Critical review of 20 representative prompts sampled from the522-prompt loom-builder dispatch corpus. Each prompt scored on structure, specificity, guard rails, measurability, and context completeness. Generated Jul 19, 2026.
The loom-builder prompt corpus (522 prompts, 20 analyzed in this report) has converged on a small set of structural patterns that distinguish high-leverage dispatches from low-leverage ones. The strongest prompts share six load-bearing elements: a ROLE/GOAL/PLAN echo handshake, a numbered 'Read FIRST' anchor list with status, a quantified success metric, RIGHT/WRONG code blocks for fragile APIs, an explicit budget calculation, and a copy-paste bash gate. The corpus-explorer S1/S2/S3e series and the ephesians4 series represent the current ceiling: they call out known gotchas, binding decisions, prior failures, and forbidden files. The sprint-1/2 and sol-clone-migration prompts demonstrate that handoff citation between dispatches is itself a pattern — later prompts compress by referencing earlier ones rather than re-stating context.
The most common anti-patterns are the absence of a falsifiable gate (replaced by 'comprehensive' or 'working') and the absence of a 'Read FIRST' anchor list, which together cause the 'file exists but doesn't work' failure class that the S3c dispatch was written to correct. Section-naming drift is the most common low-severity anti-pattern — prompts use 'Context', 'Read FIRST', 'INPUT FILES', and 'User intent (restated)' interchangeably, which makes small models pattern-match inconsistently. The lowest-scoring prompt in the sample, smoke.md, is a 3-line probe that has no role, no metric, and no gate; it should be retired from the active corpus.
The top 3 recommendations are: (1) lock the template section ordering and migrate older prompts in a separate pass, (2) make RIGHT/WRONG blocks mandatory for any prompt that touches an API likely to have moved since training, and (3) require a copy-paste bash gate on every prompt. The corpus is in a strong place — the median prompt scores 4.0+ across all five dimensions — but the gap between the strongest (5.0) and weakest (1.8) prompts is large, and closing that gap is the highest-leverage prompt-engineering work remaining.
Structural and content patterns that distinguish high-leverage dispatches from low-leverage ones. Frequency is out of 20 sampled prompts.
Prompts instruct the model to print ROLE + GOAL + PLAN (or PLAN + numbered build order) as the first output before any tool call. This forces the model to commit to a build order it can maintain as a checklist, and gives the orchestrator an early signal if the model misread scope.
Verbatim phrasing across ephesians4-* series, crawlerstudiov2-*, agents-md-enrich-batch, phase0-integration-phaseAB, sprint-1.md, sprint-2.md
Prompts supply a numbered 'Read FIRST' list of absolute file paths, in execution order. This short-circuits model spelunking and ensures the model sees binding context (specs, handoffs, schemas, proven house patterns) before touching source.
Typical: '1. CLAUDE_HANDOFF.md 2. spec.html ... 3. ddl.sql'. Files cited have status: binding, authoritative, ground truth, proven.
Prompts define done as a falsifiable check: all 12 spec assertions pass green, N/14 target files rewritten with >=15 non-empty lines, bun run build exits 0, tsc --noEmit exits 0. The metric appears in bold or in a dedicated 'Success metric' / 'Gates' block.
Quantified metrics appear in ephesians4, corpus-explorer, sprint-1/2, sol-clone migration, agents-md enrichment, phase0 integration, crawlerstudiov2-prisma-d1
For coding tasks, prompts include fenced ts/sh blocks showing the RIGHT way and the WRONG way side by side. These are surgical — they call out exactly the API call or command pattern that previously failed, so the model cannot regress to the wrong shape.
Bun.spawn vs in-process timeout, prisma adapter import path, npx vs bunx, module-level globalThis vs request-scoped singleton. Every RIGHT/WRONG cites a known prior failure.
Prompts include a budget formula: gates x 3 + files_read / 2 + files_write x 2 + buffer, and a hard cap (30, 60, 100, 150 calls). Some instruct: write output before exhaustive reads, or if you have not written output by tool call 3, STOP and write it.
