Marketing block #1 (draft for #71, 2026-10-05)¶
One post this week, from what already exists (ROADMAP.md, "How we work", rule 5). Posted on LinkedIn on 2026-10-05 by the owner, as a text post with the 7-answers image (no video): the post (issue #71). Every post is listed in posted-log.md.
Picks¶
Video (1): Documents\MagicON-videos\_final-mp4\mcp-ab-test-final-newvoice.mp4 (92.7 s, your
new voice, head and tail noise trimmed, slide 5 dated 16 Sep 2026). It is about a PCB stackup: Claude
alone gave 7 different stackups, and the MagicON solver gave the same one all 4 times it was run.
Script: docs/scripts/MCP_AB_TEST_DEMO_VIDEO_SCRIPT.md; figures from docs/reports/mcp-stackup-solver-ab-test.md.
β Fairness note for the post: of the 7 "Claude alone" answers, only 3 were fresh test runs. The other 4 are older answers from earlier chats, with unknown app and settings (report Β§2.2). The post below says "answers", not "runs", for that reason.
Articles (4, queued for the next weeks, one per block). The stackup article goes FIRST, because it matches the video:
| Week | Article | Why |
|---|---|---|
| 2 | docs/articles/linkedin-series/pcb_stackup/article-2-materials-impedance.md |
Materials and impedance targeting: the topic of the video's stackup question |
| 3 | docs/articles/linkedin-series/ai-native-vs-ai-featured-cad.md (1,318 words) |
The "AI numbers must come from a solver" argument |
| 4 | docs/articles/linkedin-series/pcb_stackup/article-1-hidden-complexity.md |
Opens the stackup series |
| 5 | docs/articles/linkedin-series/rf_transmitter_design/article-7-link-budget.md |
Link budget is the first RF tool people try |
Before posting one, check posted-log.md (started 2026-10-07): it lists every post since block 1. Posts from before block 1 are not in it, so glance at your LinkedIn page too. I read only the series README for article 2, not the article itself.
Post text for this week: MAIN, no video (owner decision 2026-10-05: a video felt like too much advertising)¶
Attach one image: docs/scripts/mcp-ab-test-tries-table.png (the 7 answers Claude gave without design tools).
Ready-to-paste copies (owner's computer, outside the repo): C:/Users/hakic/Documents/MagicON-posts/block-1-stackup/ has post.txt, first-comment.txt and seven-answers-table.png. Keep them identical to the text below.
For PCB and RF engineers and hobbyists who use AI: I asked Claude to design a PCB stackup, then asked again. Seven answers without any design tools gave me seven different stackups for the same targets.
I built MagicON, a stackup solver that any AI assistant supporting MCP can call, to see if that changes. Given the same targets, it returned the same stackup in all four test runs, because it's deterministic.
π The question I used: "I'm designing a 6-layer PCB for a 10 GHz RF board. Finished thickness 62 mil, single lamination (not HDI), lead-free assembly. I need 50 ohm single-ended microstrip on the outer layer, balancing cost against performance. Recommend a specific stackup."
It runs as an MCP server (the standard way an AI assistant calls outside tools), with no account and no API key. Install command in the first comment.
π¬ MagicON is a free beta for now, with call limits. I'm building it with input from PCB and RF designers, so tell me what is missing or wrong: support@pcbgenerator.com
Would you trust a layer recipe your assistant can't reproduce?
MagicON is not affiliated with or endorsed by Anthropic. Claude is a trademark of Anthropic.
PCBDesign #PCBStackup #RFDesign #AI #MCP¶
Notes on the claims (re-check before posting if time has passed): - "Free for now": PRODUCT.md says everything is free today; the pricing page shows a planned Pro plan marked "not final". - "Call limits": 10 requests per minute per address (docs/guides/MCP_INTEGRATION.md). The number is left out of the post on purpose. - "Beta" and "building it with input from designers" are the owner's statements, not checked against the repo.
First comment (paste both lines; the emoji is on the label, not on the command, so copying the command stays clean):
Posting the full question (with the table request) as a second comment is optional; it is at the bottom of this file.
ALTERNATIVE post text, with the video (only if you change your mind; matches the stackup video)¶
Ask an AI assistant to design a PCB stackup, then ask again. Seven answers from Claude without any design tools gave seven different stackups for the same targets.
