# Spice Bench Craft > Use when authoring SPICE characterization benches or testbench netlists, especially for servo-loop or open-loop amplifier measurements, DC-servo needs, biasing struggles or railed operating points, phase-margin and crossing practice, supply-current measurement, or reusable parameterized bench-template needs. Covers LTspice and ngspice bench craft. - Skill: `cognitohazard/spice-bench-craft` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds@latest add cognitohazard/spice-bench-craft` - Raw SKILL.md: https://api.skillmd.com/api/skills/cognitohazard/spice-bench-craft/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: cognitohazard (https://skillmd.com/u/cognitohazard) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/cognitohazard/spice-bench-craft --- # SPICE measurement bench craft Author the bench from one parameter dictionary, establish the operating point, then measure. The measurement-practice notes these sections are taken from are in `references/BENCH_NOTES.md`; read that file if you need the original wording. ## The operating point comes first An amplifier with 60-100 dB of DC gain multiplies any input offset by 1,000-100,000. Driving both inputs from ideal DC sources at the same potential does not put the output at mid-rail: the amplifier's own input-referred offset (even tens of microvolts) drives the output to a supply rail, and the gain, bandwidth and phase you then measure are those of a saturated transistor stack, not of the amplifier in its linear region. Symptoms: DC gain tens of dB lower than expected, an .op output voltage within ~100 mV of either rail, device operating regions showing triode/cutoff where saturation was intended. Always check v(out) in the .op result before trusting an AC sweep. If it is not near the intended quiescent level, the measurement is invalid. ## DC servo for open-loop AC analysis The standard bench closes the loop at DC only, with elements too large to matter in the measured band: * loop closed at DC through a huge inductor; AC injected differentially LFB out inn 1T ; 1 tera-henry: short at DC, open at any AC freq CFB inn 0 1T ; 1 tera-farad: blocks DC, grounds inn for AC VIP inp 0 DC {CM} AC 0.5 * inn receives -0.5 of the AC drive THROUGH the cap if driven; or drive * single-ended and read v(out)/v(inp-inn) — both are open-loop above * the (vanishingly low) servo corner at 1/(2*pi*sqrt(LC)). At DC the feedback forces the output to the level that zeroes the input differential, so the amplifier settles at its own correct operating point. Above a few microhertz the loop is open and v(out)/v(inp-inn) is the true open-loop transfer function. ## Reading the result - Gain: magnitude at the lowest swept frequency (sweep from well below the dominant pole). - Unity crossing: use the first 0 dB crossing; verify there is only one, or evaluate phase at every crossing, because a later crossing can carry an unstable mode that a first-crossing readout misses. - Phase: unwrap before computing margin; wrapped (modulo-360) phase can show a margin that is not there. - Supply current: measure the supply source current at the .op point, not a sum of device currents. ## Sanity checks for measurement benches 1. .op first, AC second: output within the linear region? 2. Does measured DC gain roughly match gm*ro expectations (within an order of magnitude)? A 25 dB shortfall usually means the bench is wrong, not the amplifier. 3. Re-run one point at 2x sweep density: if gain or PM change, the difference comes from the sweep density or interpolation, not from the circuit. ## Render benches from one parameter dictionary Keep DUT pin order and run conditions in data, not copied through dozens of decks. A useful dictionary has `DUT_INCLUDE`, `DUT_INSTANCE`, `VDD`, `VCM`, `CL`, `TEMP`, and analysis limits such as `FSTART`, `FSTOP`, or `TSTOP`. Render the `@NAME@` host placeholders below, preserve SPICE `{PARAM}` braces, and reject output containing an unresolved `@NAME@`. Keep the title first and `.end` last. Choose include paths (absolute or deck-relative) explicitly. ### Operating-point and supply-current archetype Use the same servo as the AC bench