


Case Study · Updated November 17, 2026
ClaudeBot
Claude as a design and engineering agent.
The entire project, defined and executed purely from language prompt, with images and text as feedback to the LLM.
All assemblies and parts produced in Autodesk Fusion with no manual 3D modeling, are then topology optimized in nTop with no human interaction in the GUI, checked in Fusion for assembly interference by Claude, sliced in Bambu Studio and printed in PETG, renders and animation in Blender.
All of that software is driven by Claude Code from plain-language prompts, on a Windows Dell Laptop with Intel Core i7-12700H, 14 cores, 16 GB of RAM and a NVIDIA GeForce RTX 3050 Ti Laptop, 4 GB reached mostly by remote desktop from a Macbook Air ot iPhone with zero manual 3D modeling.
For context, I have a background in industrial design but not mechanical engineering, simulation or robotics.
I have worked in the past for both Autodesk and nTop, so I do know what the software is capable of, though not necessarily how to use it to the level of sophistication required to execute this project.
nTop was provided as a full license, Fusion is running as a 30 day trial.
Robot programming and testing fully executed with Claude.
This page is the working record of the process and results. After the setup and a summary of what went wrong and what it cost, the process runs in chronological order, from the first design to the latest build. It includes the parts that broke, the instructions that were unclear, and what each day cost in tokens.
Status. The v9 redesign, where every link is one optimized piece, is printed in black PETG and fully assembled on its five servos.
The servo driver board burnt out at the first power-on on November 14. A replacement was fitted and powered on November 16 with no fault. The turntable tips on the yaw servo under the weight of the arm. A v10 yaw joint on two ball bearings is designed and not yet printed. The earlier v7 arm ran on all five joints.
So far. 15 days, 10 design versions, 5 servos, 3.2 million output tokens.
The Setup