Template's BUDGET CALCULATION block is reproduced across architecture-first-workflow, ephesians4-seo, crawlerstudiov2-prisma-d1, sprint-1, sprint-2
Prompts begin with STEP 0 — before any plan, the model must run 'ls -a', check for graphify-out/GRAPH_REPORT.md/CODEATLAS.md, or run a code-atlas query to discover repo structure. This is Mike's HEEC loop: Hypothesize -> Examine -> Execute -> Check.
Templates call it INTEL DISCOVERY; later prompts call it FLEET QUERY FIRST or HEEC LOOP. Always before planning.
Prompts explicitly enumerate gotchas from prior handoffs: incremental staging carry-forward already fixed, Conversation Date = created_at never exported_at, wrangler.jsonc must have nodejs_compat or every route 500s. These compress weeks of debug history into a don't-repeat block.
corpus-explorer S1, S2, S3c each carry a 'Known gotchas / invariants / Hard rules' section
Prompts forbid broad file reading: no ~ or relative paths, full absolute paths only, do NOT recursively list or read directories, maximum 20 file reads total, write output before exhaustive reads. The model is constrained to a bounded read budget.
DISPATCH SCOPE CONSTRAINTS appears as a hard append in every minimax-m3 dispatch. Smaller prompts bake it into RULES section.
Prompts end with a shell block of curl/grep/bun commands that prove success. The commands are designed to be runnable by a small model and produce a yes/no answer. 'curl -s ... | grep -c og-image # >=1' is a load-bearing gate pattern.
All ephesians4-*, corpus-explorer S1/S2, sol-clone migration, crawlerstudiov2-prisma-d1 have shell-verifiable gates
Prompts explicitly enumerate files the model must NOT modify: do NOT touch wrangler.toml, content markdown, api/register.ts, wrangler.jsonc, package.json, migrations. Do NOT run wrangler deploy. Touch NOTHING in primary loom-builder. This narrows blast radius.
sol-clone, crawlerstudiov2-prisma-d1, phase0-integration-phaseAB, all ephesians4-*, code-lens-scaffold, sprint-2 each have a forbidden-modify list
Prompts specify the exact shape of the final report: files modified + per-section status table + verification command outputs, or a CLAUDE_DIGEST.md with a table of migrated APIs (file:line, old to new), test counts, manifest path. Consumable by the orchestrator's validator.
REPORT BACK (CLAUDE_DIGEST.md) section appears in nearly every multi-step coding prompt
Later prompts cite earlier dispatches by name: see ephesians4-seo-pass for the head pattern, Sprint 0 OrchestratorDO already EXISTS extend it, Sprint 1 src/execution/ is complete REUSE its pieces. The corpus acts as a living knowledge base.
sprint-2 cites sprint-1, sol-clone cites primary loom-builder, corpus-explorer S2 cites S1's seams
Common failure modes. High-severity entries cause the 'file exists but doesn't work' class of dispatch failure.
Prompts that say 'produce a comprehensive report' or 'make a working site' without enumerating falsifiable checks. Models interpret these as license to stop at the first plausible output.
Replace 'comprehensive' with quantified metric: sections covered >=7, comparison matrix with >=10 rows, next-steps section with >=5 concrete actions. Gate on the count.
Prompts that do not instruct the model to query code-atlas or check graphify-out before reading source. The model then either reads too much or too little. Especially harmful in multi-repo dispatches.
Add a STEP 0 block: run <atlas query> against <db-path> --top 10, or check for graphify-out/GRAPH_REPORT.md before reading source. Cite the database path explicitly.
Prompts under 10 lines that delegate a non-trivial task with no context, no metric, no constraints. These are scaffold prompts for smoke-tests but have leaked into the corpus. They cannot be validated and cannot be improved without re-running.
Even smoke prompts need a 4-line contract: ROLE, GOAL, ONE GATE, REPORT. Otherwise they produce un-reviewable artifacts.
Prompts that delegate work to a known-fragile API (Playwright, Cloudflare Workers, Camoufox, Prisma adapter, DuckDB-WASM) without showing the RIGHT shape. The model falls back to training-data guesses that are out of date.
