Deft
Ask AI

Search

Search skills, authors and topics

tt-a1i / archify

archify-tuning

Improve Archify diagram accuracy, visual quality and complete authoring speed with fewer repairs. Use for generation experiments or evaluating changes to layout, routing, diagnostics, authoring instructions and acceptance rules against actual diagrams and agent traces.

View sourceVisualCraftSystems

Install

npx skills add https://github.com/tt-a1i/archify --skill archify-tuning

Trace signals

Read <round>/<type>.trace.txt with the subagent's report and actual diagram. These are hypotheses, not verdicts: confirm them against source facts and the failing input, and retain contributing factors or uncertainty. Apply the main Skill's product requirements and adjustable-rule policy before deciding what to repair.

Signal in the traceLikely causeCheck
The final candidate changed or dropped nodes, edges or lanesLost meaning or a source-backed correctionUse collect.mjs "lost" signals, then compare source facts, conditions and ownership; distinguish correction from removing real content to pass
A draft without route controls fails crossing, corridor, through-node or label checksPossible layout defect or authored placement constraintInspect node placement, dimensions and canvas first; render a scratch copy at standard to inspect routes, keeping the original failure
A crossing or corridor failure, then side pins, then explicit-pin-conflict, then a restructurePlacement, routing or a misleading repair hintTest the proposed pins on the original input; check for a legal route before concluding none exists
The same code survives a focused repairIncomplete repair evidence, a new instance or an unsuitable ruleInspect the identified element and actual harm, then reproduce the advertised fix
Width or gap values oscillate across runs, such as gapX 64, 108, 72Missing target information or interacting constraintsCheck whether the rule is appropriate and whether the suggested size can satisfy the other constraints
The agent writes meta.viewBox after a width messageUnclear canvas advice or an unsuitable fit requirementInspect actual readability, explicit geometry and accepted scrolling before suggesting a canvas change
The agent renders a scratch copy at standard, reads renderer source or greps references for a ruleMissing evidence, hard-to-find guidance or normal investigationRead why the detour was necessary; improve the diagnostic or required read only for a demonstrated gap
[truncated N chars] on a read of a referenceContent exceeds this tool's output budgetCheck the actual read and shorten or restructure the necessary material
Two or more subagents list the same unclear pointInstruction usability problemTest whether clearer guidance reduces misunderstanding or rework
A rerun is blocked by stale evidence or output pathsDelivery frictionGive the exact command in the message
A pass on the first runMachine acceptance onlyAudit source meaning and inspect the complete diagram
A clear, accurate diagram repeatedly fails a style or geometry heuristicAn overbroad acceptance ruleIdentify the intended harm; compare a narrower rule or advisory diagnostic on contrasting cases
Two selected routes get shorter but other routes become harder to followLocal placement optimization harms the whole graphCompare all relationships, labels, overview and repair cost before adopting the move

Root causes found so far

These historical cases illustrate possible causes. Check the current revision before treating their dimensions, diagnoses or fixes as current behavior.

  • Workflow same-lane neighbours were placed 8px apart, leaving little route space. Automatic drafts failed on crossings and shared corridors, and agents removed lanes and nodes. Unpinned drafts now use a 32px corridor and may take offset tracks; absolute route pins retain their existing geometry contract.
  • Dataflow flows across several stages turned at their midpoint, inside a middle stage's node. Agents deleted nodes. They now turn in the nearest clear gap.
  • Dataflow ignored perpendicular pinned sides, so agents hand-wrote via points.
  • Sequence canvases stayed 920px wide whatever the participant labels needed, and the message suggested widening viewBox, which the readability budget forbids.
  • Workflow diagnostics named edges by an internal sort order, so agents edited the wrong edge.
  • Architecture candidate selection ignored the number of bends, and legends reserved width at a smaller font than they rendered.
  • A compile-time text floor applied to tags, which browser-check exempts: a self-inflicted false failure, found two rounds late because the affected test file was not run.

Previously rejected experiments

Retain these negative results as context. Revisit only with a different supported hypothesis; they are not universal bans on the approach.

  • Routing Workflow edges by role and span order: more crossings than the canonical order.
  • Ripping up and replanning conflicting Workflow routes after the first pass: it traded crossings for shared corridors and made no draft pass.
  • Shrinking or removing content automatically to fit a budget: it hides the accuracy loss this skill exists to prevent.