Routine — Triage Brand Assets

A subroutine of crawl-fetch-ingest. After CP3 (portfolio companies) finishes, brand-asset quality is mostly good but not uniform — sometimes the strip is suspect, sometimes the foreground is too light or too dark for its frame, sometimes the resolution won’t scale. This routine is the human-in-the-loop checkpoint that catches those cases before publishing.

When to invoke

The user says something like:

“triage the brand assets”
“review the logos for {firm}”
“classify the brand assets”
“audit the portfolio logos”
“walk through the assets”

Default scope is the most recently worked-on firm. If multiple firms exist in <cwd>/data/firms/, ask which one (or accept a --firm <slug> argument).

Output

A review_status field on each company’s portfolio/{co-slug}.md frontmatter, set to one of:

Status	Meaning
`pending`	Default. Not yet reviewed.
`good-to-go`	Asset is clean and ready to publish as-is.
`not-urgent-passable`	Works but could be upgraded later (e.g., raster instead of SVG, or low resolution).
`urgent-rework`	Visibly broken or wrong (missing, wrong company, bg-strip failed, light-on-light, etc.). Will look bad in production.
`deferred-for-now`	User explicitly chose to skip this round. Re-surface next pass.

Plus optional review_notes: "..." for any human note, and reviewed_at: <ISO timestamp>.

Workflow for Claude

When invoked:

1. Run the auto-classifier

~/.claude/skills/crawl-fetch-ingest/scripts/triage-classify.py <firm-dir>

Outputs a JSON report with one entry per company describing the suggested classification + the signals that led to it (file format, size, dimensions, % opaque pixels, mean foreground luminance, bg color detected, whether a frame bgColor is wired for it in the deck, etc.).

The classifier never persists anything on its own — it only suggests. The agent must confirm with the user before writing.

2. Present the summary

Show counts per suggested classification, with the urgent + flagged cases listed by name:

Triage scan complete for calm-storm-ventures:
  good-to-go ........... 14 (no action needed unless you disagree)
  not-urgent-passable .. 8
  urgent-rework ........ 5  → Squaremind, 9am-Health, Visible, lillian-care, lindus-health
  deferred-for-now ..... 0

Walk through the 5 urgents now? (yes / show all / let me pick / cancel)

3. Conversational walk-through

For each asset in the chosen subset, present a compact dossier and ask for a classification:

[3/5] Visible
  File:           portfolio/trademark__Visible.png  (12 KB, 800×800, alpha)
  Source:         og:image from www.calmstorm.vc/portfolio/visible
  bg-strip:       ImageMagick flood-fill, 14.1% removed, detected bg #000000
  Foreground:     mean luminance 234 (very light glyphs)
  Deck frame:     dark (bgColor #000000) — light glyphs visible
  Suggestion:     urgent-rework — only 14% stripped, suggests the strip got confused
                  and most of the original raster is intact. Risk: bg color
                  detection wrong, glyph + bg both dark.
  Open the file:  /Users/.../portfolio/trademark__Visible.png

Classify as: [g]ood-to-go  [n]ot-urgent-passable  [u]rgent-rework  [d]efer  [s]kip-to-next  [q]uit

The user types a single letter. The agent persists the choice to the .md frontmatter:

review_status: urgent-rework
review_notes: ""              # any free-form text the user added
reviewed_at: 2026-05-10T15:00:00Z

Then move to the next.

4. Optional one-letter shortcuts

[m]essage — capture a free-form note for this asset (e.g., “ask client which logo variant is canonical”)
[o]pen — open <path> to view the file in Preview / browser
[h]omepage — open <homepage> to inspect the company’s brand directly
[r]eclassify-all-from-here — set the same class for this + all remaining

5. End-of-routine summary

Triage complete for calm-storm-ventures. Final tally:
  good-to-go ........... 16
  not-urgent-passable .. 7
  urgent-rework ........ 3
  deferred-for-now ..... 1

Urgent-rework list (suggested next moves):
  - Visible:       try Brandfetch tier 4 — site has no usable SVG path
  - lillian-care:  trainCase produced "Lilliancare"; rename + try logo-hunt on lilliancare.com
  - 9am-Health:    foreground luminance unstable; manual cleanup in Figma recommended

Deferred:
  - aiomics: user marked for later — re-surface next triage pass.

Auto-classification rubric

The classifier scores each asset across a handful of axes, then chooses the worst-case bucket among the axes that flagged:

Axis	Check	Worst-case bucket
Existence	File at the path in `logo:` frontmatter exists	`urgent-rework` if missing
Format	`svg` → automatic `good-to-go` for the format axis	`good-to-go` for svg
bg-strip success	If `logo_bg_stripped: true`, check `pct_transparent` (recomputed)	<30% = urgent-rework, 30–60% = not-urgent-passable, ≥60% = good-to-go
Foreground luminance	Mean luminance of opaque pixels (PNG/raster only). Reads frame context: if the deck assigns a dark frame, light glyphs are fine; if no bgColor (default white frame), light glyphs are bad.	Light glyphs on white frame = urgent-rework
Resolution	Width or height (raster only)	<80px = urgent-rework, 80–200px = not-urgent-passable, ≥200px = good-to-go
File size	Disk size	<500 B = urgent-rework (probably broken), 500 B – 4 MB = good-to-go, >4 MB = not-urgent-passable
Flagged unresolved	`status: unresolved` or `status: flagged` in frontmatter	`urgent-rework` regardless of other signals

The agent’s job during the walk-through is to confirm or override the suggestion. The classifier is wrong sometimes — that’s why the human is in the loop.

Persistence rules

Only fields review_status, review_notes, reviewed_at are written. Nothing else in the frontmatter is touched.
If review_status is already set (from a prior pass), surface that to the user with “Last reviewed YYYY-MM-DD as X. Re-classify? (y/N)”. Skip without prompting if --skip-already-reviewed.
All writes are idempotent. Re-running the routine on a fully-reviewed firm is a no-op unless the user accepts the re-classify prompts.

Note on scope

This routine reviews brand assets (portfolio company logos). The same pattern extends to person assets (headshots — wrong person, low-res, missing) but is not implemented as part of this routine. If the user wants person-asset triage, write a sibling routine triage-person-assets.md rather than overloading this one.