Every prompt that touches an API likely to have moved since training (Playwright/Cloudflare/DuckDB-WASM/prisma/bun internals) must include a RIGHT/WRONG block with citation to the installed package's docs/CHANGELOG.
Prompts that say 'read the full S0 handoff' or 'explore the entire src/ tree' or 'read both reports completely' without bounding the read. Models spend 30+ tool calls reading before writing anything.
Replace 'read everything' with a numbered list of named files + the page-range to read. 'Read pages 1-5 of CLAUDE_HANDOFF.md, sections S3 + S7'.
Prompts that name the goal but don't enumerate forbidden files. Models refactor adjacent code, breaking gates that the orchestrator wasn't watching. Especially common in 'scaffold from scratch' prompts.
Add a 'Do NOT touch' line: preserve <existing files>; extend, do not rewrite. For multi-repo: other repos are read-only references.
Prompts that end with 'report back' but no shell block to run. The model claims success and the orchestrator has to invent a check. This is the most common source of 'file exists but doesn't work' failures.
Every prompt must end with at least one bash block the model runs and pastes the output of. 'curl + grep' for HTTP, 'bun test' for code, 'wc -l' for content.
Prompts that list 8-12 'read first' files. The model reads them all, exhausts its read budget, and writes output under context pressure. The corpus-explorer S1 prompt pushes the upper limit at 6 numbered items + known gotchas.
Cap at 5 'read first' files. Move the rest to a 'Reference if blocked' appendix. Or summarize the additional files in 2-3 lines so the model doesn't need to read them.
Some prompts use 'Read FIRST', others 'Context', others 'INPUT FILES'. Some use 'GOAL', others 'User intent (restated)', others 'TASK'. The structural drift is real but minor — Mike reads them all, but small models pattern-match inconsistently.
Lock the template to a single section ordering: ROLE -> STATE ANCHOR -> INTEL DISCOVERY -> GOAL -> REQUIREMENTS -> STEPS -> GATES -> REPORT. Migrate older prompts in a separate pass.
Click a column header to sort. Default ordering is by overall score (descending).
| Prompt | Struct | Spec | Rails | Meas | Ctx | Overall | Notes |
|---|---|---|---|---|---|---|---|
20260720-corpus-explorer-s1-etl.md | 5 | 5 | 5 | 5 | 5 | 5.0 | Model-defining. Hard rules from CLAUDE.md, known gotchas block, missing-spec callout, ordered read list with status, quantified '12 assertions pass green' metric, bound to DDL + corpus.config.ts. |
20260720-corpus-explorer-s2-astro-site.md | 5 | 5 | 5 | 5 | 5 | 5.0 | Sister to S1, equally strong. Calls out the binding SPEC ADDENDUM v1.1, the seams S1 left, the exact Parquet manifest row counts. Hard rules on set:html for JSON-LD. |
20260720-corpus-explorer-s3e-thumbnail-hang-fix.md | 5 | 5 | 5 | 5 | 5 | 5.0 | Diagnoses the prior root cause explicitly, then dictates the fix with Bun.spawn RIGHT/WRONG, orphan-process verification, schema-correct status buckets. Reads like a senior engineer's incident postmortem. |
20260710-ephesians4-donations.md | 5 | 5 | 5 | 5 | 5 | 5.0 | House pattern (register.ts) cited as proven. Exact Stripe + Square API shapes verbatim. Signature verification hand-rolled for WebCrypto. E2E graceful-degradation proof with zero keys. |
20260710-ephesians4-seo-pass.md | 5 | 5 | 5 | 5 | 5 | 5.0 | Checklist-driven: each section of the universal SEO checklist maps to an implement step. Curl/grep gates verbatim. Site facts named explicitly (no invention). Don't-touch list. |