Method. A person describes the required result, originally prompt was “can you design a robot arm that I can 3d print the hoousing from easy to get electronics and motors online”. From there it was a conversation with Claude with reviews screenshots and prints, and reporting what is going right and wrong, with a photo, video or drawing on a screenshot. Claude writes the scripts, runs the jobs, places results in Fusion, loads the slicer and keeps the issue log. The person presses Print, assembles and tests.
Software. Fusion for the interface geometry and assembly checks. nTop, run headless from the command line, for topology optimization. Python with OpenCascade and manifold3d to restore exact interfaces after optimization. Bambu Studio for slicing. Blender for renders and motion studies.
Purpose. The project is a test of Claude as a design agent: how far plain-language direction can carry a real assembly from definition through optimization, printing and bench testing. The arm is the test piece. Chess is one of the tasks it should be able to do, along with pen drawing and taught sequences.
Hardware. Five Feetech STS3215 serial bus servos (12 V, 30 kg·cm), a Waveshare bus adapter, an ESP32, a 12 V supply. Upper arm and forearm are 200 mm each. Printed in PETG on a Bambu Lab H2S: grey up to v7, black for v9.
Issue Log
About fifty design issues are logged so far. Most trace to a small number of causes, split between unclear instructions and unchecked assumptions.
| Date | What Happened | Whose | What It Cost and Where It Stands |
|---|---|---|---|
| Nov 3-4 | Optimization jobs ran at full priority on the only workstation. | Claude | Repeated remote lockouts. |
| Nov 7 | Additional scope creep, reach needed for chess, one of the tasks set for the arm, was stated after the first full set had printed. | Human | One part print and a new round of runs. |
| Nov 8 | The base was specified by sketch, then by adjective, then by octopus. | Human | Five optimization runs. The printed base is still the second one (octopus was ugly). |
| Nov 8-9 | Features thinner than the minimum feature size were left to the optimizer. | Claude | Floating pieces in v5, a 3 mm bridge in v6, a 0.9 mm hair in v7. |
| Nov 9 | The arm was designed around a 35 mm box, not the real servo. No interface was checked against the part. | Claude | The whole v6 print set. Thirteen of the first twenty issues. |
| Nov 9 | Design started before the servos had arrived, from a brief written in another conversation. | Human | Nothing was measured against a real servo until day seven. |
| Nov 9 | An assembly instruction was given backward and a part snapped. | Claude | One printed link. |
| Nov 9 | A file was named “final”. | Claude | Nothing has been called final since. |
| Nov 9 | The servo test ran with the power supply unplugged. | Human | A few minutes, it was late… |
| Nov 9 | Tall posts lose their supports mid-print. | Claude | Mid. Open. |
| Nov 11 | The yaw servo is not fixed to the printed base. | Claude | Critical. Looks resolved in the v9 assembly, power test pending. |
| Nov 11 | The turntable has no bearing and rides on the servo horn alone. | Claude | Critical at v7. Only partly resolved in v9: a printed bearing was added, see November 16. |
| Nov 11 | The jaw cannot close on the fixed finger. | Claude | Critical. Looks resolved in the v9 assembly, power test pending. |
| Nov 11 | The yaw servo runs hot holding still. | Claude | Mid. Open, to be rechecked on Base v9. |
| Nov 11 | Cable routing was in neither the brief nor the design. | Human and Claude | Mid. Still open. Leads are taped on. |
| Nov 11 | The aesthetic intent (one piece, organic, optimizer-shaped) was first written down on day nine. | Human | All of v8, built and discarded in one evening. |
| Nov 11 | v8 solved a snapping fork with bolted-on blocks on an arm whose whole point was optimized form. | Claude | The same evening. |
| Nov 12 | A request to radius edges throughout the model meant hard edges only. | Human and Claude | A cut jaw strut and four tries at the base. |
| Nov 12 | The tray and bay rims on the base are still sharp. | Claude | Low. Open. |
| Nov 13 | The v9 forks had not been through an assembly. | Claude | Critical. Closed on November 14: all four forks assembled without breakage. |
| Nov 14 | The back legs of the v9 base are thinner than they need to be. | Claude | Mid. Reinforced in the v10 base, not yet printed. |
| Nov 14 | The servo driver board burnt out at power-on with no servo connected. | Unknown | Mid. Cause still unknown. A replacement board was powered on November 16 with no fault. |
| Nov 14 | The servo leads have no path at the fork joints: they leave the connectors along the face the next link folds onto. | Claude | Critical. Open, needs a joint revision. Soft stops protect the leads until then. |
| Nov 14 | Every fork joint rubs slightly where the idler boss passes the cutaway in the cradle. | Claude | Mid. Clearances opened up in the v10 links, not yet printed. |
| Nov 15 | In simulation the turntable is the softest part: its cradle tilts under the overturning load and accounts for about 5 mm of the 12 mm tip deflection at the design loads. | Claude | Mid. Open. A simulation result only. On the built arm the fault is the turntable tipping on the servo, see November 17. |
| Nov 15 | The finished part files carry zero-volume flaps left by the finishing booleans, which broke the simulation mesher. | Claude | Low. Open. Harmless for printing; cleaned copies were used for the simulation. |
| Nov 15 | The grip teeth on the gripper body were filled in on a simulation copy without asking first. | Claude | Five failed solver runs before the cause was found. The gripper body result is an approximation. |
| Nov 15 | The first stress animation drew the deflection at ten times scale, and the jaw sank through the table. | Claude | One re-render at five times scale. |
