agentleFS
DITlieDDITlieD/ELAI-archive/.agents/skills/see/SKILL.md

see

Skill24 starsChanged 2 months agoMITView on GitHub

Harness leg-1 image preprocess for the J-1 perception gateway. Triggers when the user is about to attach an image to a message, asks to read or analyze a screenshot / diagram / mockup / photo / figure, says "look at this image", "read this screenshot", "analyze this mockup", "check this figure for the fine print", or invokes /see by name. Runs scripts/preprocess.py locally BEFORE the image is attached so the bytes are conditioned (resized to the target model sweet spot, JPEG q85, dense-UI overlapping tiles, global thumbnail) instead of sending raw pixels. Pillow absent degrades to passthrough plus a one-line stderr warning, never a silent fallthrough. The Rust PerceptionGateway is the runtime leg; /see is the harness leg and they share semantics, not code.

6 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

2 rules. 0 are common, 1 is distinctive, and 1 is very specific to this project.

  • 0

    Common

    Most of the 4 edited versions we read keep these.

  • 1

    Distinctive

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Does NOT fire on:

    line 23
  2. Very specific:

    Run scripts/preprocess.py --input IMAGE --target-model Codex|gpt|qwen-vl --output OUT.jpg BEFORE attaching the image.

    line 32

“Common” is measured: kept by at least 6 in 10 of the 4 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

/see — harness leg-1 image preprocess

/see is the HARNESS-side leg of the J-1 perception gateway (docs/thefinalglance/J-plan/J-1_perception_gateway.md). It conditions an image LOCALLY, with stdlib + pillow, BEFORE the image is attached to a Codex message — so the model receives bytes already sized to its sweet spot instead of raw pixels. The Rust PreprocessSpec (crates/orchestrator/src/perception/preprocess.rs) is the RUNTIME leg on the sidecar; the two share SEMANTICS, not code. No FFI, no shared module, no import of preprocess.rs from preprocess.py.

WHEN THIS SKILL FIRES

Fire /see (set LEG-1 PREPROCESS mode for the rest of the attachment) when the user:

  • invokes /see this image, /see with a path, or /see before attaching a screenshot / diagram / mockup / photo / figure.
  • says "look at this screenshot and tell me what's wrong", "read this architecture diagram", "analyze this UI mockup", "I'm attaching a photo — what's in it", "check this figure for the fine print".
  • is about to attach any image to a message for the model to reason over (drag-drop, paste, or path reference into a harness session).

Does NOT fire on:

  • pure-text questions ("what does the SSB single-writer rule mean?") — no image.
  • single-file code reads ("read cli/src/args.rs and fix line 88") — no image; that is a normal edit.
  • design-judgement asks on a mockup like "improve this UI design" — spec-eval owns the design verdict; /see only conditions the bytes that spec-eval then reasons over.
  • tasks already inside another skill's protocol (planning-workflow's own reads, paper-to-plan's source ingestion, incremental-eval's multi-artifact walk) — those skills own the read discipline; /see only fires when an image is the artifact.

LEG-1 PREPROCESS PROTOCOL

  1. Identify the target vision model the message is bound for (default Codex; override per dispatch).
  2. Run scripts/preprocess.py --input IMAGE --target-model Codex|gpt|qwen-vl --output OUT.jpg BEFORE attaching the image.
  3. Attach the conditioned output (the resized JPEG q85) to the Codex message. The thumbnail (STEM.thumb.jpg) and dense-UI tiles (STEM.tiles directory) are emitted alongside for dense-UI reasoning; attach them when fine text or small click targets matter.
  4. Cite the conditioning in the answer ("image preprocessed to EDGEpx longest side, JPEG q85, N tiles + 512px thumbnail") so the reasoning trace shows the bytes were conditioned, not raw.

Sweet-spot table (EXTERNAL figure SRC-3, design rationale ONLY — HD-1/ID-5)

These longest-side edges are EXTERNAL design rationale from J-HARNESS-GROUNDING-MAP.md SRC-3 (vlm-vision-research-summary), NOT ELAI-measured numbers. They MUST NOT appear in any user-facing claim or metrics TOML as a measured value. They are pinned as config with a dated source line; re-verify on model swaps.

target modelsweet-spot longest sideJPEG qualitytile overlapthumbnail
Codex-class1568px850.25512px
GPT-class2048px850.25512px
Qwen-VL-class1280px850.25512px

Resize is downscale-only (never upscale), aspect ratio preserved. PNG to RGB JPEG q85 is the conditioning step (~60% image-token savings, negligible quality loss — EXTERNAL figure SRC-3, design rationale only). Dense-UI tiling uses 0.25 overlap on each axis plus a 512px global thumbnail so the model has both the fine-grained tiles and the layout overview.

DEGRADED PASSTHROUGH (JC-4, stub_ban)

If pillow is absent (ImportError) OR any preprocess step fails (decode error, resize error, write error), the script forwards the raw input bytes UNCONDITIONED to the output path and emits ONE explicit stderr warning ("pillow not installed; /see degraded to passthrough — install pillow for leg-1 conditioning", or the specific failure). It NEVER silently sends raw bytes while claiming conditioning (plan stub_ban line 102-103; mirrors PreprocessOutcome::DegradedPassthrough in preprocess.rs:131-148). The script always exits 0 on the passthrough path so the harness can still attach the raw image; the warning is the audit trail. Pillow is verified before use (JC-4: pillow 12.1.1 confirmed on this machine; the absent path is the acceptance gate, not this machine's state).

BOUNDARY — what /see is NOT

  • /see is the HARNESS leg (local pillow). The Rust PerceptionGateway is the RUNTIME leg (sidecar, T2/T5). Same semantics, different language.
  • /see does NOT implement: the two-call perception scaffold (leg 3, T2), the hard-limit router (T3), the SSB event (T4), or legs 2/4 sidecar VLM (T5).
  • /see FEEDS B-24's output validator and E-36's document leg — it never re-implements either (GD-1, plan line 12).
  • /see does NOT add pillow as a Cargo dep, a sidecar dep, or a runtime Rust dep. Pillow is a Python harness-side dep only.

CITATIONS (cite, do NOT fork — GD-1 / stub_ban)

  • J-1 plan: docs/thefinalglance/J-plan/J-1_perception_gateway.md (T6 scope lines 74-77).
  • Rust PreprocessSpec: crates/orchestrator/src/perception/preprocess.rs (same sweet-spot consts, same JPEG q85, same tile_overlap, same thumbnail — different language).
  • SRC-3 vlm-vision-research-summary (EXTERNAL figures, design rationale only — HD-1/ID-5).
  • AGENTS.md anti-slopsquat (JC-4 pillow verify-before-use).

END /see SKILL

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.