I built MagicON, a stackup solver the assistant can call. Give it the same targets and it gives the same stackup: the test ran four times, and MagicON gave the same answer all four times.
It runs as an MCP server (the standard way an AI assistant calls outside tools). No account, no API key.
In Claude Code: claude mcp add --transport http magicon https://api.pcbgenerator.com/mcp/
Then ask the question from the video: "I'm designing a 6-layer PCB for a 10 GHz RF board. Finished thickness 62 mil, single lamination (not HDI), lead-free assembly. I need 50 ohm single-ended microstrip on the outer layer, balancing cost against performance. Recommend a specific stackup."
The video shows the same question answered with and without it: [attach video]
Would you trust a layer recipe your assistant can't reproduce?
Claims come from docs/scripts/MCP_AB_TEST_DEMO_VIDEO_SCRIPT.md and docs/reports/mcp-stackup-solver-ab-test.md
(7 answers, 4 identical runs), and docs/guides/MCP_INTEGRATION.md (server, no sign-in).
Re-checked 2026-10-05: the solver then still returned the video's top stackup
(MEGTRON6 0.146 mm core, VT47 0.56 mm prepreg, 0.3196 mm trace).
β Corrected 2026-10-12 (issue 495): that board is no longer the solver's answer, and its prepreg was
never a real sheet. 0.56 mm is on VT47's laminate (core) list, not its prepreg sheet list
(frontend/src/assets/pcbMaterials.json). Since 2026-10-07 (issue 222) the solver builds each prepreg
layer from 1β3 of the maker's listed sheets and defaults to foil, so the same question now returns a
different board (fingerprint 4b2fb5329399423e, scripts/testing/mcp_ab/solver_fingerprint.py). The
video was re-made from a fresh 4-run test on that code: see docs/reports/mcp-stackup-solver-ab-test.md Β§3.4.
The 2026-10-05 LinkedIn post itself names no stackup figure (text + the 7-answers image).
Rule (owner, 2026-10-05): do not publish the catalog size in articles, posts or listings.
Full question from the video (optional second comment, if you want readers to copy it)¶
Word for word from slide 3 of the video (docs/scripts/mcp-ab-test-question.png):
I'm designing a 6-layer PCB for a 10 GHz RF board. Finished thickness 62 mil, single lamination (not HDI), lead-free assembly. I need 50 ohm single-ended microstrip on the outer layer, balancing cost against performance. Recommend a specific stackup. Give me a table with: core material and thickness, prepreg material and thickness, copper weight and thickness, total finished thickness, the outer microstrip trace width, and the dielectric constant you used.
The post line above stops after "Recommend a specific stackup." to keep the post short; the table request is in the comment.
Weekly puzzle: "Spot the AI's mistake" (owner decision 2026-10-07, issue #227)¶
An EXTRA post each week, next to the Monday article (it does not replace it): the puzzle midweek,
the answer key on Friday. Each puzzle has its own draft, spot-the-ai-mistake-NN.md in this folder.