| Nov 16 | The turntable is held mainly by the output gear of the yaw servo. The printed bearing is plastic rubbing on plastic and only partly helps. | Claude | Critical. First logged as the turntable flexing. Confirmed on November 17 as the turntable tipping on the servo. A v10 yaw joint is designed, not yet printed. |
| Nov 16 | The control software, written in a separate Claude conversation, was working from the old v2 dimensions (120 and 100 mm links), not the printed v9 arm (200 mm links). The design conversation had never handed the v9 geometry over. | Claude | Mid. A day of distorted pen drawings. Link lengths corrected and a geometry sheet written on November 16. Calibration of the joint zeros is still open. |
| Nov 17 | The turntable tips toward the weight of the arm and sits out of square on the yaw servo. Its leading edge rubs on the base. Play measured by hand is about 3 degrees, about 8 mm at the pen at 150 mm reach. | Claude | Critical. Open. The turntable is carried only by the servo horn. v10 puts it on two preloaded ball bearings so the servo only turns it. Designed, not yet printed. |
| Nov 17 | The arm came off the yaw servo spline twice in one day of pen drawing. Nothing held the turntable down. | Claude | Critical. Open. Re-seated by hand each time. v10 adds a ring that holds the turntable down. |
| Nov 17 | The fault was first read as play inside the servo, then a single bearing was proposed that would not have resisted the tipping. | Claude | Two corrections from Human before the redesign started. |
| Nov 17 | Pen drawing tests: pen height is off by 1.5 to 4.5 mm and does not repeat, and sideways strokes under 10 mm vanish. | Claude | Mid. Open. Worked around in software at three times the drawing time. Expected to improve with v10, to be measured again. |
| Nov 17 | The servos cannot tell that the turntable has slipped or come off: position and load read normal. | Claude | Low. Open. v10 has index marks on the yaw joint. |
| Nov 17 | The elbow cannot fold far enough to bring an upright pen closer than 217 mm to the base. | Claude | Low. Open. Worked around by tilting the pen. Not changed in v10. |
| Nov 17 | The elbow was designed, documented and limited in software as folding one way only. On the built arm it folds both ways, and it always should. The requirement was never written down. | Claude | Critical. The design range is now 130 degrees each way, with one corner of the forearm fork trimmed in v10. Not yet printed. The software limit is still to be changed. |
| Nov 17 | The front feet of the base reach 75 mm in front of it, so a chess board or drawing paper under the arm sits on them at an angle. Found in prototype testing. | Claude | Mid. Cut back to match the back feet in the v10 base, not yet printed. |
| Nov 17 | The reinforcements added to the feet of the v10 base stopped short of the body, leaving a gap at all four feet. They had been trimmed after their check and not checked again. The first fix then joined them with sharp corners. | Claude | Critical. Caught by Human in the slicer, twice, before anything was printed. The feet now run into the body through 10 mm fillets. |
| Nov 17 | The simulation treated the ring between turntable and base as frictionless and able to hold the turntable down. The real ring can only push up, and it rubs. | Claude | Low. Open. The simulation could not have shown rubbing or lift-off. The v10 joint has not been simulated yet. |
| Not started | Chess. |
Token Use
Counted from the session log of the main Claude Code conversation, by calendar day in US Eastern time. Output is text and code written by the model. Cache read is the conversation so far being re-read on each of 3,236 requests, which is why it dominates. Related work in other conversations is not counted. Model is the Claude model that did the work in Claude Code that day: Claude Opus 5.5 until November 8, Claude Fable 5.1 from then, and Claude Fable 5.2 from November 11, with some later work still on 5.1. Totals run through November 17 and cover 3,236 requests. The latest day is counted up to the last update of this page.
| Day | Main Work | Model | Output | Cache Write | Cache Read | Uncached Input |
|---|---|---|---|---|---|---|
| Nov 3-4 | First runs | Opus 5.5 | 130K | 901K | 36.8M | 0.5K |
| Nov 5 | v2 redesign | Opus 5.5 | 122K | 1.00M | 34.7M | 0.4K |
| Nov 6 | Cradle fixes, fit tests, order list | Opus 5.5 | 324K | 1.46M | 111.1M | 0.9K |
| Nov 7 | Chess links, base v1 | Opus 5.5 | 259K | 957K | 95.2M | 0.6K |
| Nov 8 | v4 to v6, plate, Blender, hex base | Opus 5.5, then Fable 5.1 | 213K | 1.24M | 73.2M | 5.7K |
| Nov 9 | v6 assembly, joint test, v7, bench test | Fable 5.1 | 448K | 2.40M | 253.4M | 8.6K |
| Nov 10 | Turntable v4, metadata | Fable 5.1 | 60K | 427K | 15.7M | 2.2K |
| Nov 11 | Full assembly, v8, v9 | Fable 5.1, then Fable 5.2 | 358K | 2.38M | 106.2M | 2.2K |
| Nov 12 | v9 review fixes, base radius, this page | Fable 5.2 | 356K | 1.51M | 122.6M | 0.7K |
| Nov 13 | Links printed, base plate, promo video, optimizer setup published | Fable 5.2 and 5.1 | 273K | 1.21M | 146.1M | 5.3K |
| Nov 14 | v9 assembly record, issue log, stress simulation runs | Fable 5.2 | 219K | 1.14M | 70.1M | 0.5K |
| Nov 15 | Stress and deflection animation, turntable and gripper simulation, single page, page menu | Fable 5.2 | 161K | 670K | 62.3M | 0.3K |
| Nov 16 | Replacement driver board powered, turntable support logged, v9 geometry handed to the control software | Fable 5.2 | 90K | 635K | 8.9M | 0.1K |
| Nov 17 | Turntable tipping confirmed, pen drawing findings logged, v10 yaw joint, base and links designed, elbow range corrected | Fable 5.2 | 250K | 704K | 47.8M | 0.38K |
| Total | 3.22M | 16.6M | 1,184.0M | 28.6K |
Timeline
Twelve days from the first design to the assembled v9 arm. The full account of each step is in The Process, further down this page.