20260707-crawlerstudiov2-prisma-d1.md | 5 | 5 | 5 | 5 | 4 | 4.8 | Honest about known defect (two Prisma singletons) and gives the consolidation choice to the model. 'Do NOT run wrangler deploy' is repeated three times for emphasis. v5 vs v6 vs v7 Prisma decision is delegated correctly. |
20260707-phase0-integration-phaseAB.md | 5 | 5 | 5 | 5 | 4 | 4.8 | BINDING DECISIONS D1/D4/D5/D6 enumerated, source cited (ARD + ADR). 'Deleting/skipping a fixture = task failure' is exemplary. Work items are surgical edits with full paths. |
sol-clone-pi-080-migration.md | 5 | 5 | 5 | 5 | 4 | 4.8 | Quantified 4-point success metric with end-to-end smoke proof. Pre-080 archive directory mandated before edits. 'gates verify untouched' for the primary repo. Clone is not git — handles it. |
sprint-1.md | 5 | 5 | 5 | 4 | 5 | 4.8 | 9 modules built in pure-core / thin-shell architecture. Module-level contracts cited (ExecutionTrace, DispatchEnvelope). 'Tests MUST NOT hit network' is the kind of constraint that prevents hours of waste. |
sprint-2.md | 5 | 5 | 5 | 4 | 5 | 4.8 | Extends Sprint 1 — explicitly forbids duplicating execution/. PURE CORE / THIN SHELL rule. HMAC token format must be shared across Loop 1 + Loop 3 (cross-loop contract surfaced). |
20260720-corpus-explorer-s3c-ssr-search-content.md | 5 | 5 | 5 | 5 | 4 | 4.8 | Calls out the prior S3 failure by name (file-existence gates pass while search returns nothing). New metric is 'real Pagefind result for a real distinctive phrase, not a claim'. The forensic tone is exemplary. |
20260710-astro-local-build-runbook-review.md | 5 | 5 | 4 | 4 | 4 | 4.4 | Runbook distillation from a real session. Cites 'proven: all routes 500 to 200' for the nodejs_compat claim — strong ground-truth discipline. Knowledge captured is reusable across the whole Astro-on-Workers pipeline. |
agents-md-enrich-batch.md | 4 | 5 | 4 | 5 | 4 | 4.4 | Quantified 14/14 success metric. HEEC loop, FLEET QUERY FIRST, exact repo list. Missing: per-repo shell gate for line count. |
20260710-ephesians4-blog-content.md | 4 | 5 | 4 | 3 | 4 | 4.0 | Strong: voice/format anchors, exact 4 file paths, schema, dates, scripture. Weak: no shell-verifiable gate; '4 markdown files exist' is the only check. |
20260710-ephesians4-seo-validate.md | 4 | 4 | 4 | 4 | 4 | 4.0 | Companion validator dispatch — assumes the seo-pass output. Lightly-scoped but role-appropriate. |
architecture-first-workflow.md | 4 | 4 | 4 | 4 | 4 | 4.0 | Compact dispatch that encodes citations, budget, and YAML validation in 30 lines. Reuses the template's BUDGET formula verbatim. |
sprint-3.md | 4 | 4 | 4 | 3 | 4 | 3.8 | Cites Sprint 1 + Sprint 2 contracts. Weaker on 'done' — needs more explicit gates. |
20260707-crawlerstudiov2-agents-md.md | 4 | 4 | 3 | 4 | 3 | 3.6 | Good H2 section spec with required sections. Weak: 'explore the full src/ tree' invites broad reads, no explicit blast-radius for the existing README/Caddyfile/components.json. |
webgraph-phase1-foundation.md | 4 | 4 | 3 | 3 | 4 | 3.6 | Good META-EXECUTION PROTOCOL loop (Hypothesize-Execute-Evaluate-Commit). Weak: no code-atlas STEP 0, no shell-verifiable gate at the end. |
smoke.md | 1 | 3 | 1 | 2 | 2 | 1.8 | Canonical anti-example. 3 lines, no ROLE, no metric, no gate. Used as a sanity probe to confirm the harness responds. Belongs in _archive, not the main corpus. |
20260720-corpus-explorer-s1-etl.md20260720-corpus-explorer-s2-astro-site.md20260720-corpus-explorer-s3e-thumbnail-hang-fix.md20260710-ephesians4-donations.md20260710-ephesians4-seo-pass.mdsmoke.mdagents-md-organization-pass.mdwebgraph-phase1-foundation.mdOrdered by impact. The first three address the largest variance gap in the corpus.