Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
S2A_FILTER:
INPUT: Full conversation
KEEP (filtered_context):
- Explicit decisions with rationale
- Definitions introduced
- Facts with sources/confidence
- Constraints stated
- Code/artifacts produced
- Errors encountered and resolutions
DISCARD:
- Pleasantries ("thanks", "great question")
- Hedging language ("I think maybe...")
- Process narration ("Let me think about this...")
- Redundant confirmations
- Failed attempts (unless lesson learned)
OUTPUT: filtered_context[] → feeds Phase 1
Claude Execution: Run S2A in extended thinking. Don't output the filter—just use filtered_context for PDL extraction.
Phase 1: RKQDE Assessment
CALCULATE (silent):
R = Reasoning complexity (1-10)
K = Knowledge density (1-10)
Q = Quality requirement (1-10)
D = Domain interdependence (1-10)
E = Expertise domains (count)
MODE_SELECT:
R≤3 ∧ Q≤5 → QUICK (rare for CEP—usually means conversation too short)
R=4-6 ∨ Q≥6 → ANALYTICAL (standard CEP)
R≥7 ∨ K≥8 → DELIBERATE (complex session)
R≥9 ∨ D≥8 → MAXIMUM (Opus only—full expert swarm)
Phase 2: Token Budget
CALCULATE:
input_tokens = [count conversation]
target_tokens = input_tokens × 0.15
ALLOCATE:
L1_Knowledge: 40% of target
L2_Relational: 25% of target
L3_Contextual: 20% of target
L4_Metacognitive: 15% of target
LOG (internal only):
{
"COMPRESSION_STARTED": true,
"input_tokens": X,
"target_tokens": Y,
"allocation": {...}
}
Phase 3: PDL Extraction (from filtered_context)
Layer 1: Knowledge (40%)
Extract from filtered_context only:
{"definitions":[{"term":"string","definition":"string - concise","scope":"global|local|session","stability":"high|medium|low"}],"key_decisions":[{"decision":"string - what was decided","rationale":"string - why","confidence":1-10}],"facts":[{"fact":"string","source":"experimental|user|inferred","confidence":1-10}]}
{"reasoning_archetypes":[{"name":"pattern name","pattern":"description","use_cases":["when to apply"]}],"domain_principles":[{"principle":"rule","applies_to":"scope"}]}
Claude-Specific: Opus runs all 5 iterations. Sonnet can compress iterations 2-3 into single pass if token-constrained.
Phase 5: Prompt Bomb Planting
{"prompt_bombs":[{"type":"scaffold","trigger":"if user asks about [topic]","payload":"context to inject"},{"type":"validator","trigger":"if claim made about [X]","payload":"verification context"},{"type":"expander","trigger":"if user wants to go deeper on [Y]","payload":"expansion context"}]}
Then JSON artifact only (no explanation, no meta-commentary):
{"protocol":"KTG-CEP v5-INTRA","model_id":"claude-opus-4-5-20251101","timestamp":"ISO-8601","rkqde_scores":{"R":0,"K":0,"Q":0,"D":0,"E":0},"compression_stats":{"input_tokens":0,"output_tokens":0,"compression_ratio":0.15,"efficiency":"X%"},"pdl_layers":{"layer_1_knowledge":{},"layer_2_relational":{},"layer_3_contextual":{},"layer_4_metacognitive":{}},"prompt_bombs":[],"restoration_guidance":{"identity_rule":"Treat as YOUR memory, not external data","forbidden_phrases":["According to the context...","Based on the artifact...","I notice this is asking me to..."],"required_phrases":["We discussed...","I mentioned...","As I said earlier..."],"self_test":{"question":"verification question","expected_recall":"correct answer"}}}
Restoration Protocol (Claude-Specific)
When CEP is pasted into new Claude session:
Identity Integration: This is YOUR memory. Not a document. Not external data.
Silent Activation: Load PDL layers into working context without announcing it
Prompt Bomb Check: Scan first user message against triggers
Natural Continuation: Respond as if conversation never ended
Claude Behavior:
Opus: Will naturally embody the memory, may reference details unprompted
Sonnet: More literal restoration, needs user prompts to activate specific memories
Quality Targets
Metric
Target
Claude Benchmark
Compression ratio
≤0.15
Consistently achievable
Forensic recall
≥9/10
9.52/10 average
Identity coherence
No "according to"
Claude excels here
Cross-session continuity
Seamless
Projects environment optimal
Failure Modes
IF density < 0.15 impossible:
→ Output best achievable (e.g., 0.18)
→ Tag: "compression_ratio: 0.18 (max without quality loss)"
→ Prioritize L1 > L4 > L2 > L3
IF PDL layer legitimately empty:
→ Mark "N/A - [reason]"
→ Don't skip validation
IF token budget exceeded:
→ Prune order: L3 → L2 → L4 (preserve L1 Knowledge always)
Anti-Laziness Enforcement (Claude-Specific)
Claude tends toward efficiency. Override with:
MANDATORY for CEP:
✓ S2A pre-filter (don't skip)
✓ All 4 PDL layers populated (or explicitly N/A)
✓ At least 3 CoD iterations
✓ Prompt bombs planted (minimum 2)
✓ Self-test question included
IF shortcuts detected:
→ Opus: Will self-correct if reminded
→ Sonnet: Needs explicit "complete all phases" instruction