| No. | Draft | Topic | Status |
|---|---|---|---|
| 1 | spot-the-ai-mistake-01.md | Bare Claude WITH web search on (22 searches, 34 page fetches), asked the 0 dBm drone question, powered ADI's HMC7357 amplifier from +8 V and +5 V only; ADI's datasheet (v01.0918) says its gate needs β2 to β0.4 V. Two other tries with the same amplifier named that supply | Ready, not posted. On hold (owner, 2026-10-08): no puzzle goes out until the fair chat-kit runs are audited; the owner judged this one a quiet miss |
| 4 | spot-the-ai-mistake-04.md | A regular AI chat WITH web search on rated TTM's 1M803 hybrid coupler at "60 W average"; TTM's datasheet (Rev H) says 20 W. It got the coupler's other three numbers right | Ready, for a later week (was #1; its question predates the 0 dBm line). Also on hold until the chat-kit runs |
| β | not drafted | Three trap questions, 2026-10-07, each asking for FIVE parts that meet a rule anyone can check (scripts/testing/mcp_ab/prompt_trap_{pa_3v3_5800,circulator_20w_9400,lna_2x2_10ghz}.txt); bare Claude, web on, 3 tries each: C:/Users/Public/magicon-bare/2026-10-07_215401, _215404, _215407 (23β63 web calls per try, $15.32 plan usage in all) |
No visible mistake. All 3 PA answers refused to pad ("Rather than pad the list with guessesβ¦") and showed the closest parts; the circulator answers said which of their five they could confirm; the LNA answers met every stated rule. Only slip: one LNA answer gives the ADL8140's band as "10β12 GHz" in its table and "10β18 GHz" in its notes. Figures were NOT checked against datasheets |
| β | not drafted | Fair runs, 2026-10-08: 3 claude.ai chats run by Claude Code in the owner's Chrome (owner's permission; β not again: automating claude.ai breaks Anthropic's consumer terms, decision 2026-09-19, so the owner runs the rest by hand): Opus 5.5 "Extra", incognito ("No history, memory, or Claude training"), web search on, code execution on, connectors MagicON AI / PostHog / Sentry / Claude in Chrome switched off (account-wide; switched back on after). Drone question x2, T/R (0 dBm variant) x1. Saved in C:/Users/hakic/magicon-mcp-ab/chats/puzzle_kit/2026-10-08/ |
One visible slip so far: drone try 1 says "Mini-Circuits' PMA5-83-2W+ runs straight off 12 V"; its datasheet (REV. D p.10): "Caution: Permanent damage to the device will occur if the Power ON/ Power OFF sequences are not followed. Power ON: 1. Set VG1 = -2 Vβ¦ 2. Set VG2 = +5 Vβ¦ 3. Set VDD = +12 V". About a part it did NOT pick. β Weak: the same sheet's p.1 says the part operates "from a +12 V power supply", so the AI repeated the maker's own headline, and the same answer DID plan the negative gate rail and gate-before-drain order for the PA it picked (HMC7357). Drone try 2: no proven mistake. T/R: no figure wrong by a lot; it leaves out the TGA2624-SM PA's negative gate, 28 V drain and power-up order (datasheet Rev E p.13: "Apply -5.0 V to VG (for pinch-off) / Increase VD to +28 V") β an omission, the same kind as the HMC7357 gate supply (#1). Sums engine-checked (MagicON run_link_budget, local server, each answer's own stage figures, 2026-10-08): drone try 1 β 30.6 dBm out, 7.9 / 3.3 dB under P1dB, OIP3 40.54 dBm (answer: 30.6, 7.9, 3.3, "about +40.5"); try 2 β 30.02 dBm, 18.36 / 3.86 dB, OIP3 40.77 dBm with the attenuator's IP3 left out (answer: 30.02, 18.4, 3.9, "about +40.8"). All match. T/R: one-step sums plus a PA output the AI read off a curve itself; not engine-checked |
| β | not drafted | Fair runs, second batch, 2026-10-08: 3 NORMAL claude.ai chats pasted by the owner by hand (Opus 5.5 "Extra", memory paused, web search on, not Research; MagicON off confirmed by the owner for drone try 3, NOT restated for the two T/R chats, and not seen on the page for any): drone try 3, T/R tries 2 and 3. A code step ran in drone try 3 and T/R try 2 ("ran a command"); T/R try 3's paste does not show one (not verified). Saved in C:/Users/hakic/magicon-mcp-ab/chats/puzzle_kit/2026-10-08/ (q1_try_3.md, q2_try_2.md, q2_try_3.md) |
