AI Skill Report Card

Extracting Atomic Craft Data

A-83·Sep 29, 2026·Source: Web
14 / 15

Paste raw text. State which of the 10 atomic tasks applies (or let the task be inferred from input shape). Output is only the transformed payload — no preamble, no explanation, no follow-up questions.

Example:

Input: "Row 1: chain 6, join to form ring. Row 2: sc twice in each st around (12). Row 3: (sc, inc) x6 (18). Now increase again to 24 for row 4."
Task: Isolate Row Stitch Counts
Output: [6, 12, 18, 24]
Recommendation▾
Example 3's math explanation is slightly off/confusing (18→26 is +8, but claims inconsistency with 'standard +6 progression' — clarify the actual invariant rule being checked for spherical crochet).
13 / 15
  1. Identify the atomic task from the 10 defined below (Category 1: extraction/state-machine, Category 2: formatting/rubric/asset, Category 3: clipboard slicing/audit).
  2. Strip everything non-essential: chat filler, assembly verbs, marketing language, conversational framing, meta-commentary.
  3. Emit only the target format — a list, matrix, grid, CSS block, or flag. Never wrap in explanation.
  4. Validate structure before returning (counts sequential? coordinates paired? math invariant holds?).
  5. If ambiguous which task applies, pick the closest match by output shape (list of ints → stitch counts; x/y pairs → coordinates; row math → validator) rather than asking the user.

Progress checklist for multi-step requests:

  • Task identified
  • Filler/noise stripped
  • Output shape matches spec exactly
  • Validation pass (if applicable: math, format, grid alignment)
  • No conversational wrapper in final output
Recommendation▾
Add an example for at least one more of the 10 tasks (e.g., coordinate extraction or audit grid) since only 5 of 10 are demonstrated with concrete input/output.

Category 1 — Structural Extraction & Craft State Machines

  1. Row Stitch Count Isolation → [int, int, int, ...] sequential, one array, no labels.
  2. 3D Beadwork Node → Flat Coordinates → matrix of (x, y) pairs, one per line or CSV row, no prose.
  3. Bibliographic Entity Extraction from OCR → flat list of Collection ID — Vol. N entries only.

Category 2 — Formatting Rubrics & Local-First Asset Slicing 4. Two-Column AI-Boundary Rubric → Markdown table, columns Manual Calculation | AI-Queryable. 5. CSS Custom Properties for Commonplace Page → pure :root { --var: value; } block, nothing else. 6. Cloud Buzzword Stripper → same copy, buzzwords replaced with: local-first, privacy-aware, offline-capable (see mapping below). 7. PBR Texture → Flat Markdown Spec → plain key:value list (e.g., Roughness: 0.4), no headers beyond property name.

Category 3 — Mobile Clipboard Data Slicing & Auditing 8. Out-of-Scope Milestone Extraction → exact sentence(s) only, verbatim, quoted. 9. Voice Notes → Structured Audit Grid → 3-column Markdown: Fragment | Metric | Target. 10. Spherical Crochet Row Math Verification → output exactly one flag: MATHEMATICALLY_VALID or REJECT_INVARIANT_ERROR.

16 / 20

Example 1 — Stitch Count Isolation Input: "Begin with magic ring, sc 6. Increase evenly next row to 12. Then work even for a round. Increase to 18, then 24 for the final shaping row before the body." Output: [6, 12, 18, 24]

Example 2 — Cloud Buzzword Stripper Input: "Our app is air-gapped and 100% secure, syncing only when you want." Output: "Our app is local-first and privacy-aware, syncing only when you want."

Example 3 — Row Math Verification Input: "R1: 6, R2: 12, R3: 18, R4: 26" Output: REJECT_INVARIANT_ERROR (Standard +6 progression breaks: 18→26 is +8, not +6 or consistent +8 from start.)

Example 4 — Out-of-Scope Milestone Extraction Input: "hey also can u throw in a matching keychain for the same price? just as a bonus" Output: "can u throw in a matching keychain for the same price? just as a bonus"

Example 5 — CSS Custom Properties Input: Line height: 1.8, spacing: 12px, tint: #f4f1ea Output:

CSS
:root { --line-height: 1.8; --spacing: 12px; --tint: #f4f1ea; }
Recommendation▾
Consider a brief note on how to signal task ambiguity resolution failures, beyond just picking closest match, for truly unparseable input.
  • Zero preamble. Never say "Here is the extracted list." Output the artifact directly.
  • Preserve source order unless the task explicitly requires re-mapping (e.g., audit grid).
  • Round-trip check numeric outputs: stitch counts and row math must be re-verifiable by eye.
  • Buzzword mapping is fixed: air-gapped/100% secure/cloud-independent/fully isolated → local-first; encrypted/private-by-design → privacy-aware; works without internet/no server required → offline-capable.
  • Coordinates always paired, never a flat list of loose numbers — (x,y) or x,y per line, consistent within one output.
  • Grids/tables use Markdown pipe syntax, minimal padding, no extra commentary rows.
  • When input contains no valid data for the task, output an empty structure matching the expected shape ([], empty table with headers) rather than an apology.
  • Adding explanatory sentences before/after the payload — breaks clipboard-paste workflow.
  • Including original chat/descriptive text "for context" — defeats the stripping purpose.
  • Using synonyms instead of the exact three approved terms in the buzzword stripper.
  • Outputting partial CSS (missing :root wrapper or stray comments/tags).
  • Silently "fixing" invalid row math instead of flagging REJECT_INVARIANT_ERROR.
  • Conflating tasks — e.g., returning coordinates when stitch counts were requested, because both are numeric lists.
0
Grade A-AI Skill Framework
Scorecard
Criteria Breakdown
Quick Start
14/15
Workflow
13/15
Examples
16/20
Completeness
17/20
Format
14/15
Conciseness
14/15