context-as-images
Skill0 starsChanged 16 days agoMITView on GitHub
Use when designing, tuning, or debugging an LLM agent pipeline, local proxy, or context-caching layer that resends large, low-precision context (system prompts, tool docs, chat history, logs, file dumps) on every request to a vision-capable model, and token cost from that resend is a concern; also use when deciding whether an image-based compression trick is worth adding, or when such a trick is producing wrong recall of exact values (IDs, hashes, secrets).
In this repository23 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- agent-orchestrationSkill
- ask-for-approvalSkill
- claim-verification-loopSkill
- code-councilSkill
- code-reviewSkill
- grillingSkill
- grow-test-coverageSkill
- investigateSkill
- loop-until-verifiedSkill
- plan-and-buildSkill
- prompt-cachingSkill
- proofreadSkill
- review-allSkill
- review-before-mergeSkill
- review-panelSkill
- to-questionnaireSkill
- track-review-findingsSkill
- trim-contextSkill
- unslopSkill
- verify-findingsSkill
- whatnowSkill
- wizardSkill
The file
We split this skill into the rules it gives an agent
5 rules. About 0 are common, 5 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 19
Use when a pipeline resends the same bulky, low-precision block every turn (system prompt, tool docs, old history, log dumps) to a model …
- Distinctive: line 23
Do NOT use when:
- Distinctive: line 38
Check profitability on what remains: image only where its image-token cost beats its text-token cost, not as a blanket policy.
- Distinctive: line 44
Keep a text-only escape hatch for any consumer that cannot read images or where exact recall outweighs cost.
- Distinctive: line 47
Stop condition: ship the substitution for a content type only once step 5 shows a measured net token reduction on real requests.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Context as images
Overview
A vision-capable model is billed for an image by its pixel dimensions, not by how much text is packed inside it. Dense, repetitive text (code, JSON, logs) can render at roughly 3+ characters per image token, versus about 1 character per text token. Rendering the right content as an image page instead of text can cut that content's token cost by more than half. It is a compression trick, not a universal one.
When to use
Use when a pipeline resends the same bulky, low-precision block every turn (system prompt, tool docs, old history, log dumps) to a model that reads images.
Do NOT use when:
- content is sparse prose (near 1 char/token as text): imaging it can cost MORE. Check the ratio, do not assume.
- content has byte-exact values the model must recall precisely: hashes, IDs, secrets, UUIDs. Image recall of exact strings is lossy (as low as 0/15 correct on some models, 13/15 on the best). Keep those in text.
- the receiving stage cannot reliably read images (text-only subagent, non-vision model, strict API mode).
The pattern
- Identify the bulky, low-precision, resent-every-turn block(s); recent turns and anything byte-exact are excluded.
- Pull byte-exact identifiers out first; keep them in accompanying text, never inside the rendered image.
- Check profitability on what remains: image only where its image-token cost beats its text-token cost, not as a blanket policy.
- Render the block to an image page and substitute it for the original text.
- Measure actual input tokens per request, with and without the substitution, using a real token counter, not an assumption.
- Keep a text-only escape hatch for any consumer that cannot read images or where exact recall outweighs cost.
Stop condition: ship the substitution for a content type only once step 5 shows a measured net token reduction on real requests; drop it if savings are flat or negative.
Quick reference
| Content | Image it? |
| Dense code / JSON / logs, resent every turn | Yes |
| Sparse prose, one-off content | No, verify ratio first |
| Hashes, IDs, secrets, exact values | No, keep as text |
| Old history beyond the last few turns | Often, if dense |
| Recent or active turns | No |
Common mistakes
- Imaging without measuring. Images are not always cheaper for sparse prose; measure real tokens first.
- Imaging exact values. Vision recall of hashes or IDs is unreliable; keep those in text.
- No escape hatch. A stage that cannot read images must fall back to text, not drop content.
Source
- Repo:
teamchong/pxpipe, commit8ba82b713a1e823bc1c09b7a68e47f63caa7b426. - License: MIT. Wording adapted with attribution; no code or prose copied.
- File read:
README.md(root). - Adapted: the mechanism of trading text tokens for image tokens on bulky, low-precision context, its profitability gate, and its byte-exact-value escape hatch, restated in original wording.
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
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.