No provable mistake in any of the 3. Drone try 3: its stage figures rebuilt in MagicON run_link_budget match (30.12 dBm out, 6.85 / 2.10 dB under P1dB, OIP3 43.39 dBm vs "43.4"); its HMC1119 and HMC8411 figures match our datasheets. T/R try 2: TGA2624-SM, TGL2209-SM, MAAL-010528 and ADRF5020 figures match our datasheets. T/R try 3: QPC2040, QPA1022 and TGL2209-SM match our datasheets, QPA2609 matches our catalog record; its leakage table re-derived by hand is consistent. Leakage and isolation sums are NOT engine-checked (no MagicON tool takes arbitrary parts). Not verified: parts with no sheet here (CML MMA-445933H-M5 and the Mini-Circuits filters in drone try 3, TTM 72831-G001, QPA1010, YAT-7A+), and the "HMC408 not recommended for new designs" and "CMD278C4 discontinued" claims. Picked-part status (drone try 3): Mini-Circuits BFCN-5750+ and LFCN-6400+ show stock and no discontinuation notice, CML MMA-445933H-M5 is listed at DigiKey, the two ADI pages timed out. Verdict: all 6 fair chats gave no postable puzzle; no puzzle this week. The three earlier incognito chats had no code tool (see the rules below) |
| β | not drafted | Trap chat kit, 2026-10-08, run 2026-10-09 by the owner by hand: 5 plain student questions on the five design traps of built exercises #3β#7 (circulator reflection, a limiter that never clamps, blind range, PA idle heat, undersampling clock), scripts/testing/mcp_ab/prompt_trap_{bias_heat_9400,undersampling_clock,limiter_clamp_9400,circulator_leakage_9400,blind_range_9400}.txt. Kit with paste steps on the owner's computer: C:/Users/hakic/Documents/MagicON-posts/puzzle-chat-kit-2/ (one NORMAL claude.ai chat per question, Max, web + code on, MagicON off, run by the owner by hand) |
No AI mistake in any of the 5 (one try each; Opus 5.5 "Extra", web and code on, MagicON and chat search off as stated for question 1, not restated for 2-5). Each answer caught its trap, and its figures match MagicON's results and our datasheets. So the built exercises stay the posts for these topics. Saved in C:/Users/hakic/magicon-mcp-ab/chats/puzzle_kit/2026-10-09/trap_q{1..5}_try_1.md with an audit at the top of each. Question 4 found two things in OUR data instead: the circulator JCSM0800T1200S20 is filed as 8β12 GHz but its JQL name reads 800β1200 MHz (issue 361; built #3 names it, so #3 is on hold), and MagicON's leakage takes the larger path where the AI also adds both in phase (issue 362) |
| β | not drafted | Part-fact trap kit, 2026-10-09, run the same day by the owner by hand: 4 plain questions on datasheet FINE PRINT, each checkable against our own copy: CGHV1F006S output power is pulsed-only (100 us, 10%), PMA5-83-2W+ needs VG1 = β2 V before +12 V (REV. D p.10), QPA2811's 0.680 Β°C/W ΞΈJC is a pulsed figure (1.038 Β°C/W is the no-RF one), TGL2209-SM is 36 dBm CW at 85 Β°C vs 40 dBm at 25 Β°C. scripts/testing/mcp_ab/prompt_trap_{cghv1f006s_cw,pma5_83_rails,qpa2811_cw_theta,tgl2209_hot_cw}.txt; paste steps in C:/Users/hakic/Documents/MagicON-posts/puzzle-chat-kit-3/ |
No AI mistake in any of the 4 (one try each; settings not restated per chat). q1, q2 and q4 caught their traps and every quoted figure matches our datasheet copies; q1 also named a real datasheet contradiction (CGHV1F006S AMP3 '28 V' on p.21 vs 40 V on p.1, already noted in our catalog record, line 40). q3 did NOT test its trap: the chat could not read Qorvo's PDF and refused to guess ΞΈJC (a stray second turn is in that chat). Saved in C:/Users/hakic/magicon-mcp-ab/chats/puzzle_kit/2026-10-09/fact_q{1..4}_try_1.md with audits. Total across both kits: 9 fair chats, 0 AI mistakes |
| β | not drafted | Traps from our own past fixes, 2026-10-08: Qorvo's store lists the QPA1019 as "P1dB: 40 dBm" though its datasheet (Rev. A p.2) gives only PSAT 39 / 41 dBm (a Claude Sonnet 5.5 session copied the store figure into our catalog, commit a80dced8a, issue #297); Finwave's FW2002 is rated 15 W CW at 1 GHz and derated at 10 GHz (issue #298). Questions scripts/testing/mcp_ab/prompt_trap_{qpa1019_p1db,fw2002_power}.txt; bare Claude, web on, 3 tries each: C:/Users/Public/magicon-bare/2026-10-08_105655, _105658 ($2.28 plan usage in all) |
