What
The opus-5-prompting-interview workstream (#1699) operationalized the Opus 5 prompting guide's correction-shaped directives (I8-a/b/c, I17 family, I21). Two candidate audit rows raised by the corpus digests were not built, and the interview's merged triage (.work/opus-5-prompting-interview/validation/) is no longer on local disk, so whether they were consciously rejected or simply not picked is unconfirmed. Recording both as deferred-with-trigger rather than letting them vanish:
- Vision-workaround re-validation (guide, "Capability improvements" → Vision): "Re-validate any prompt-side vision workarounds you tuned for prior models; they may no longer be needed." Digest
02-capabilities.md proposed a re-validation pass / detection of prior-model vision scaffolding (transcribe-first scaffolds, single-shot image-read compensations). No catalog row covers it; I8 base (prior-model workarounds) is scoped fable-5, so an opus-5 target skips it. A 2026-08-08 sweep found no such scaffolding in this repo, so the value is consumer-facing only.
- Late-window defensive scaffolding (guide, "Capability improvements" → Long-context work): instruction following "stay[s] consistent throughout the window" on Opus 5, which the digest read as an argument against late-window re-anchoring prompts / instruction repetition / degradation-justified pre-emptive compaction. Candidate row: flag scaffolding justified on window-degradation grounds (cost-ground justifications stay). Detection is weak (justification prose, not a literal), so it would be
behavioral, opus-5-scoped.
Why deferred, not built
Both are behavioral, weakly detectable, and unproven in value; per ADR-0006's promotion-gate posture, each would enter as a model-scoped row with fixtures. Neither has an attested operative instance in this repo (measured 2026-08-08).
Triggers
- A consumer repo surfaces an actual instance of either shape (evidence of value).
- A second model guide states either claim (promotion-gate movement per
reference/criteria.md "Model scoping").
- The Opus 5 guide's raw-
.md body changes (catalog-wide recheck trigger fires anyway; MD5 baseline 8579d63fc9f793784b8c56320fd74e71, 11,225 bytes, byte-identical 2026-07-25 → 2026-08-08).
- The interview triage resurfaces (e.g. from the knowledge-corpus slice or a session transcript) showing an explicit rejection — then close this as decided.
Related
What
The
opus-5-prompting-interviewworkstream (#1699) operationalized the Opus 5 prompting guide's correction-shaped directives (I8-a/b/c, I17 family, I21). Two candidate audit rows raised by the corpus digests were not built, and the interview's merged triage (.work/opus-5-prompting-interview/validation/) is no longer on local disk, so whether they were consciously rejected or simply not picked is unconfirmed. Recording both as deferred-with-trigger rather than letting them vanish:02-capabilities.mdproposed a re-validation pass / detection of prior-model vision scaffolding (transcribe-first scaffolds, single-shot image-read compensations). No catalog row covers it; I8 base (prior-model workarounds) is scopedfable-5, so anopus-5target skips it. A 2026-08-08 sweep found no such scaffolding in this repo, so the value is consumer-facing only.behavioral,opus-5-scoped.Why deferred, not built
Both are behavioral, weakly detectable, and unproven in value; per ADR-0006's promotion-gate posture, each would enter as a model-scoped row with fixtures. Neither has an attested operative instance in this repo (measured 2026-08-08).
Triggers
reference/criteria.md"Model scoping")..mdbody changes (catalog-wide recheck trigger fires anyway; MD5 baseline8579d63fc9f793784b8c56320fd74e71, 11,225 bytes, byte-identical 2026-07-25 → 2026-08-08).Related
plugins/claude-config/skills/audit-instructions/reference/criteria.md(catalog these rows would join)plugins/playbooks/reference/model-adaptation/opus-5.md(behavior-delta chapter; carries neither claim, payload-discipline consistent)