The Process + Results
The full record of the build, day by day, and the current parts as 3MF files.
Results: The Current Parts
This is iteration v9: six printed parts for five Feetech STS3215 servos.
Each 3MF file carries its own description, load cases and optimizer settings in its metadata.
The parts were printed in PETG on a Bambu Lab H2S with 4 walls, 30% gyroid infill and tree supports.
Iteration v9 is printed but not tested when these files were posted. Earlier iterations were printed and assembled, and the problems found in them are listed below. Treat the files as a work in progress.
Download All Six Parts (ZIP, 5.6 MB)
| Part | What It Does | File | Size |
|---|---|---|---|
| Base | Holds the yaw servo and the turntable track | Base_261112_v9e.3mf | 2.3 MB |
| Turntable | Rides the track and holds the shoulder servo | Turntable_261112_v5.3mf | 0.6 MB |
| Upper Arm | 200 mm between joints | Upper_arm_261112_v9f.3mf | 0.8 MB |
| Forearm | 200 mm between joints | Forearm_261112_v9f.3mf | 0.8 MB |
| Gripper Body | Holds the jaw servo | Gripper_body_261112_v9f.3mf | 0.8 MB |
| Jaw | The moving jaw | Jaw_261112_v9f.3mf | 0.4 MB |
The Parts in 3D
The six printed parts of iteration v9, shown from the 3MF files themselves. Pick a part, or view them posed as assembled. The servos are not shown. This iteration has since been printed and assembled.
This 3D view is a derivative of the 3MF Viewer from the 3MF Consortium. Open This Part in the 3MF Viewer (New Tab)
How the Shape Is Decided
Every structural part of ClaudeBot is the output of a simulation, not a sketch.
The optimizer is given a volume it may use, the regions it must keep, the regions it must avoid, a material, the restraints and the loads. It meshes the volume, solves the load cases, removes the material that carries the least strain energy, and solves again, up to a hundred times.
The part that comes out is the shape that best carries those loads with the material allowed.
That makes the process deterministic. The same inputs give the same part. If a part changes between iterations, it is because an input changed: a load, a keep-out, a volume fraction.
Claude never drew a strut or a hole.
The design work went into defining the problem, in Fusion and in Python, and the geometry followed from it. Form follows function.
The Setup in Fusion
The problem is defined in Fusion before nTop sees it.
The parametric model holds what the optimizer is not allowed to change: the servo cradles with their 2 mm rib relief, the pad plates with the M3 horn pattern, the bearing rings, the feet and the wire ways.
Each link is a coloured placeholder jointed to its servo, so the joint axes, the link lengths and the swing of each joint are fixed here.
From this model a script writes the coloured STEP that nTop reads: grey for the volume it may use, green for what it must keep, red for what it must avoid, and face colours for the restraints and loads.