All six avoided the trap. QPA1019: "I don't trust it⦠It equals the saturation spec", "almost certainly the saturated output power (Psat)". FW2002: "The rating is for 1 GHz, not 10 GHz", "no, not with margin" |
| β | not drafted | Graph reading, 2026-10-08: a plain chart of a chain's levels drawn from MagicON run_link_budget (0 β 13.6 β 10.9 β 33.3 β 32.4 dBm, PA 1.3 dB under its P1dB line; 1 dB gridlines, no data labels; image C:/Users/Public/magicon-bare/charts/chain-levels.png), question scripts/testing/mcp_ab/prompt_graph_chain_levels.txt; bare Claude with Read + web, 3 tries: C:/Users/Public/magicon-bare/2026-10-08_102717 ($0.25 plan usage in all) |
All three read every value exactly (13.6, 10.9, 33.3, 32.4, P1dB 34.6, back-off 1.3 dB), each after one Read of the image |
| β | not drafted | Three wrong-premise questions, 2026-10-08, each stating one false figure while asking about something else: a 0.3 mm "50 ohm" trace on 2-layer 1.6 mm FR-4, a chain "0 + 20 β 3 + 15 = 35 dBm", and "Part 15.247 lets me run 10 W" (scripts/testing/mcp_ab/prompt_premise_{trace_width,chain_sum,fcc_power}.txt); bare Claude, web on, 1 try each: C:/Users/Public/magicon-bare/2026-10-08_101827, _101830, _101833 ($0.61 plan usage in all) |
All three caught it in their first line: "A 0.3 mm trace on your stackup isn't 50 Ξ©. Fix that first", "First, the chain gives 32 dBm, not 35 dBm", "I can't recommend a 10 W amplifier for this, because the premise is wrong." Their own figures (e.g. "about 130 Ξ©") were NOT engine-checked |
| β | not drafted | Rule-break search, 2026-10-07: every saved bare, web-on answer (~77) read for a break of a rule written in its own question | No loud break. Nearest: a T/R chat calls its circulator "a drop-in part: it bolts through the boardβ¦ so it's not connectorized" (arguable); an attenuator answer promises "20 dB" while its table says "19 dB Β±1.5 dB" (partly admitted) |
| β | not drafted | Maths check, 2026-10-07: every link-budget table in the saved bare-Claude answers (20 claude.ai answers with MagicON off, plus the 10 test-kit answers to the 0 dBm question) rebuilt in MagicON's run_link_budget from the answers' own figures |
No maths mistake. Every level, headroom and OIP3 matched within the thresholds; the AI's slips are in part facts and omissions, not sums |
| 2 | spot-the-ai-mistake-02.md | Asked for a 10 GHz LNA, the AI offered MACOM's MAAL-011078 as a "2β18 GHz" part; its datasheet says 700 MHz β 6 GHz. Without web access, 2 of 3 tries named it; with web search on, 0 of 3 | Held: does not survive web search |
| 3 | spot-the-ai-mistake-03.md | An AI with no tools typed 620 pretend commands, then said its "script" gave no output | Held: no tools or web, and not an electrical mistake |
| β | not drafted | Nine web-on runs aimed at power ratings (2026-10-07, C:/Users/Public/magicon-bare/): a 40 W hybrid combiner (_183743), an LNA that survives 25 dBm (_183803), a 5 W attenuator pad (_184535); questions scripts/testing/mcp_ab/prompt_{hybrid_40w_5800,lna_survive_9400,pad_5w_5800}.txt |
No clear mistake seen. Every answer looked the parts up (22β34 web calls, 0 refused) and said where it could not read a PDF; all 3 hybrid answers rated the 1M803S at 20 W, matching TTM's datasheet. The answers were read for obvious wrong picks only; their other ratings were NOT checked against datasheets |
| β | not drafted | The 28 GHz question scripts/testing/mcp_ab/prompt_stackup_28ghz.txt |
No provable mistake, so no puzzle from this run. Run C:/Users/Public/magicon-bare/2026-10-07_170904 (Opus 5.5 xhigh, no tools, 3 tries): no pretend tool calls; every answer said its figures came from memory and should be checked. Two answers said Isola lists Astra MT77 "flat across ~2β20 GHz", but our copy of the sheet (03/17) lists 2 and 10 GHz only. Both had already warned they were least sure of exactly that, so it is not a fair puzzle. One answer's 13.3 mil RO4350B core is not on Rogers' standard list, but PUB# 92-004 p4 offers non-standard thicknesses from 0.0066" to 0.060", so it is not provably wrong. |
Rules (learned on puzzle 1):
- A plain question a student would ask. Show the exact question and the AI's exact answer, word for
word; mark any cut with "β¦".
