Shaping Frame is a session-level cognitive frame for Claude Desktop that tracks epistemic weight during complex working sessions — what is still a hypothesis, what has been decided, and what Claude generated but you have not yet validated. It operates every turn, silently, and surfaces only when material starts crossing a threshold without your explicit approval. If you have ever finished a long Claude session and realized a phrase you used early on had quietly become an architectural constraint by the end, Shaping Frame is built for exactly that problem.Documentation Index
Fetch the complete documentation index at: https://mintlify.com/XxYouDeaDPunKxX/shaping-frame-for-claude/llms.txt
Use this file to discover all available pages before exploring further.
The Core Failure Mode: Silent Promotion
The failure mode Shaping Frame targets is silent promotion. A phrase, suggestion, assumption, or imported fragment gets repeated, reused, or built upon across turns until it starts behaving like a decided constraint — even though you, the operator, never explicitly approved it. It does not announce itself. It accumulates through conversational inertia: cited once, referenced again, used as a premise in the next proposal, and suddenly acting like a decision. Silent promotion is especially dangerous with three kinds of material:- AI-generated proposals — Claude produces a suggestion in response to your input. You do not reject it. It gets referenced again three turns later as if it were agreed.
- External material you imported — a document, transcript, or note you brought into the session. It is relevant because you brought it. That relevance is not the same as approval.
- Model priors — Claude’s training patterns, conventions, and defaults. They enter the session without being introduced, carry no tag, and sound authoritative.
What Shaping Frame Is
Shaping Frame is a session-level cognitive frame. Its job is to track epistemic weight while the conversation is still moving: ideas, constraints, imported material, Claude-generated proposals, model priors, and operator decisions. It separates material by source class — who or what introduced it — and tracks movement through weight states, from a raw spark through candidate, tracked, and crystallized, with explicit handling for rejected material, quarantined elements, and items that need verification before they can ground a decision. The frame operates every turn. It does not emit output on every turn. It fires a checkpoint when material crosses an epistemic threshold: the point where an unvetted element starts behaving like a decision. Between those crossings, silence is normal and expected. The frame’s operating model consists of two layers:| Layer | Artifact | Role |
|---|---|---|
| Always-on floor | CUSTOM_INSTRUCTION.txt (in Instructions for Claude) | Active every turn, every session |
| Full frame | shaping-frame.skill | Loaded for working sessions; provides the full rule set, source classes, weight states, checkpoints, and reference files |
What It Is Not
Understanding what Shaping Frame is not is as important as understanding what it is. The frame is narrow by design.- Not a file tool. It does not read, write, or manage files. It tracks epistemic weight in the conversation layer.
- Not a code tool. It has no execution environment and produces no code artifacts.
- Not a general prompt style. It is not a persona, a tone modifier, or a way of asking Claude to be more careful. It applies specific threshold logic to specific source classes.
- Not a memory layer. It does not persist state between sessions. Each conversation starts fresh, with the frame active from turn one.
- Not a writing template. It does not structure your output or enforce a document format. It tracks what is behind your output, not the shape of the output itself.
The Package
Shaping Frame ships as two artifacts that must be used together.shaping-frame.skill — the skill archive. It contains SKILL.md (the full frame), plus three reference files: frame.md (a compact reminder, read at session start and on drift risk), dogana.md (the checkpoint checklist, read when a threshold fires), and snapshot.md (the surface snapshot format, read when a surface trigger fires). Upload this to Claude Desktop via the Skills panel.
CUSTOM_INSTRUCTION.txt — the anchor. This is not part of the .skill archive. It must be added separately to Instructions for Claude in your Claude Desktop settings. Its purpose is narrow but critical: it tells Claude to execute a full read of the skill and keep the frame active as a working posture. Without it, Claude may treat the skill like ordinary reference material — reading it once, summarizing it, and applying only a distilled version. The anchor prevents that compression.
Both artifacts are required. The skill provides the full frame. The anchor ensures the frame runs as intended.
Host Limits
Claude Desktop does not provide a hook layer for this skill. There is no automatic pre-write interception and no persistent runtime enforcement guard. The frame is reconstructed through the conversation context and the custom instruction anchor.In long or fast sessions, some threshold crossings will be missed. The frame reduces drift by making threshold crossings visible — it does not eliminate drift. This is a host constraint, not a skill defect.
Quickstart
Install the skill and activate the custom instruction anchor in under five minutes.
Epistemic Weight
Understand the weight states, thresholds, and checkpoint logic at the core of the frame.