📓 Dev Diary · newest firstCross-check · DevelopmentMulti-assistant review · ImprovementSame look as the product: W / M / L skins, dark / light / system theme.
Automatic Cross-Checking
Dev Diary
One assistant, many tasks. Two checking stages keep the work moving and the mistakes from repeating.
Automatic cross-checking · Development
Thin-slice ticket, no more than 150 lines of code in scope.
A local zero-token check script is run first and must fail (red).
One stateless run, 8 steps by default.
The outer script re-checks: green means the ticket is claimed automatically and the next ticket starts.
Multi-assistant review · Improvement
Two rounds with no progress trip a circuit breaker (STOP), or a new kind of error shows up.
One focused root-cause review.
The fix goes into the check scripts, the dispatch rules or a skill card (2 good examples + 1 bad example).
The same kind of mistake does not happen twice.
“Once the framework is stable, the human is only there for ideas and final calls.”
Dev loop Advanced the eleven-step schematic component recognition pipeline and the four-layer line-first engineering drawing engine, now covering analog circuits, piping diagrams, and power single-line diagrams; produced fifty de-identified comparison visuals for side-by-side review.
Dev loop Rebuilt mission control as a realistic seven-stage dual-track rocket flow across five rocket classes: rollout, stacking, ignition, max dynamic pressure, and stage separation on the shared ascent, then first-stage booster recovery for the summary reply and second-stage payload deploy or full ship catch for task completion.
Evolution loop Hardened local model exam prompts with an explicit untrusted-payload boundary header, a flat top-level JSON rule, and a two-line node routing checklist, so local models resist prompt injection and routing confusion without slow reasoning chains.
Human Agreed to build the complete suite together for internal work: keep only the chat entry and mission panel visible by default, light up advanced and engineering modules only when licensed, and strip internal developer views in one step when packaging a release.
Engineering AI as two loops, not vibe coding
Dev loop Line-first circuit reading: trace every wire, junction and crossing before placing parts. Real schematics fully matched went from 118 to 139, with zero regressions.
Dev loop A zero-token layer of part physics (what passes DC, what blocks it) now tells apart two identical transistors: no feedback resistor means a digital inverter; add feedback and a coupling capacitor and it is a high-gain analog amplifier. Matches rose to 144 (82.66% on the dev set).
Dev loop The dev loop: thin work tickets, a zero-cost check that must fail first, one stateless AI run, then an outside script decides pass or fail. The AI can no longer grade its own work.
Evolution loop The evolution loop: two rounds with no progress trip a breaker, one deeper diagnosis finds the root cause, and the fix goes into the checker and the skill notes (two good examples plus one bad example, with why it is wrong).
Human More AI budget arrived, and the limits got stricter, not looser. The human's line of the day: once the framework is stable, people only bring ideas and make the final call.
Rules that carry their own reason
Dev loop Machines now pick their own lead by hardware: the desktop with the big graphics card takes the heavy local model, the laptop runs light, cool checks, and takes over if the desktop sleeps.
Dev loop Six different AI tools now go through one dispatch script: small slices under 150 lines, timed automatically, every run written to one shared ledger, so it is easy to see which tool is fast and which is accurate.
Evolution loop Rules moved out of long side documents and into the assistant's built-in memory. Each rule now carries why it exists and what went wrong last time, so it generalizes to new situations instead of being followed word by word.
Evolution loop New hard stop: if the same ticket shows no progress for two rounds, stop, cut it smaller or ask for a diagnosis. Retrying blindly just burns budget.
Show, don't list: examples beat rules
Dev loop The Mac build was installed on two Macs: install, open, reply and close all passed. Screen recording failed because the permission was off, so we attached a plain terminal transcript as proof instead of faking a video.
Dev loop New installs now default to Auto Route: cloud first, local model as fallback. Nobody sees "no assistant available" on first launch any more.
Evolution loop A long list of rules still produced robotic, question-heavy replies. The fix was a living example sheet: the same situation answered the bad way and the good way, side by side. Examples teach faster than rules.
Human The human set the local model's role: try locally first, go to the cloud only when stuck. The local model may suggest, but never decides.
Fix the ruler before blaming the student
Dev loop A second assistant started work with access to one shared folder only, talking through a message board. It closed more than ten tickets in a day. Routine progress is written silently to a status file; only done, blocked or needs-a-decision triggers a message.
Dev loop Same ticket, two coding tools: one delivered in about 6.5 minutes, the other hit its usage limit after 3 minutes on a US$20/month plan. Lesson: spread work across tools instead of betting on one.
Evolution loop Every test setup failed the same exam questions. The root cause was the scoring script, not the models. After fixing the scorer, a 14b model went from 17/34 to 20/34.
Human The human named the public template: an Evolution loop plus a Development loop. "Every optimization is about getting the most out of each token."
Parallel work, zero conflicts
Dev loop Two, then three coding assistants worked at the same time, each owning files nobody else touched: 0 conflicts, about 7 to 8 minutes per ticket, 12 changes merged in one day.
Dev loop First real install on an old laptop: installed in 8.7 seconds, clean uninstall, chat working. The English/Chinese switch shipped with English rewritten for English readers, not translated word by word; 39/39 tests passed.
Evolution loop Two assistants under-reported their own file re-reads and one claimed an existing file was missing. New rule: every report carries numbers that can be cross-checked.
Evolution loop A 4b model scored 4/24, worse than a 1.7b model at 15/24. It was not weaker; it kept thinking out loud and got cut off mid-sentence. Check for truncation before judging a score.
Human "First make the entry app something we are happy with." Dispatch requests and help requests became fill-in forms that a script checks for missing fields.
Write down what done means
Dev loop One change bounced back and forth 20 rounds before it merged. New rule: write the acceptance criteria before dispatch, list every issue in one review, and allow at most 3 send-backs before rethinking the design.
Dev loop A tiny desktop widget shows what runs in the background and pauses everything with one button. It checks status every 10 seconds with no AI at all, so it costs zero tokens. 17 changes merged that day.
Evolution loop Instead of training a small local model, it gets a capability handbook: one card per task type. No card, no guess; it says "I don't know".
Human "Code decisions should not wait on me." The human now decides only three things: how it feels to use, spending or publishing, and granting bigger permissions. Everything else merges after all tests pass and a strong reviewer model signs off.
Remember the rules, keep things moving
Evolution loop The most important rules were squeezed into a hot list of no more than 30 lines, read at the start of every session. Tested on 62 questions: 62 correct, using roughly a third of the words of the full rulebook.
Dev loop A small local model took more than 5 minutes per reply because its thinking had no limit. With a cap, the same task took 6 seconds.
Dev loop The new home page passed every test yet stayed stuck on "Loading..." because of one mismatched quote the tests never checked. Passing tests is not the same as working; open it and look.
Human Design goal in one line: simple to start, hard to put down. Coding goes to strong coding tools; small local models handle small, automatically checkable chores.