so the `.op` result is the exact bias point the AC run will linearize around. Read `V(out)`, device regions, and `I(VDD)`; source-current sign follows the source orientation. ```spice * generated operating-point bench .include @DUT_INCLUDE@ .param VDDV=@VDD@ VCMV=@VCM@ CLV=@CL@ .temp @TEMP@ VDD vdd 0 {VDDV} VIP inp 0 DC {VCMV} @DUT_INSTANCE@ CL out 0 {CLV} LFB out inn 1T CFB inn 0 1T .op * LTspice: .options logopinfo * ngspice: .save V(out) I(VDD) @m1[gm] @m1[gds] @m1[id] .end ``` ### Open-loop AC archetype Keep the DUT instance, load, temperature, and servo identical to the `.op` deck. Single-ended `AC 1` makes `V(out)/V(inp,inn)` the unambiguous loop gain; use the differential denominator even when it is numerically one. ```spice * generated open-loop AC bench .include @DUT_INCLUDE@ .param VDDV=@VDD@ VCMV=@VCM@ CLV=@CL@ .temp @TEMP@ VDD vdd 0 {VDDV} VIP inp 0 DC {VCMV} AC 1 @DUT_INSTANCE@ CL out 0 {CLV} LFB out inn 1T CFB inn 0 1T .ac dec @POINTS_PER_DECADE@ @FSTART@ @FSTOP@ .save V(out) V(inp) V(inn) I(VDD) .end ``` For balanced drive, give `VIP` `AC 0.5 0`, replace `CFB inn 0 1T` with `CFB inn ndrive 1T`, and add `VIM ndrive 0 DC 0 AC 0.5 180`; keep the inductor as the only DC feedback path. Always compute gain and phase from `V(out)/V(inp,inn)`, unwrap phase, inspect every 0 dB crossing, and repeat at twice the point density. ### Closed-loop transient and load-step archetype Use ordinary resistive feedback for a closed-loop transient bench; the DC-only servo belongs to open-loop characterization. Parameterize both the command step and the load step so one renderer covers settling and load regulation. ```spice * generated closed-loop transient bench .include @DUT_INCLUDE@ .param VDDV=@VDD@ VLO=@VIN_LO@ VHI=@VIN_HI@ CLV=@CL@ .temp @TEMP@ VDD vdd 0 {VDDV} VIN inp 0 PULSE({VLO} {VHI} @TDELAY@ @TRISE@ @TFALL@ @TON@ @PERIOD@) @DUT_INSTANCE@ RFB out inn @RFB@ RG inn 0 @RG@ CL out 0 {CLV} ILOAD out 0 PULSE(@ILOAD_LO@ @ILOAD_HI@ @TDELAY@ @TRISE@ @TFALL@ @TON@ @PERIOD@) .tran 0 @TSTOP@ 0 @MAXSTEP@ .save V(inp) V(out) I(VDD) I(ILOAD) .end ``` Before extracting slew or settling, confirm the initial `.op` is linear and the requested step does not turn the measurement into an overload test. ## ngspice batch-output practice This server invokes ngspice with batch mode and a raw output (`-b -r`), which suppresses top-level `.meas`. Move measurements into a `.control` block as the dot-less interactive `meas` command; a dotted `.meas` inside `.control` is not valid. ```spice .control run meas ac gain_10 find vdb(out) at=10 meas ac unity when vdb(out)=0 cross=1 .endc ``` Prefer saved traces plus the server's analysis tools when they express the metric. If using `wrdata`, its columns repeat the scale vector for every dumped vector: dumping `v(out)` and `i(VDD)` yields `scale, v(out), scale, i(VDD)`, not one shared scale followed by both values. Parse repeated scale/value groups, and remember that `wrdata` writes only the text table; add an explicit `write` too if later server analysis needs a rawfile. ## Routing: pick one interface per session Pick the interface once, based on what this session can do, and stay with it. - **You can run code** (shell + Python): do all circuit work through `from ltspice_mcp.api import Api`. Look up arguments instead of guessing: `api.reference()` lists the six ops, `api.reference("")` gives the full argument tree with enums and one example; `help(api.)` shows the same. Use `run_experiments` for any LTspice run, for declared sweep/corner/Monte-Carlo matrices, and for anything that must outlive the call (`jobs` waits or cancels). Use `edit_schematic`/`verify_circuit` for `.asc` work; they do orthogonal routing and collision checks, so never hand-write `.asc` files. `analyze_results` reads any rawfile, including one you ran yourself (`raw_path=`). Keep intermediate results in your interpreter; print the crossing or the worst corner, not the full table. - **You cannot run code**: use the six MCP tools available to you and nothing else. - One exception, and say so when you use it: a hand-run ngspice one-off is fine only when you pass its raw straight to `analyze_results(raw_path=...)`.