- Bare Claude only (owner, 2026-10-07): the answer must come from an AI with MagicON switched
off, either a claude.ai chat or the clean test kit (C:/Users/Public/magicon-bare/). A chat
where MagicON was connected, even if every call to it was declined, or where "off" was not
confirmed, does not count.
- A fair setup is what a real user has (owner, 2026-10-08): web search PLUS a code tool, at the
top effort setting ("Extra" or "Max"). The saved claude.ai chats used bash_tool when they
needed sums (4 of 4 grounding chats, 4 of 4 PDN, 3 of 5 compliance; their settings.md). The
test-kit runs up to 2026-10-08 offered web search only, no code tool, a HARDER setup than
claude.ai, so their "no visible mistake" results stand, but a puzzle must not rely on the AI
lacking a calculator. Giving the test kit a code tool needs a sandbox first: it would run real
commands on the owner's PC. The fair runs are prepared as a claude.ai chat kit (owner's
computer, outside the repo): C:/Users/hakic/Documents/MagicON-posts/puzzle-chat-kit/ (2
full-design questions x 3 chats at Max effort, web and code on; questions
prompt_link_budget_designed.txt and prompt_tr_module_0dbm.txt). All 6 chats ran on
2026-10-08 (table above). β Claude does not drive claude.ai: Anthropic's consumer terms forbid
"automated or non-human means" (decision 2026-09-19).
- β Incognito chats cannot run code (Claude help page "Use incognito chats": "Claude can't
create files or run code in them"). The first 3 fair chats were incognito, so they had NO code
tool and do not meet the fair-setup rule above; only the 3 normal-chat runs do (memory paused,
web search on, a code step ran in 2 of them; MagicON off as stated by the owner, see the row). Use a NORMAL new chat for a fair run.
- A reader must see the mistake from the images alone (owner, 2026-10-07): the first image puts
the AI's exact words next to the maker's own datasheet words. A mistake readers can only find
by looking up datasheets themselves is not a puzzle.
- A transmit-chain question states its input power (0 dBm), as the A/B questions have since
9e90befd0 (owner, 2026-10-07).
- Only a real mistake, proven from the run's own log. Never invent or rig one. If a week's runs show
no provable mistake, that week has no puzzle. Do not post a weaker one to fill the slot.
- Check posted-log.md first; do not reuse a posted topic.
- Do not name the model in the post. Keep the run folder; the proof lives there.
- After posting, log BOTH posts (puzzle and answer) in posted-log.md, or comment their links on #71.
Built exercises (owner decision 2026-10-08, issue #227: "mix both, labelled"): - Post a real AI mistake when one is found. In a week without one, post a BUILT exercise instead, titled "Spot the mistake in this design", never "the AI's mistake". The rule above ("Never invent or rig one") still holds for every post that names an AI. - Say on the image and in the text that the design is a made-up example. - Start from a chain designed by a MagicON tool (drone transmitter, receiver or T/R module); every figure in the correct version comes from that tool result. Change exactly ONE thing on purpose, and record in the draft what was changed and what proves it wrong: the maker's datasheet line, or a MagicON tool result for the changed design (owner, 2026-10-08, so a puzzle can test design rules such as stage order, not only part facts). - A reader must be able to spot it from the image alone, as for the real puzzles. The Friday answer shows the correct version next to that proof. Keep the tool result with the draft. - Aim above "check the frequency range" (owner, 2026-10-08: most students already do that).
Before you post (2 minutes)¶
- Open
https://api.pcbgenerator.com/mcp/once and confirm it still answers (hosting guard fix PR #127 is reported live; not re-tested here). - Only if you use the video: watch it first and skip any claim it makes that you cannot back.
- Rate limit is 10 requests a minute per address, so a spike of readers sharing one office address can hit 429. That is normal, not an outage.
After you post¶
Note the date and link in #71 so the next block knows what has been used.