The Setup in nTop
Screenshots of the upper arm setup as example, opened in nTop with the optimizer paused.
Each notebook is generated by a script from a coloured STEP and a load table, so these views exist for every part.








What Each Part Was Given
| Part | Design Space | Kept | Load Cases | Volume Fraction | Min Feature | Elements | Iterations | nTop Blocks |
|---|---|---|---|---|---|---|---|---|
| Base | 848 cm³ | 213 cm³ (25%) | 9 | 0.26 | 3.5 mm | 2.5 mm | 50 | 289 |
| Upper arm | 326 cm³ | 119 cm³ (36%) | 4 | 0.36 | 4 mm | 2 mm | 100 | 227 |
| Forearm | 386 cm³ | 121 cm³ (31%) | 4 | 0.32 | 4 mm | 2 mm | 100 | 224 |
| Gripper body | 141 cm³ | 63 cm³ (45%) | 3 | 0.48 | 4 mm | 2 mm | 100 | 229 |
| Jaw | 48 cm³ | 23 cm³ (48%) | 3 | 0.55 | 4 mm | 2 mm | 100 | 207 |
Design space is the volume after the keep-outs are removed. Kept is the finished part. The volume fraction counts the keeps, so the free region is thinner than the number suggests. The material is PETG at 2 GPa. Every link also carries a 25 N fork-spread case, so the one-piece fork is sized to be pushed over the servo.
The Setup Files
The complete setup is published so it can be checked, rerun or changed.
for each part, the design space volumes, the restraints, every load case as numbers, the optimizer settings, and the full nTop block graph with every block, its wiring and its values. About 230 blocks per part.
nTop notebooks and coloured STEP files (ZIP, 1.6 MB)
the five .ntop files that produced the printed parts, with the STEP each one reads. Open in nTop 6.3 with the STEP in the same folder. The optimization runs on open; a 200 mm link takes about an hour.
Both are under the same CC BY 4.0 license as the parts.
Build Your Own
Everything needed to replicate or modify the arm. The part files are above.
Software
| Tool | Used For |
|---|---|
| 3MF | The file format for every part, with the description, load cases and optimizer settings stored in the file |
| nTop | Topology optimization of every structural part, and the stress checks |
| Autodesk Fusion | Interface geometry, joints and the assembly |
| Blender | Renders and animation |
Parts
| Part | Quantity | Link |
|---|---|---|
| Feetech STS3215 serial bus servo, 12 V, 30 kg·cm, 2-pack | 3 packs (5 used, 1 spare) | Amazon |
| Serial bus servo driver board for ST/SC servos | 1 | Amazon |
| ELEGOO ESP-32 development boards, USB-C, 3-pack | 1 pack (1 used) | Amazon |
| SHNITPWR 72 W adjustable power supply, 3.5-12 V, 6 A | 1 | Amazon |
| SELOKY LM2596 buck converter module, 12 V to 5 V, 2-pack | 1 pack (1 used) | Amazon |
| 5264 3-pin servo cables, 3 x 300 mm and 3 x 900 mm | 1 pack | Amazon |
| CENTROPOWER 5.5 x 2.1 mm DC barrel connectors, 10 pairs | 1 pack | Amazon |
| Rubycon 1000 µF 16 V low-ESR capacitors, pack of 15 | 1 pack | Amazon search |
| Dupont jumper wire kit, 120 pieces, 20 cm | 1 kit | Amazon search |
| M2 socket cap screw assortment, 1260 pieces (M2 x 8 and M2 x 10 hold the servos) | 1 kit | Amazon search |
| M3 to M6 stainless socket cap screw kit, 700 pieces (M3 x 6 to M3 x 12 for horns and idlers) | 1 kit | Amazon |
| ELEGOO Rapid PETG filament, 1.75 mm, black, 2 kg | 1 | Amazon |
All parts were bought on Amazon. Nine rows link to the listing that was bought, or the closest match found for it. Three rows are Amazon search links, where the exact listing could not be found. None of the links are affiliate links.
License
ClaudeBot is open source under Creative Commons Attribution 4.0 International (CC BY 4.0). Copy it, change it and build on it, including commercially. Credit “ClaudeBot by CDFAM” and link to cdfam.com/case-studies/claudebot. The license covers the printed part designs and this write-up. The servos, electronics and software tools belong to their makers. Everything is provided as is, with no warranty.
The Process
Before November 3, 2026
Where the Parts Were Defined
The arm began on a Mac, in a separate Claude Code project, as a Python script. A first version was drawn around MG996R hobby servos and is kept for reference. Version 2 moved to five Feetech STS3215 bus servos and is fully parametric: one generator script writes every printed part, with the servo dimensions held as fixed constants and the clearances as adjustable parameters.



The joint was defined there, in the style of the SO-101 arm. Each link bolts straight to the servo’s tapped 20 mm output face with four M2.5 screws and rides the rear 20 mm boss in a 20.6 mm printed bore. No horns. The servo dimensions were taken from the manufacturer’s CAD, down to a case bottom that is not flat: a 2 mm rib and a 1.6 mm pedestal, each needing its own relief in the cradle.
The working rules were set there too. Every geometry change is regenerated and checked by boolean interference against the servo solid, at working and extreme poses. Parts must be single, watertight solids. A fit coupon prints first. Print files are 3MF, never STL. Verified travel at that stage: shoulder ±90°, elbow ±88°, wrist ±85°.
The topology optimization was specified in the same project, as a package of design spaces, keep bodies and loads in STEP, and handed to a second conversation on the CAD machine. That handoff is the point where the November 3 entry below picks up. The first optimized parts came back on November 5 and needed three interface trims, all traced to errors in the design spaces. The joint defined here expects M2.5 screws straight into the servo. The servos that arrived carried an M3 horn, which is what the November 9 assembly ran into.
The same project holds the control side: ESP32 firmware that bridges Wi-Fi to the servo bus, and a Python stack with one server that owns the bus and apps for sequencing, teaching poses, chess and pen drawing. Lessons recorded there so far: send one profiled move and let the servos run the ramp, because waypoints streamed over Wi-Fi come out jerky; a stalled servo is the hottest case, and two overheat alarms came from the jaw creeping into its stop under power; keep each joint’s zero near the middle of the encoder range. Cartesian motion, and with it chess, is waiting on calibration.
Tokens: not counted. This work was done in a separate conversation.
November 3-4, 2026
First Runs
The first headless nTop runs on the upper arm. An early instruction was a correction: stop the fine-resolution runs and keep adjusting the coarse 3 mm one until it looks like it might work.

