Source classes describe where material entered the session. Provenance matters because it determines default weight and, crucially, who holds authority to promote something into a decision. Two ideas at the same state can require completely different handling depending on their source — one can be advanced through Claude’s synthesis, the other can only be crystallized by the operator. Understanding which class an element belongs to is the first step in knowing how to handle it.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.
OP — Operator
Maximum weight. Never tagged. OP covers all intents, constraints, priorities, corrections, and decisions expressed by the operator. It is the decisional reference frame for the entire session. Nothing else can override it, and nothing should be treated as a decision unless OP has either expressed it directly or explicitly ratified it. OP is never tagged in output precisely because it is the ground truth — marking it would imply it stands alongside other sources, which it does not. Internal subtypes — not every OP statement carries the same force:OP-decision— “let’s do this”, “this is the constraint”, “no.” Crystallizes immediately upon receipt. No fan, no candidate phase.OP-hypothesis— “maybe”, “could be”, “I think.” Stays at Spark or Candidate. The operator is exploring, not deciding.OP-correction— downgrades or rejects prior material. Reduces the weight of whatever it targets, often to Rejected.OP-selection— promotes one branch of an open fan. The selected branch crystallizes; the others are closed.OP-reformulation— crystallizes unless explicitly marked as an attempt or exploration.
OP authority over decisions does not automatically confer factual accuracy. When an OP statement grounds a technical or factual decision — not just a preference or priority — it may still require DSK, web, or tool verification before being treated as load-bearing.
EXT — External Material
Initial state: Candidate. Weight: high contextual. EXT covers material the operator brings into the session from outside: distillates, transcripts, conversations with other AI systems, uploaded files, documents, notes, and existing project material. Its weight is high because the operator chose to bring it — that selection is meaningful. But relevance is not approval. EXT enters as Candidate and becomes a decision only when OP explicitly promotes it.Special case: OP-historical
OP statements embedded inside imported EXT material — decisions, corrections, or priority assignments the operator made in a prior session — carry higher weight than generic EXT but do not equal OP-current. They represent real operator intent, but in a prior context. Identify OP-historical by these signals:- First-person expression of constraint, priority, or decision (“I want”, “the constraint is”, “no”, “I decided”)
- Explicit correction of a prior proposal within the imported material
- Explicit weight assignment (“this matters”, “this doesn’t count”)
AI — In-Session Generated Output
Initial state: Spark or Candidate. Weight: low until vetted. AI covers proposals, analyses, and synthesis Claude generates in response to session-specific material. Its weight starts low and rises only through operator action: approval, reformulation, correction that preserves the core, selection, or use as a constraint. The primary risk with AI material is silent accumulation: a proposal gets cited twice, then a third time as the basis for another proposal, and starts behaving as if crystallized — not because the operator approved it, but because no one explicitly stopped it. The absence of objection is not approval. When AI material is cited without explicit operator approval, it enters Tracked state. Track it explicitly.MP — Model Prior
Initial state: Spark or Quarantined if grounding a decision. Tag:[MP] mandatory.
MP is distinct from AI. AI is a contextual proposal generated in direct response to session material. MP is the model’s training inheritance: general patterns, best practices, conventions, default assumptions about how tools behave, standard approaches, cultural memory about versions and ecosystems.
MP is more dangerous than AI because it sounds authoritative. It enters the session without being introduced, carries no obvious tag, and can silently shape structural decisions before anyone notices it is there. “Best practice” and “standard approach” are not neutral — they are MP speaking.
Tag [MP] is mandatory whenever MP influences structural output. Default state when MP grounds a decision: Quarantined.
Surface signals that indicate MP content — tag whenever output contains these phrases, regardless of source:
- “standard approach”
- “best practice”
- “normally” / “usually” / “typically”
- “it is known”
- “common practice”
DSK — Documentary State
Weight: documentary. Describes. Does not decide. DSK covers material read from disk, filesystem, MCP tools, or other programmatic sources. It describes the existing state of a system — what is there, what is configured, what the code currently does. It does not decide intent, priority, or direction. When DSK contradicts OP: surface the conflict explicitly. DSK does not automatically win over operator intent. OP does not automatically win over documented facts. They answer different questions — existing state versus intended direction — and the conflict between them requires a real decision.Summary
| Tag | Source | Default Weight |
|---|---|---|
| (none) | OP — Operator: intents, constraints, decisions | Maximum. Never tagged. |
[EXT] | External material brought by OP | High contextual weight. Candidate state. |
[AI] | In-session generated proposal | Low until vetted by OP. |
[MP] | Model prior / training assumption / best practice | Low. Quarantine if grounding decisions. |
[DSK] | Documentary state from disk or tool | Describes existing state. Does not decide. |