Prompt architecture: single-owner rule for instruction content across the ~45k-word prompt surface #1252

Activity

1. waleedkadous

   added

   area/cross-cuttingTouches multiple areas — needs coordinated handlingTouches multiple areas — needs coordinated handling

   on Jul 25, 2026
2. waleedkadous commented on Jul 27, 2026

   

   waleedkadous

   on Jul 27, 2026

   ContributorAuthor

   More actions

   Architect guidance for the spec phase (from Waleed, 2026-07-27)

   Basis: this work is grounded in Anthropic's post The new rules of context engineering for Claude-5-generation models. Read it before writing the spec — the single-owner rule and fleet tiering in this issue's body are applications of its guidance.

   Scope reminder — the full prompt surface, not just CLAUDE.md. The trims must cover:

   • CLAUDE.md / AGENTS.md (and the hot-tier governance injections)
   • the architect role docs
   • the builder prompt surfaces (spawn prompts, porch phase prompts, protocol docs delivered to builders)
   • the consult prompts (what each CMAP lane model receives)
     Both trees where applicable (codev/ and codev-skeleton/).

   The hard problem is measurement. A particular challenge: how do we know a trim helped (or at least didn't hurt)? The spec must propose a measurement approach for the impact of prompt changes, not just an inventory and a restructuring plan. Think about what is observable per-surface (e.g. instruction-compliance incidents, gate-rejection rates, CMAP round counts, tokens per phase, duplication counts) and what a before/after comparison would look like. A trim proposal without a way to evaluate it is not accepted.

   Field input from a production Fable-5-generation deployment (an external adopter running long multi-lane programs; details scrubbed, signal preserved) — treat these as evidence for the spec, not requirements:

   • Fable-class models are uncalibrated for cost by default, not less capable: they maximize verification depth and delegation breadth unless given an explicit budget — and they honor stated constraints precisely; they just never invent them. Unbudgeted prompts produced unbounded process (audits-of-audits, compounding review layers); a budget stated in the prompt was honored immediately.
   • Two externally-imposed rules that worked on first contact and are candidates for default role-prompt content: 'ceremony is priced per artifact' (full audit apparatus only where the artifact warrants it; otherwise one time-boxed review pass) and 'escalate, don't widen' (disclose out-of-brief findings with receipts; fixing them unasked is the failure mode even when the fix is right).
   • Prompt templates should carry budget fields (scope boundary, review-pass count, time-box) — this maps to the fleet-tiering axis: stronger models need budgets more, not less.
   • Strong models mint dense internal vocabularies (finding codes, round letters) that make reports opaque — an explicit legibility contract in role prompts is a candidate rule.
   • (Their infra recommendations — verification primitives, status truthfulness, liveness detection — are Tower/porch feature work, out of this issue's scope.)
3. waleedkadous commented on Jul 28, 2026

   

   waleedkadous

   on Jul 28, 2026

   ContributorAuthor

   More actions

   Outcome: harvested, not fully merged

   What shipped (PR #1281, the minimal harvest — merged as aa534081):

   • The 17-file drift reconciliation — the one live bug this project surfaced. The codev/ shadow copies (which outrank the installed skeleton in the four-tier resolver) had drifted from codev-skeleton/ in 17 files, so the served SPIR/ASPIR builder prompts were missing their entire Multi-PR Mechanics and Verify Phase sections. All 17 are now byte-identical to the skeleton, with architect rulings E1 (max_iterations: 3) and E2 (check cwds moved to .codev/config.json porch.checks) applied.
   • The knowledge artifacts: the 77-row shadow-tree audit with per-file rulings, the word-count baseline + after-measurement (N1: 21,856 → 20,324 served always-on words, −7.0%), the behavioral baseline (B1: 51.88% REQUEST_CHANGES, n=160, self-excluded), the measurement tooling (scripts/measure-prompt-surface.sh + prompt-behavior-metrics lib/CLI/test), and the full project review at codev/reviews/1252-prompt-architecture-single-own.md.

   What did NOT ship, and why: the structural enforcement machinery — shadow-tree deletion with equivalence proofs, the scar-rule registry with line-exact CI enforcement, the ownership map with completeness contract, dedup via shared partials — was built, fully reviewed (PR #1278, closed unmerged, preserved on builder/spir-1252), and then deliberately declined at the pr gate on risk/reward: a successor project will first shrink the prompt surface radically (judgment-style rewrite, >50% reduction target, frontier-models-only), and enforcement should be rebuilt fit-for-purpose around the post-shrink surface rather than carried across it.

   Pointers:

   • Successor charter: issue Prompt-surface judgment rewrite: acceptance = blog-principles conformance per file (size is reporting-only) #1280 (radical prompt-surface shrink; carries forward the scar-registry concept, the eight ratified rule wordings, the ownership-map design, and this project's audit + baselines as its "before" measurement).
   • Porch bug found during close-out: issue porch: .codev/config.json command override is silently ignored for forge-intercepted checks (pr_exists) #1282 (.codev/config.json porch.checks command override silently discarded for the forge-intercepted pr_exists check — checks.ts:262 intercepts after the override merge; only cwd/skip survive).
   • Reference branch: builder/spir-1252 with PR [Spec 1252] Prompt architecture: single-owner rule across the prompt surface #1278 (closed unmerged) — the complete machinery and its review history.
4. waleedkadous

   closed this as completedon Jul 28, 2026