Settled in these two days: 1 mm mesh tolerance is enough for an FDM part, and an optimization at full priority can leave the computer unreachable by remote desktop.
Tokens, November 3-4: 130K output · 901K cache write · 36.8M cache read · 0.5K uncached input
November 5, 2026
A Redesign From a Second Brief
The v2 design brief was written in a separate Claude conversation and handed over as a package. That conversation did not know what the first runs had taught about meshing and smoothing. This one did, and it did not know why the geometry was the way it was.
Runs settled into a pattern that held for the rest of the project: a coarse gut-check at 3 mm elements, a look in the nTop GUI, then a 1.5 mm refinement.
Tokens, November 5: 122K output · 1.00M cache write · 34.7M cache read · 0.4K uncached input
November 6, 2026
Fit Tests, and the Cables Nobody Mentioned
The keep bodies from Fusion had a notch in each cradle end wall, where two edges met at zero thickness, and a clash at the wrist. A manual patch in nTop was tried first and looked bad. The Fusion source was then corrected and fit-test coupons were printed in PLA.

That evening, with the servos still in the post, came the question of whether the design had any allowance for wiring. It did not. The brief had not asked for cable routing and Claude had not raised it. Both halves of that are in the Issue Log above.
A quieter finding: nTop counts the fixed interface bodies in its volume fraction, so the first runs starved the free material. The fraction is now set per link.
Tokens, November 6: 324K output · 1.46M cache write · 111.1M cache read · 0.9K uncached input
November 7, 2026
A Chess Board Sets the Reach
The first full set of links printed in the morning. Then a reach requirement arrived. The reach looked small, and five minutes later the first concrete task was named: the arm should be able to play chess.
The links were stretched to 200 mm to reach a full board, the elbow fold opened to 150 degrees, and the keep-out volumes swept by each joint were added as plugs. The third run died silently after four hours.


The base request that evening: a low center of gravity, feet on the outside for screws or clamps, concrete ballast inserts, stable through 360 degrees. Then, once the design space was drawn, a note that the arm would not only be used for chess.
Tokens, November 7: 259K output · 957K cache write · 95.2M cache read · 0.6K uncached input
November 8, 2026
A 22-Hour Plate, a Render, and an Octopus
Overnight link runs v4 and v5 failed in two ways. A minimum feature size of 6 mm diverged after iteration 55. Then the 5 mm horn plate and 3.5 mm feet came out as floating pieces, because the optimizer cannot fill a slab thinner than its minimum feature. One disconnected strut was fixed by hand in Meshmixer and sent back. The verdict from the bench that morning was that the process had been really messy.
A file naming rule followed (Part_YYMMDD_vN), the v6 links converged, and five parts went onto one plate with dates printed into them.


The base brief changed three times in one afternoon. First: a hexagonal, symmetrical base, more organic in form. Then six ballast pockets. Then: it can be chunky, it does not need to be light, it should look like an octopus holding onto a rock. Four optimization runs produced loose pieces or flat webs. Topology optimization was dropped for that base and an octopus was lofted by hand. It has not been printed. The arm still stands on Base v2.
nTop also locked its owner out of the computer again by taking every core. Jobs now run at below-normal priority.
Tokens, November 8: 213K output · 1.24M cache write · 73.2M cache read · 5.7K uncached input
November 9, 2026
The v6 Assembly: No Joint Could Be Assembled
The 22-hour plate finished at 09:09. By mid-afternoon it was clear that not one servo joint could be assembled without rework. Every problem traced to one cause: the arm had been designed around a 35 mm box standing in for the servo, and no interface had been checked against the real part.




