anentrypoint codebase-vision-report
Analyzes a codebase by rendering it as dense images (via gitoutput's --images mode, which uses mdpix's bitmap-atlas renderer), reading every rendered page, and producing a report grounded in what was actually read. For a text-only, non-image variant, see codebase-text-report — same workflow shape, reads gitoutput's default .txt chunks instead of rendering pages.
Arguments
Invoke as /codebase-vision-report <target> :: <prompt>, or Skill(skill="codebase-vision-report", args="<target> :: <prompt>").
target— a local directory path or a remote Git URL (anythinggitoutputaccepts as itssourceargument).prompt— the question or report focus (e.g. "find security issues", "summarize the architecture", "does this codebase handle X correctly").
Split args on the first ::. If no :: is present, treat the whole string as target and use a default prompt: "Summarize this codebase's architecture and key components."
Optional trailing flags in args (space-separated, after the prompt or anywhere in target): --image-profile <claude|gpt|gemini|grok> (default claude — this skill runs inside Claude Code, so claude is almost always correct), --exclude-pattern <glob> / --include-pattern <glob> (repeatable, passed through to gitoutput to scope a large codebase).
Step 1 — Generate images
Always invoke gitoutput at its densest defaults — claude profile and minification on — so the codebase packs into the fewest possible pages: never pass --no-minify, and only override --image-profile away from claude when the user explicitly names a different target model. Both are gitoutput's own defaults for --images, so simply omitting --image-profile/--no-minify already gets the minimum-page-count behavior; this is a deliberate choice, not an oversight, and following the command below exactly preserves it.
Run gitoutput via npx with --images --json so the result is machine-parseable, writing to a fresh output directory (use a timestamp or the target's basename plus a short random suffix to avoid colliding with a prior run of this skill in the same working directory — never reuse a bare default name across invocations):
npx github:AnEntrypoint/gitoutput "<target>" --images --json --output "./codebase-vision-<slug>-<suffix>" [--image-profile <profile>] [--exclude-pattern <p> ...] [--include-pattern <p> ...]
Capture stdout and parse it as JSON: {outDir, pageFiles, factsheetFile, manifestFile, pageCount, truncated}. A non-zero exit or a JSON body containing error means generation failed — report the error to the user directly, do not proceed to steps 2-3 on a failed generation.
If pageCount is 0 (empty/trivial target — nothing matched, or the target had no analyzable content), stop here and report "nothing to analyze" rather than attempting to read zero images or fabricating a report.
If truncated is true, the codebase exceeded the chosen profile's page budget and only the first pageCount pages are present — note this explicitly in the final report (e.g. "This analysis covers the first N pages of the codebase; it was truncated because the codebase exceeds the <profile> profile's page limit. Narrow the scope with --include-pattern/--exclude-pattern for full coverage."). Never present a truncated analysis as if it were complete.
Step 2 — Batch-read every page
Read each path in pageFiles, in order, using the Read tool directly on the PNG path — no separate vision API call is needed; this skill runs inside a vision-capable Claude Code session and Read already renders images.
For large pageCount (roughly above 20-30 pages, since each page costs real context), read in batches: read a batch of pages, write a running note of what each batch covered (file names visible, key structures noticed) before moving to the next batch, so the full codebase is covered without holding every page's full detail in context simultaneously. Do not silently skip pages to stay small — if the codebase is large, read all of it in batches rather than sampling a subset and reporting as if it were the whole thing.
If factsheetFile is non-null, also Read it — it holds exact-value strings (hashes, UUIDs, URLs, numeric IDs) that vision models are prone to misreading from the rendered pages. Cross-reference any exact value the report cites against the factsheet rather than trusting what was read off a page image.
Step 3 — Write the report
Synthesize a report that directly addresses prompt, grounded in what was actually read across all pages (and the factsheet for exact values) — never a generic description that could have been written without reading the pages. Structure:
- A one-paragraph direct answer to
prompt. - Supporting detail organized by the codebase's actual structure (as seen across the pages), citing specific files/functions/patterns observed.
- A truncation/coverage note if step 1 flagged
truncated: true. - Any exact values (hashes, IDs, URLs) cited must match the factsheet, not a page-read guess.
Write the report to a file (<outDir>/report.md, next to the generated images) and also present it directly in the response. Leave the generated outDir (images, factsheet, manifest, report) on disk for the user to re-inspect — do not delete it after the report is written.