Lintel on GitHub Copilot. Generated from
skills/generate-style-learn/SKILL.md; edit the canonical file, then runli-copilot init.
- Resource root:
../../..from this skill's base directory (the Lintel source withbin/,lib/,skills/). Write plans, state and evidence into the working repository's.claude/tree, never into the resource root.- Skill-relative paths: this skill's own
scripts/,references/anddata/folders (and a<base>that the workflow defines as its own directory) mean../../../skills/generate-style-learn/in the Lintel source, not this generated folder.${LINTEL_SKILLS_DIR:-skills}means the skills root,../../../skills. Abin/li-runstep runs in the working repository, so use$LINTEL_SKILLS_DIR/generate-style-learn/there.- Shell steps: run Bash snippets with Bash (Git for Windows'
bash.exeon Windows, neverSystem32\bash.exe). Save a snippet to a temporary.shfile and runbash "<resource root>/bin/li-run" <file>; it preparesLINTEL_SOURCE_ROOT,LINTEL_REPO_ROOTand the profile context.- Tools: Read=
view, Write=create, Edit=edit, Bash=bash/powershell, Grep=grep, Glob=glob, AskUserQuestion=ask_user, TodoWrite=the plan checklist, Task or a named role=taskwith that custom agent, WebFetch=web_fetch.- Other Lintel workflows are native skills: invoke
/li-<name>rather than reading their files. Named roles such asCodeReviewerare custom agents.
You are the generate-style-learn skill — extracts palettes from selected existing artifacts.
What this skill does
Analyzes 1+ supported artifact-files (PPT/DOCX/web) and extracts a reusable style palette
to <out-dir>/<name>.json plus companion <out-dir>/<name>-STYLE.md (human-readable).
Choose the source paths and owned output explicitly; a palette name is not a
personal-directory lookup.
Never modifies input-files. Read-only analysis.
Use case: operator gets a customer-brand-deck → wants to extract the palette + apply it to future generation-runs without manually curating tokens.
Follow owned source and output selection. Missing or unwritable output fails visibly; no destination is selected on the operator's behalf.
When to use
- "Customer sent a deck — extract their style so I match it in the next report" →
/li-generate-style-learn deck1.pptx --name customer-A - "Compare two style-references" → run the skill on each, diff palettes
- "Operator's own brand snapshot" → compare explicitly selected palettes against the supplied ground-truth artifacts
When NOT to use
- Live style-edit — this is extraction, not an editor
- Single-color-pick — use the named
brand-sourceselected asset evidence procedure to read the requested token from the explicit palette/profile; no diagnostic command - Customer-specific live-stream — this is batch
Inputs
- Required
<paths>— 1+ artifact files (varierande format ok) - Required
--name <slug>— palette-name for output (e.g.,customer-Aornordic-minimal) - Optional
--format <ppt|web|word|auto>— explicit format hint (default: auto-detect) - Optional
--overwrite— replace existing<name>.jsonif present - Required
--out-dir <path>— explicitly owned repository-relative palette output directory
Extraction (per format)
PPT (.pptx):
- Open via python-pptx OR pptx-genjs
- Iterate theme colors → most-frequently-used fill/text colors
- Classify into primary/secondary/accent/text/background (60-30-10 rule)
- Extract heading font + body font + most-common sizes
- Layout distribution (% per layout type)
- Logo position from Slide Master
- Visual patterns (accent bars, animation style)
DOCX (.docx):
- Open via python-docx OR docx-templater
- Theme colors from docDefaults + styles.xml
- Font hierarchy (heading 1/2/3 + body)
- Margins + line-spacing
- Header/footer styles
Web (HTML/CSS):
- Parse
<style>+ linked CSS - Extract :root CSS variables (preferred source)
- Fall back: most-frequent computed colors in DOM
- Typography from font-family declarations
PDF or another unsupported format:
- Lintel supplies no PDF reader here. A
.pdfextension is not an extraction capability and a print/export result is not readable source evidence. - Use an actually available external read operation only with explicit source and operation authority, recording its coverage and limitations. Otherwise report that extraction is unsupported/unverified and leave that input open.
- Do not install a reader, restore the removed PDF dependency, invent style observations or mark a partially inspected set complete.
Auto-detect: the extension is a format hint, not proof of a reader.
Output schema (palette JSON)
{
"name": "nordic-minimal",
"source_files": ["deck1.pptx", "deck2.pptx"],
"extracted_at": "2026-05-28T18:00:00Z",
"colors": {
"primary": "#002855",
"secondary": "#4A90D9",
"accent": "#F2A900",
"text_dark": "#1B1B1B",
"text_light": "#FFFFFF",
"background": "#FFFFFF"
},
"fonts": {
"heading": "Segoe UI Semibold",
"body": "Segoe UI Light",
"heading_size_px": 32,
"body_size_px": 13
},
"layout_distribution": {
"content": 0.55,
"section-header": 0.15,
"title": 0.10,
"two-column": 0.10,
"data-viz": 0.10
},
"visual_patterns": {
"logo_position": "top-right",
"accent_bar": "left-edge-vertical",
"bullet_style": "square",
"animation": "fade"
}
}
Output schema (companion STYLE.md)
Human-readable summary describing the style in plain language:
# <name> style
**Extracted from:** deck1.pptx, deck2.pptx (2026-05-28)
## Color story
Primary navy (#002855) with bright accent (#F2A900). High-contrast text-on-light.
60-30-10 ratio: navy dominant, blue secondary, gold accent.
## Typography
Segoe UI family across the board — Semibold for heading (32px), Light for body (13px).
Tight line-height, deliberate hierarchy.
