TT Claude Memory

feedback_compound_launch_prompts_amplify_context_degradation_risk

Properties1
descriptionCompound multi-sub-task launch-prompts authored at >75% context show empirically higher defect rates than single-action prompts at fresh context; split + per-sub-task author-side mirror catches defects pre-dispatch

2 min read

When authoring launch-prompts for dispatch to fresh autopilot sessions, prefer single-action scope over compound multi-action scope, especially when the authoring session is at >75% context. Compound launch-prompts amplify substrate-error risk proportionally to (sub-action count × context-degradation).

Why: empirical precedent from BATCH improvement 20 (authored 2026-05-22T21:00 BST at 91% parallel-session context as a 3-action compound: R8 review + 4 email polish + Sprint 1 server launch-prompt authoring). Outcome:

  • Confirmed substrate error at §0 Check 1 (partner-outreach-drafts path drift, 5 occurrences pointing at docs-admin/external-partners/... when actual file at docs-admin/)
  • Transcript showed degradation symptoms (table-repetition, partial-then-full rendering of methodology blocks)
  • Forced deprecation 9 hours later via Rich-directive 2026-05-23T06:47 BST + split-and-reauthor pattern (batch-imp-19 v2.0 R8 + batch-imp-21 v1.0 Sprint 1 server)

Contrast with batch-imp-15 / 16 / 17 / 18 — all single-action, all authored at fresh-or-near-fresh context, all clean enough to dispatch without rewrite. The variable that changed: compound design + degraded context, not the underlying methodology.

How to apply:

  • Above ~75% context: restrict launch-prompt authoring to single-action scope. If a 2+ action compound is needed, defer to a fresh session.
  • Above ~90% context: defer ALL launch-prompt authoring to fresh session. Even single-action prompts show substrate-error rates at that degradation level.
  • During launch-prompt authoring at any context level: apply §13.5 author-side mirror — verify every test -f / ls path in the §0 PRECONDITION GATE BEFORE committing the file. Compound prompts have N× as many paths to verify.
  • When you catch yourself merging 3 user-requested actions into one compound: pause and propose split + defer instead. The discipline cost of refusing the compound is small; the cost of dispatching a defective compound (or having to deprecate-and-rewrite it later) is high.

Related: feedback-research-artefact-forward-traceability (frontmatter discipline) + feedback-post-commit-body-content-verification (POST-COMMIT BODY-CONTENT VERIFICATION codifies the recursive landing check) + feedback-substrate-search-locations-file-is-canonical (substrate-sweep discipline) + arch-state §15 A-283 (the deprecation-and-rewrite batch where this pattern was confirmed) + batch-imp-20 v1.1-deprecated (the empirical precedent).

Companion: when a compound is genuinely necessary (e.g., the user explicitly requests 3-in-1 and refusal is not appropriate), apply per-sub-action MEMORY hot-reload checkpoints + per-sub-action §0 dry-run BEFORE moving to next sub-action. That makes the compound a sequence-of-singles internally even though it’s authored as one file.

Graph View

Backlinks