journal-cover-letter-skill-trainer

Journal Cover Letter Skill Trainer v0.2.0

Improve the journal-cover-letter-skill through controlled experiments. Learn transferable editorial reasoning from expert letters without copying their wording, treating them as factual authorities, or overfitting one manuscript.

Non-negotiable principles

• Treat the manuscript and author-confirmed case materials as the factual source of truth. An expert letter is a benchmark for editorial decisions, not a gold standard.
• Generate the baseline before exposing its generator to the expert letter.
• After the expert is revealed, never reuse that evaluator context for a claimed-blind candidate generation. Start a fresh generator context for every round.
• Change the target route for a stated reason. Do not patch only the current letter.
• A single case may create an experimental candidate, but cannot by itself prove a general improvement.
• Protect every route that was not targeted. A Research improvement must not silently weaken Review or Bibliometrics, and vice versa.
• Preserve the target Skill's name. Version the candidate and release; do not rename the Skill to encode a version.
• Keep private manuscripts and real letters outside public source trees and release packages.
• Never describe a comparison as objective proof. Report the rubric, judges, isolation level, evidence, uncertainty, and human decision.

1. Determine the run mode

Set:

training_intent: STRENGTHEN_EXISTING_ROUTE | PROPOSE_NEW_ROUTE
target_route: ORIGINAL_RESEARCH | REVIEW_SYNTHESIS | BIBLIOMETRICS | user-defined route
storage_mode: SESSION_ONLY | LOCAL_PRIVATE_CORPUS
isolation_level: STRICT_AGENT_ISOLATED | FRESH_CONTEXT_BLIND_GENERATION | HUMAN_GATED_REVEAL | REFERENCE_CONTAMINATED
promotion_policy: HUMAN_APPROVAL

Use STRICT_AGENT_ISOLATED when host-managed workers or API requests enforce separate contexts and controlled inputs. Use FRESH_CONTEXT_BLIND_GENERATION when a human starts a new temporary chat or conversation for every baseline and candidate generation, supplies only the candidate Skill plus manuscript-side materials, and verifies that memory, project files, custom instructions, and inherited conversation do not expose the expert. Use HUMAN_GATED_REVEAL when the expert is withheld until the first baseline is sealed but later candidate generation remains in the revealed conversation. If any claimed-blind generator has already seen or been quoted material from the expert letter, use REFERENCE_CONTAMINATED for that round.