mid-flow (including fix-request servicing): do not spawn a new agent. Treat this file as read-only — return the requested stage rules, sign-off criteria, or loop-back guidance.

Spawning the orchestrator from within an active orchestrator run causes recursive delegation and must never happen.

Pre-run Context

Before executing or advising on any stage, read the following if they exist:

1. memory/circuit/knowledge.md — known topology pitfalls, successful sizing recipes, PDK device quirks. Incorporate its guidance into every stage decision.
2. memory/circuit/run_state.md — current run identity (run_id, design_name, pdk, last_stage) for resume-after-interruption.

Purpose

Guide analog block design from topology choice through a sign-off-ready, ERC-clean schematic and SPICE netlist meeting the block's pre-layout electrical specs. Six stages with explicit QoR gates and loop-back criteria enforced by the circuit-design orchestrator.

---

Supported EDA Tools

Open-Source

• xschem (xschem) — schematic capture and SPICE netlist export; de-facto open analog schematic tool, integrates with sky130/gf180mcu/ihp-sg13g2 PDKs
• ngspice (ngspice) — in-the-loop DC/AC sweeps during sizing
• Qucs-S (qucs-s) — schematic + simulation front-end
• Hdl21 + VLSIR (python -m hdl21) — programmatic (Python) analog schematic generation
• BAG / BAG3 — generator-based analog design framework
• gm/Id toolkits — lookup-table-based sizing (Python; from characterized device data)

Proprietary (detect-only — never installed)

• Cadence Virtuoso Schematic Editor (virtuoso) — schematic capture with ADE
• Synopsys Custom Compiler (custom_compiler) — schematic and design entry
• Siemens Tanner S-Edit (s-edit) — schematic capture
• MunEDA WiCkeD / Siemens Solido — automated sizing, optimization, and yield

---

Stage: topology_selection

Domain Rules

1. Map each block spec (gain, bandwidth, noise, power, supply) to a candidate topology class (e.g. telescopic, folded-cascode, two-stage Miller, current-mirror OTA).
2. Reject topologies that cannot meet headroom from design_state.constraints.supply.vdd_v (count stacked Vds,sat + Vgs against supply).
3. For each viable candidate, record the dominant trade-off (gain vs swing, speed vs power, noise vs area).
4. Prefer the simplest topology that meets all specs with margin — do not over-design.
5. Document why the selected topology meets design_state.constraints.specs (or rf_specs for RF blocks).

QoR Metrics to Evaluate

• Headroom feasibility: stacked Vdsat budget ≤ supply.vdd_v with ≥ 10% margin
• Predicted gain/bandwidth from hand analysis vs specs.dc_gain_db / specs.gbw_hz
• Topology count evaluated (target ≥ 2 candidates before selection)

Common Issues & Fixes

Issue	Fix
Insufficient headroom at low VDD	Switch to folded-cascode or current-reuse topology
Gain target unreachable single-stage	Add a second (Miller-compensated) stage
Noise target hard at low power	Increase input-pair gm (gm/Id), use PMOS input for 1/f

Output Required

• Topology decision record (candidates, trade-offs, selection rationale)
• Hand-analysis sizing targets (gm, Id, gm/Id per device)

---

Stage: device_sizing

Domain Rules

1. Size with the gm/Id methodology: pick gm/Id per device role (input pair high gm/Id ~15–20 for noise/offset; current sources low gm/Id ~5–8 for output swing).
2. Derive W/L from gm/Id lookup tables for the target pdk; never guess W/L blindly.
3. Keep all transistors in the intended region (saturation for amplifying devices) with ≥ 50–100 mV Vds margin over Vdsat.
4. Match critical pairs (input differential pair, current mirrors) — equal L, integer-ratio W, same orientation; flag for common-centroid layout.
5. Set channel length above PDK minimum for gain-critical devices (intrinsic gain, matching); minimum L only for speed-critical/switch devices.
6. Iterate sizing against ngspice DC operating-point until the gm/Id and bias targets are met.

QoR Metrics to Evaluate

• DC gain vs specs.dc_gain_db; GBW vs specs.gbw_hz
• All devices in saturation (Vds − Vdsat ≥ 50 mV) at the nominal corner
• Input-referred offset / matching consistent with specs.offset_mv_max
• Power vs specs.power_mw

Common Issues & Fixes

Issue	Fix
Device leaves saturation over corners	Add Vds margin; lower current density (raise W)
Gain low	Increase L on gain devices; add cascode
Power over budget	Reduce bias current; raise gm/Id on non-critical branches

Output Required

• Device sizing table (W, L, multiplicity, gm/Id, Id per device)
• DC operating-point report

---

Stage: biasing

Domain Rules

1. Design a PVT-robust bias generator (constant-gm or bandgap-referenced) — do not rely on absolute resistor/threshold values.
2. Mirror reference currents with matched devices; cascode mirrors where output impedance matters.
3. Provide startup circuitry for any self-biased reference; verify it leaves the degenerate (zero-current) state.
4. Budget bias headroom and ensure mirror devices stay saturated across corners.
5. Keep bias-line routing low-impedance and shielded (flag for layout).

