Engineering Weave Drafts
Engineering Weave Draft Skills
Decomposes the end-to-end task of "AI reads and processes a weaving draft" into discrete, independently buildable sub-skills, organized into four sets. Each set covers one layer of weaving expertise a human weaver applies sequentially: reading the grid, understanding loom mechanics, expanding abstract block patterns, and calculating physical yarn/reed requirements.
When asked to decompose a weaving/draft-related AI capability, don't just list features — map them to the four-set structure:
- Set 1 — Draft Architecture: threading/treadling/tie-up reading, shaft counting, drawdown simulation, float scanning
- Set 2 — Loom Mechanics: rising/sinking shed, direction conventions, lift plans, table/dobby conversion
- Set 3 — Block Theory: profile drafts, block substitution, unit vs. non-unit weaves, shaft-count formulas
- Set 4 — Physical Engineering: WPI/sett, reed/sleying, take-up/shrinkage, width-in-reed, floating selvedges, balance, flaw-scanning
For each set, produce a numbered skill inventory table: # | Skill | What It Does | Why It Matters | Needs Tools?
Progress:
- [ ] Identify which set(s) the request falls under
- [ ] State foundational concepts the AI must internalize (with the "common misconception" called out explicitly)
- [ ] Build the skill inventory table (10-20 rows per set)
- [ ] Flag which skills are pure computation vs. which need vision/web/image-gen tools
- [ ] Give a recommended build order (highest-consequence + lowest-complexity first)
- [ ] List key distinctions/pitfalls as a final section
- [ ] Offer to spec out 2-3 specific sub-skills in more depth
Research first when uncertain: weaving terminology has subtle, high-stakes distinctions (e.g., rising vs. sinking shed, unit vs. non-unit weaves). If unsure whether a rule is universal or structure-specific, search/verify rather than assert — wrong formulas here cause wasted yarn and unwearable cloth, not just wrong text.
Set 1 — Draft Architecture
- Threading (top grid) = warp-to-harness map; Treadling (right grid) = treadle press sequence; Tie-up (small grid) = harness-to-treadle map
- Drawdown = matrix product of treadling × tie-up × threading; this is pure math, not interpretation
- Fork early: WIF file (structured, parse directly) vs. image (needs vision + grid-line detection, much harder, lower confidence)
Set 2 — Loom Mechanics
- Rising vs. sinking shed is about which direction shafts move, not a property of the shed itself. Rising shed (jack/table/dobby): pressing lifts tied shafts. Sinking shed (counterbalance/countermarch): pressing lowers them.
- Tie-up symbols aren't universal: O often = rising, X often = sinking — never assume, flag ambiguity.
- Threading reads right-to-left (mirrors standing at the loom); treadling reads top-down (reading convention) but weft builds bottom-up on cloth — this mismatch flips twill diagonal direction if not tracked carefully.
- Table/dobby looms use lift plans, not tie-ups — no treadles, just direct shaft selection per pick.
- Table-loom conversion algorithm: rotate the tie-up 90°, read each resulting row as one pick, each number in the row as a shaft/lever to raise.
- Countermarch is the exception to simple tie-up inversion: every shaft must be explicitly tied up AND down per treadle, or it floats mid-shed.
Set 3 — Block Theory
- A profile draft is a block map (one shaft = one block), not a literal threading. It requires a substitution template to expand into real threading/treadling/tie-up.
- Unit weaves (Summer & Winter, Atwater-Bronson lace) bind pattern and background inside each unit — infinitely extensible. Non-unit block weaves (overshot, crackle) float the pattern weft over the block — float length caps how far you can extend.
- Shaft-count formulas are structure-specific, never universal:
- Summer & Winter: shafts − 2 = number of blocks
- Turned twill: 4 shafts per block
- Overshot: 2 shafts per block, but adjacent blocks share one shaft (A:1,2 / B:2,3 / C:3,4 / D:4,1)
- Crackle: 3 shafts per block, overlapping
- Satin: 4-6 shafts per block
- Tabby picks bind pattern floats — not optional on structures that require them.
Set 4 — Physical Engineering
- WPI ÷ 2 = plain-weave sett only. Twills need closer sett depending on float length — never apply WPI/2 universally.
- Width in reed = finished width + weft contraction (cloth shrinks narrower during weaving) — omitting contraction undersizes the cloth.
- Take-up/shrinkage has no universal percentage — it depends on structure, material, twist, and finishing. Never default to a flat 10%.
- Floating selvedges are structure-conditional: needed when an edge warp thread would otherwise be raised/lowered for 2+ consecutive picks (typical in twills); not needed in plain weave.
- "Balance" has two unrelated meanings: (1) threading/treadling symmetry (point repeats needing a closing end), (2) EPI = PPI sett balance. Never conflate them.
- A drawdown can look correct but fail a formal "hang-together" / interlacement validation — visual inspection alone isn't sufficient.
Example 1: Input: "What skills does an AI need to convert a floor-loom draft to a table loom?" Output: Skill inventory scoped to Set 2 — Shed-Type Classifier, Tie-Up Inverter, Lift-Plan Generator, Tie-Up→Lift-Plan Rotator (90° rotation algorithm), Loom-Constraint Checker — each with why-it-matters tied to the specific failure mode (e.g., skipping the classifier risks weaving the inverse shed).
Example 2: Input: "How does an AI know how many shafts a profile draft needs?" Output: Route to Set 3's Shaft-Count Calculator skill; return the structure-specific formula table (not a single universal number), and flag that the formula changes entirely based on weave structure name.
- Always separate pure computation skills (matrix math, formula application — low risk, build first) from tool-dependent skills (vision parsing, web lookups, image rendering — higher risk, build last, needs confidence gating).
- Every skill inventory should end with a build order ranked by consequence-of-error ÷ complexity-to-build, not just logical sequence.
- Call out the single highest-stakes misconception per set explicitly (e.g., "rising/sinking shed is about shafts, not the shed") — these are the errors that produce physically wrong cloth, not just bad explanations.
- When uncertain about a structure-specific rule (shaft formulas, sett ratios, tie-down rules), verify via search rather than guess — state explicitly when a rule is universal vs. structure-dependent.
- End every set breakdown with an offer to spec out 2-3 specific sub-skills in full detail (pseudo-code, decision trees, rule libraries).
- Don't apply a rule learned for one structure (e.g., Summer & Winter's shaft formula, or plain weave's WPI/2 sett) universally across all weave structures.
- Don't treat "block weave" and "unit weave" as synonyms — overshot is a block weave but explicitly not a unit weave.
- Don't silently invert a tie-up without checking for countermarch looms, which require explicit up/down ties rather than simple inversion.
- Don't skip the WIF-vs-image fork — assuming an image-only pipeline when a structured file format is available wastes effort on the hardest, least reliable sub-skill (vision grid parsing).
- Don't present take-up/shrinkage or floating-selvedge requirements as fixed defaults — both are structure- and material-conditional and must be derived, not assumed.