The list: horn screw bores too small; counterbores printed closed on the bed face; four side screw holes aimed into the servo body, for a mounting flange that does not exist; a fork that could not slide over the horn; a bearing bridge left to the optimizer as a single 3 mm strut. A part also snapped during assembly after Claude gave the fitting direction backward.


The interface test should have been the first print of the project. It came after six design versions. Once it passed, the four links were rebuilt on the tested joint and re-run as v7 the same afternoon. A new finishing step unions the exact interface solids back into the optimized mesh and cuts the exact holes, because smoothing at 1 mm had been closing 3.4 mm holes.


Also today: the first v7 upper arm was saved with “final” in its file name. The response was to ask why anyone would tempt fate like that.
Tokens, November 9: 448K output · 2.40M cache write · 253.4M cache read · 8.6K uncached input. The heaviest day, and a day the usage limit was hit mid-task.
November 10, 2026
A Turntable Edited by Hand, a Failed Print, and Metadata
The turntable was imported into Fusion as an editable solid, tidied by hand, exported, and re-checked against the printed base: zero overlap at any yaw angle. The upper arm print failed as spaghetti and was reprinted alongside it.
From this day every 3MF carries a readable metadata block: author, project, date, material, printer, software and, for optimized parts, the loads and restraints. It is plain text on purpose, so that reading it does not require an nTop license.
Tokens, November 10: 60K output · 427K cache write · 15.7M cache read · 2.2K uncached input
November 11, 2026 · Day
First Full Assembly: It Moves, and a New List of Flaws
Base v2, Turntable v4 and the v7 links went together with all five servos on the bus. Every joint moves. The turntable had to be clamped to the base, which makes yaw hard to test, and the yaw servo overheats while holding still.




Logged, by instruction, without redesigning yet: the idler side of every joint adds no stiffness; the fork arms are bare 4 mm plates the optimizer never touched; the yaw servo is not fixed to the base at all, because the script carried flange screw holes for a flange this servo does not have; the turntable rides on the servo horn alone; the base cable slot does not line up with the servo connectors.
November 11, 2026 · Night
v8 Built, Checked, Rejected. v9 Started From the Original Intent
After the first full assembly, the request was a full redesign from every flaw found in the build. Claude’s answer, v8, fixed the snapping fork by splitting it: a separate idler arm bolted on with two M3 screws and captive nuts, and a plain wall under the turntable. It passed every clearance check. It was rejected on sight.
These components are so square and brutal. I would rather risk snapping in the assembly process than have these bolted-on bodies look so bolted on and unrefined. The support under the turntable does not look optimized, it looks rudimentary.
Review of v8, November 11



The design intent had never been written down: every link one piece, organic, shaped by the optimizer, no blocky primitives. Once it was, v9 followed overnight. The forks are grown by the optimizer. The base is modeled resting on the table, so the supports under the turntable track come out of the optimization: 16 to 22% solid under the track, against 4% on the printed base. Wire ways, servo fixing screws and foot screw holes are cut afterward.
Two runs failed on the way. A minimum feature size of 6 mm cannot form fork arms inside 4.5 mm plates. The gripper body lost its fork because the jaw’s keep-out was drawn as a full disc. All links then passed: one piece each, no thin members, zero overlap with the servos, no clashes across 128 chess poses.

Tokens, November 11 (day and night): 358K output · 2.38M cache write · 106.2M cache read · 2.2K uncached input
November 12, 2026
v9 Reviewed, Edges Radiused, Printed Arm Running
The printed arm (v7 links, a plain jaw, Base v2) is clamped to a spare printer bed and moves under bus control. This is the latest physical build, shown in the video at the top of the page.



The overnight v9 model was reviewed in Fusion in the morning. Three things were wrong. The jaw servo was off the jaw axis. The grip holes and tie slots were closed over. The interface blocks had sharp edges where they met the optimized surfaces.



The radius request went wrong first. It asked for edges to be radiused throughout the model. Claude rounded the whole body, which cut through a jaw strut. The request was then narrowed to hard edges only. The second version rounds 1.5 mm only within 2.5 mm of an interface block.
The base took four attempts. The first lost its 2 mm tray walls, because a 1.5 mm ball cannot round a wall thinner than about 3 mm. The second had ragged rims. The third ran for five hours and ran out of memory. The fourth leaves the thin shells sharp and passes the hole, wall and overlap checks at 213.1 cm³.