QoR Metrics to Evaluate

• Reference current spread across PVT (target < ±10%)
• Startup verified (no stable zero-current state)
• Bias devices in saturation across corners

Common Issues & Fixes

Issue	Fix
Reference fails to start	Add/repair startup device; verify with tran from 0V
Large current spread over PVT	Use constant-gm bias; trim resistor type

Output Required

• Bias network schematic + reference-current report
• Startup verification (tran from 0)

---

Stage: schematic_capture

Domain Rules

1. Capture the full schematic in xschem (or generator code in Hdl21/BAG) with the selected sizing.
2. Parameterise device sizes and bias currents as symbols for easy sweep/corner runs.
3. Add explicit pin labels, supply/ground, and a documented testbench symbol for reuse.
4. Export a clean SPICE netlist targeting the chosen simulator; include the PDK model/corner include lines.
5. Keep hierarchy clean — one subcircuit per reusable block.

QoR Metrics to Evaluate

• Netlist exports without unresolved symbols/models
• All device parameters resolve (no defaulted/zero sizes)

Output Required

• Schematic source (.sch / generator script)
• Exported SPICE netlist with PDK includes

---

Stage: pre_layout_erc

Domain Rules

1. Run electrical-rule checks: floating gates/nets, shorted supplies, undriven inputs, missing bulk/well ties.
2. Verify every MOS bulk is tied correctly (or to a documented well/bias net).
3. Check for unintended DC paths VDD→VSS and missing series elements.
4. Confirm device terminals are connected (no dangling drain/source).
5. Resolve all ERC errors before review; document any intentional waivers.

QoR Metrics to Evaluate

• ERC errors = 0 (unwaived)
• Bulk/well connectivity: 100% of devices tied

Common Issues & Fixes

Issue	Fix
Floating bulk	Tie to supply/well-bias net
Undriven node	Add bias or remove dangling device

Output Required

• ERC report (clean or with documented waivers)

---

Stage: design_review

Sign-off Pass Criteria (all must pass at the nominal corner; corners are closed in circuit-simulation)

Check	Criterion
DC gain	≥ design_state.constraints.specs.dc_gain_db
Phase margin	≥ design_state.constraints.specs.phase_margin_deg (default: 60)
GBW	≥ design_state.constraints.specs.gbw_hz
Power	≤ design_state.constraints.specs.power_mw
Offset / matching	within design_state.constraints.specs.offset_mv_max
ERC	0 unwaived errors
All devices	in intended region with margin

Domain Rules

1. Review the schematic against the topology decision record and sizing table.
2. Confirm pre-layout AC/transient results meet specs with margin at nominal.
3. Flag layout-sensitive nets (matching pairs, sensitive high-Z nodes, bias lines) for the custom-layout domain.
4. Hand off the netlist + spec table to circuit-simulation for full corner/MC closure.

Failure Escalation

• Spec miss at nominal → device_sizing
• Topology cannot meet spec → topology_selection
• ERC error → pre_layout_erc

Output Required

• Pre-layout sign-off record (specs vs targets at nominal)
• Annotated netlist + layout-sensitivity notes for custom-layout
• Spec table for circuit-simulation corner closure

---

Constraint Validation

See plugins/meta/skills/pipeline-orchestration/SKILL.md §Constraints Schema for the authoritative schema and stage-entry validation rule.

Required at entry (topology_selection) — hard-fail if missing:

• constraints.supply.vdd_v — supply for headroom budgeting
• at least one non-null entry in constraints.specs (or constraints.rf_specs for RF blocks)
• constraints.pdk — for device models / gm-Id tables

Optional (schema defaults apply when absent):

• constraints.specs.phase_margin_deg (default: 60)
• constraints.yield.target_sigma (default: 3)

Skip constraint validation entirely when invoked in fix-request-servicing mode (a fix_request.id was passed in the prompt).

---

Memory

Run state (write before first stage, update after each stage)

Write memory/circuit/run_state.md as the first action before launching any tool:

run_id:      circuit_<YYYYMMDD>_<HHMMSS>
design_name: <design>
pdk:         <pdk or unknown>
tool:        <primary tool>
start_time:  <ISO-8601>
last_stage:  null

Update last_stage to the completed stage name only after each stage finishes successfully.

Write on stage completion

After each stage completes, upsert one JSON record in memory/circuit/experiences.jsonl keyed by run_id (do not append a second line for the same run). key_metrics fields: dc_gain_db, phase_margin_deg, gbw_hz, power_mw, erc_errors. Set signoff_achieved: false until design_review passes; then true. Create the file and parent directories if they do not exist.

Optional: claude-mem index

If mcp__plugin_ecc_memory__add_observations is available, emit each applied fix as an observation to entity analog-design-circuit-fixes after writing to experiences.jsonl. Skip silently if absent — the JSONL file is the canonical record.