placard

原文涵盖的主题

• Placard — per-artifact digests + recursive project maps
• When to use
• Reading protocol (when consuming an already-digested project)
• Digesting protocol (when creating/refreshing digests)
• Tiers
• Relationship to knowledge graphs
• Schema

Placard makes a project cheap for an LLM to understand, on two planes: Navigation plane (cheap): every content-bearing directory gets a PROJECT.placard.md map. Each map says, for each child directory, why you'd descend into it (a one-line dirsummary rolled up from the child). Walk the tree top-down paying only summary-sized tokens until you commit to a file. Content plane (heavy, on demand): each expensive artifact gets a faithful digest beside it (foo.pptx → foo.pptx.placard.md) that carries a summary AND a ## Full text section holding the cleaned, faithful, once- and-for-all text extraction of the original. Read the digest instead of re-ingesting the original — extract once, read many. Open the original only for pixels, exact layout, or a figure's actual content. Read map → (descend maps) → digest (summary + full text) → original, opening an original only when a digest's escalation contract says it is justified. Placard is non-destructive: it never edits originals and never mutates a shared store. It only adds .placard.md sidecars and PROJECT.placard.md maps. The CLI (placard.py) does the deterministic work — hashing, staleness, extraction and cleaning (digest), crash-safe publis

这里只展示节选，请以完整原文和当前要求为准。