Issue tiers were defined today, ten days in: Critical (cannot be assembled, cannot carry load, parts interfere), Mid (a feature does not work, such as closed holes), Low (finish, such as edge radii). The v9 links and Turntable v5 are loaded in Bambu Studio. The print settings were then changed by hand to four walls and 30% gyroid, with parts moved toward the bed center where fewer errors occur.
Tokens, November 12 (includes building this page): 210K output · 1.43M cache write · 52.7M cache read · 0.5K uncached input
November 13, 2026
The v9 Links Come Off the Printer
The four v9 links and the turntable printed overnight in black PETG, on one plate, with tree supports. This is the first print of the one-piece forks, where each link wraps both sides of its joint without a bolted-on arm.
In the photos the parts still carry their supports. They have not been cleaned up or fitted to the servos, so nothing is known yet about fit, cable routing or how the forks slide on.


One link cleaned up. The tree supports came away and left the fork, the cradle and the pad plate as designed. The first full set of v7 links, printed in grey on November 9, is shown beside it for comparison: bolted idler arms then, one piece now.

The base was not on this plate. It was reworked on November 12 to radius its hard edges, and at 248 by 246 mm it fills a bed on its own. It is sliced and next to print: 9 hours 19 minutes and 251 g of PETG.
The project also went public on this day: the case study, a 3MF viewer, the part files under an open license, and the parts list.
v9 Assembled, November 14
The full v9 set (base, turntable and four links) printed cleanly in black PETG and went together on five servos on November 14. All four one-piece forks slid on and bolted up with nothing broken, which is the step that snapped a fork on v7. The redesigned jaw looks resolved in the assembly. The yaw turntable looked resolved too, but on November 16 it turned out to need more support on the base (see the Issue Log).
The arm did not move under power that day. The servo driver board burnt out when it was switched on, before any servo was connected. A replacement was fitted and powered on November 16 with no fault.


The issues found during the assembly, in the order they were logged.






Every issue, with the date it was found, whose it was and where it stands, is in the Issue Log near the top of this page.
Latest: Stress and Deflection Simulation, November 15
Each printed v9 part was run as a linear static analysis under its own design load case, and the results were put back into the assembly. The clip shows the arm straight out and horizontal: eight seconds of the whole arm as the load comes on, then a close-up of each part at full load. Color is von Mises stress on a scale of 0 to 8 MPa. Deflection is drawn five times larger than it is, with the unloaded shape in grey.
| Part | Load Case | Peak von Mises | Flex Within the Part |
|---|---|---|---|
| Base | Arm reaching toward the board: 25 N down, 4.5 N·m overturning | 1.0 MPa | 0.01 mm |
| Turntable | The same case as the base, entering through the shoulder servo pocket | 9.1 MPa | 0.55 mm |
| Upper arm | Gravity x2.5 with a x3 margin, plus elbow servo stall | 10.7 MPa | 1.83 mm |
| Forearm | Gravity x2.5 with a x3 margin | 6.9 MPa | 0.77 mm |
| Gripper body | 20 N grip on the fixed finger, plus jaw servo stall | About 5 MPa (approximate) | About 1.2 mm (approximate) |
| Jaw | 20 N closing force on the pad, plus 5 N drag | 11.8 MPa | 1.39 mm |
The peaks are 9 to 12 MPa at loads well above normal use. PETG yields at roughly 45 to 50 MPa by the handbook; that figure has not been measured on these prints. On the links the highest values sit at the edge of the fixed drive pad, where the rigid restraint in the model concentrates stress more than a bolted servo horn would. A joint cannot load a link beyond the stall torque of its servo, 2.94 N·m, so the servos give up before the printed parts do.
The turntable is the softest part in the chain. Its cradle moves 0.55 mm, and that tilt carries along the whole arm: about 12 mm at the finger tip at the design loads, of which about 5 mm comes from the turntable.
This is a per-part analysis with the parts chained rigidly at the joints. It is a long way short of a coupled assembly solve, and the servos are not modeled. Five parts were solved in nTop. The turntable was drawn in CAD and never had a load setup, so it was given the base’s. The gripper body would not mesh in nTop, because the grip teeth on the fixed finger hang on by webs thinner than an element. Its field is an approximate voxel solve written in Python, on a copy with the teeth filled in, and is labeled as approximate in the clip.
No part is close to failing in the simulation. Four items from the day are in the Issue Log: the soft turntable, stray flaps in the part files, and two mistakes in how the simulation and the clip were made. The powered test of the assembled arm is still to come.