## Layout pattern
Content-heavy (55%) with regular section-header punctuation (15%). Balanced
two-column + data-viz (10% each).
## Visual signature
Logo top-right, vertical accent bar left edge, square bullets, fade animations.
## Use this style
In future generation: `/li-generate ... --palette <name>`. Catalog at
<out-dir>/<name>.json; supply this source explicitly or register it in the
authorized configuration before name-based reuse. Extraction does not activate it.
Workflow
Step 1 — Validate inputs
out_dir="${OUT_DIR:?select an owned repository-relative output directory}"
[ -z "${NAME:-}" ] && { echo "Usage: /li-generate-style-learn <files> --name <slug> --out-dir <path>"; exit 2; }
[ "$#" -gt 0 ] || { echo "Required: 1+ artifact files"; exit 2; }
for f in "$@"; do
[ -f "$f" ] || { echo "Missing: $f"; exit 2; }
done
Complete the shared owned-path checks before reading or creating anything.
Capture original output states for both named files. Validate NAME as a single
slug, not a path; missing/linked/out-of-root sources or outputs fail. The actual
configured P07 path is an alternative source only when verified and authorized.
Step 2 — Per-file extraction
Use the extension as a hint, then establish the actual supported read operation. Apply the format-specific method above. Unsupported/unreadable inputs stay explicitly unverified; continue independent supported inputs without claiming that the entire requested set was processed.
Step 3 — Aggregate across files
If multiple files: merge extractions:
- Colors: union of palettes; classify into 6 roles via 60-30-10
- Fonts: most-common heading + body
- Layout distribution: weighted average
Step 4 — Write outputs
Keep the generated strings as palette_json and style_md. With explicit repo,
out_dir, name, boolean overwrite, and the captured original_output_states
keyed by repository-relative destination, publish through existing P03:
import re
import sys
import context_safety as safety
try:
root = safety.checked_root(repo)
directory = safety.selector_path(out_dir)
if not re.fullmatch(r"[A-Za-z0-9][A-Za-z0-9_-]*", name):
raise ValueError("Palette name must be a single slug")
outputs = {
f"{directory}/{name}.json": palette_json.encode("utf-8"),
f"{directory}/{name}-STYLE.md": style_md.encode("utf-8"),
}
for path in outputs:
previous = original_output_states[path]
if previous is not None and not overwrite:
raise ValueError(f"Output already exists; explicit --overwrite required: {path}")
for path, payload in outputs.items():
previous = original_output_states[path]
safety.atomic_write(root, path, payload,
mode=previous["mode"] if previous is not None else 0o600,
expected=previous, check_expected=True)
if safety.read_owned(root, path, len(payload))[0] != payload:
raise ValueError(f"Output failed readback: {path}")
except (KeyError, ValueError, OSError, UnicodeError) as error:
print(f"ERROR [lintel/style-learn]: {error}", file=sys.stderr)
raise SystemExit(2)
This is publication, not extraction or license approval. Both files must pass readback before reporting completion. Retain and report any partial output on failure at its original owned destination; do not switch folders or erase evidence.
Step 5 — Surface confirmation
Style 'nordic-minimal' extracted.
Source files: deck1.pptx, deck2.pptx
Colors: 6 roles classified
Fonts: 2 family (Segoe UI Semibold heading, Segoe UI Light body)
Layout distribution: 5 categories
Saved: <out-dir>/nordic-minimal.json
<out-dir>/nordic-minimal-STYLE.md
To use: /li-generate ... --palette nordic-minimal
Reusable patterns
Follow the reusable pattern consumer contract. The
palette and STYLE.md output is unchanged and is observation, not policy. Palette tokens can
reach a universal draft only through the frontend-style-extract adapter as
visual.palette.<token> defaults with exact #RRGGBB values. Fonts, licensing and
accessibility are never inferred as confirmed from an image or page.
Status protocol
- DONE — palette + STYLE.md written and every required selected source actually processed
- DONE_WITH_CONCERNS — required sources were processed; advisory details remain uncertain (e.g., logo-position reliability)
- BLOCKED — any required source/reader is missing or unverified, name is absent, or output permission fails; retain partial observations without claiming the full request passed
- NEEDS_CONTEXT — the format or authorized reader is unresolved; a format flag does not supply a missing reader
Pause-points
--overwritenot set but palette name exists: hard-block for operator-confirm- Multiple wildly-different styles in source files: surface "sources don't agree, palette will be averaged — proceed?"
Integration
Reads:
- Source artifact files (
<paths>)
Writes:
<out-dir>/<name>.json<out-dir>/<name>-STYLE.md
Consumed by:
/li-generate --palette <name>(downstream format-builders)/li-generate-design --palette <name>(in shared pipeline)
Anti-patterns
- Modify source files — extraction is read-only.
- Inferred color-roles without justification — if the 60-30-10 ratio is unclear, surface the ambiguity + ask for operator-guidance rather than guessing.
- Overwrite without confirm —
--overwriteflag required for replacement.
Failure recovery
- File format-detection fails: surface available formats + ask for a
--formathint - Aggregation conflict (multiple wildly-different styles): default to most-recent file's style + flag with warning
- Output missing/unwritable: BLOCKED with the actual selected path and error; preserve inputs and any owned partial output. No current-directory, HOME or neighboring-run fallback; a new destination requires explicit selection.
Recommended next steps after invocation
- Verify extraction quality: open
<name>-STYLE.md+ manually check claim-vs-source - Test palette:
/li-generate "test brief" --palette <name>→ verify match - For multi-source aggregation: re-run with
--diff <existing-palette>to see what new sources added (future feature)