Independent use will determine whether SpecJudge can use CLAUDE.md, AGENTS.md, and nested repository instructions to select cheaper coding models without materially reducing task quality.

state: expiredheat: lowuncertainty: highconvergesscott: mediumcoding-agents model-routing repository-contextSpecJudge

What is this?

SpecJudge is presented as a local tool that reads repository instruction files such as CLAUDE.md and AGENTS.md to recommend the least expensive coding model likely to handle that repository, while flagging insufficient context. The supplied results establish that these files encode repository-specific guidance, can be nested with closest-file precedence, and are used across multiple coding agents, but evidence on their value conflicts: some testing reports benefits, while other studies report higher token use and little or negative quality improvement. The snippets do not independently document SpecJudge’s creator, implementation, benchmark results, or whether instruction files contain enough signal for reliable model routing, so its core claim remains unverified here.

Why it matters to Scott

SpecJudge combines Scott’s existing CLAUDE.md-as-repository-spec pattern with his task-aware, cost-tiered model-routing work, while adding a concrete repository-level routing and insufficient-context decision. Independent results could affect how he builds routing into coding-agent harnesses—especially given his warning that bloated instruction files can increase cost and reduce success—but the tool’s claims and creator remain unverified, limiting relevance for now.

ip:concept.claude-md-patterndev:concept.task-aware-model-routingip:concept.fat-agents-md-anti-patternip:concept.evaluation-driven-developmentradar:concept.model-routingradar:multi-model-orchestrator-worker-agentsradar:karpathy-claude-md-rulesradar:concept.agent-evaluation

queries asked of Scott's wikis

repository-aware coding model routing
cheap-model escalation for coding agents
CLAUDE.md and AGENTS.md as machine-readable repository specifications
nested instruction precedence in agent harnesses
task quality measurement for coding-agent routing
insufficient-context detection and routing confidence

Measured heat

no measured readings yet — the hourly heat pass fills this in

How the heat travelled

no chain yet — the hourly chain pass fills this in

Evidence (3) — ⭐ canonical anchor

source	object	author	comments
🟠 reddit ⭐	Built a local tool that reads your CLAUDE.md and tells you which model that repo actually needs — and labels the answer when your context file isn't enough to tell ClaudeAI	jokiruiz	2
🟠 reddit	Stop blindly using Claude Sonnet. I built an open-source evaluator (SpecJudge) to prove local models can beat it if you use Spec-Driven Development LocalLLaMA	jokiruiz	9
🟠 reddit	I built a tool that tells you whether your project actually needs Opus — it's now a one-command install in spec-kit ClaudeAI	jokiruiz	5

Interpretation history

2026-08-20T09:37:53Z
After repeated reviews, SpecJudge has attracted no independent use, inspectable benchmark, or external implementation, and there is no concrete confirming event expected. The artifact has faded without advancing the repository-driven routing hypothesis, so the episode can close pending genuinely new evidence.
2026-08-18T09:34:02Z
The extra comments add no inspectable independent use, implementation, or controlled quality-cost result, so SpecJudge remains a low-traction maintainer claim rather than evidence for repository-driven model routing.
2026-08-16T09:30:11Z
No new evidence has appeared beyond the maintainer’s repeated claims, and the minor engagement changes add no independent use, benchmark, or release verification. The case remains an unvalidated implementation awaiting external results.
2026-08-14T08:39:34Z
The latest maintainer post broadens SpecJudge’s claimed implementation to one-command spec-kit installation and repository-context-based model ranking, but it remains the same creator’s unverified line of evidence. Without independent use or controlled quality-and-cost results, the case still does not show that repository instructions can route coding tasks safely.
2026-08-14T08:22:26Z
evidence attached: reddit.post.1vo0d71 — The tool directly implements repository-instruction-based model selection, materially bearing on the open routing hypothesis.
2026-08-12T10:30:08Z
The refreshed discussion is mostly dismissive or generic model-choice commentary and adds neither independent use nor a controlled quality-cost result. SpecJudge remains an unvalidated single-maintainer routing artifact rather than evidence that repository instructions can safely select cheaper models.
2026-08-10T10:24:41Z
The new post adds a maintainer-reported Haiku-versus-Sonnet anecdote, but changing both the model and specification makes it incapable of validating SpecJudge’s routing. The case still awaits independent use or a controlled quality-and-cost benchmark.
2026-08-10T10:22:08Z
evidence attached: reddit.post.1vkglyc — The post provides direct usage context for SpecJudge's repository-instruction-driven model selection hypothesis.
2026-08-08T14:35:02Z
The additional activity adds no independent implementation or quality-cost evidence; the case remains a single-maintainer claim awaiting external use or benchmarks.
2026-08-08T14:29:32Z
grounded: converges/medium — SpecJudge combines Scott’s existing CLAUDE.md-as-repository-spec pattern with his task-aware, cost-tiered model-routing work, while adding a concrete repository
2026-08-08T14:23:50Z
case created — The MIT-licensed CLI is a concrete new routing artifact, but its quality and cost benefits currently rest only on the developer's low-engagement announcement.

Decision trace

08-20 19:37expireAfter repeated reviews, SpecJudge has attracted no independent use, inspectable benchmark, or external implementation, and there is no concrete confirming event expected. The artifact has faded withou
08-20 19:37alert_silentThe only trigger is staleness and the reobserved posts remain substantively unchanged; there is no new event worth surfacing before a future independent result.
08-20 19:37alert_routeThe only trigger is staleness and the reobserved posts remain substantively unchanged; there is no new event worth surfacing before a future independent result.
08-18 19:34repriceThe extra comments add no inspectable independent use, implementation, or controlled quality-cost result, so SpecJudge remains a low-traction maintainer claim rather than evidence for repository-drive
08-18 19:34alert_silentOnly stale engagement changed; no consequential event or validation occurred, so this can wait for routine review or a concrete external result.
08-18 19:34alert_routeOnly stale engagement changed; no consequential event or validation occurred, so this can wait for routine review or a concrete external result.
08-16 19:30repriceNo new evidence has appeared beyond the maintainer’s repeated claims, and the minor engagement changes add no independent use, benchmark, or release verification. The case remains an unvalidated imple
08-16 19:30alert_silentThis review was triggered only by staleness; no consequential new event or confirming evidence occurred, so the case can wait for independent adoption or controlled quality-and-cost results.
08-16 19:30alert_routeThis review was triggered only by staleness; no consequential new event or confirming evidence occurred, so the case can wait for independent adoption or controlled quality-and-cost results.
08-14 18:39repriceThe latest maintainer post broadens SpecJudge’s claimed implementation to one-command spec-kit installation and repository-context-based model ranking, but it remains the same creator’s unverified lin
08-14 18:39alert_silentThe new delta is a creator-reported packaging and implementation update, not independent validation or a consequential access change backed by a release artifact; it can wait for routine review.
08-14 18:39alert_routeThe new delta is a creator-reported packaging and implementation update, not independent validation or a consequential access change backed by a release artifact; it can wait for routine review.
08-14 18:22alert_silentThis is a creator-only Reddit claim about easier installation and task-level escalation logic, with no linked release artifact, independent use, or quality/cost results. The implementation ideas are r
08-14 18:22surface_candidateThis is a creator-only Reddit claim about easier installation and task-level escalation logic, with no linked release artifact, independent use, or quality/cost results. The implementation ideas are r
08-14 18:22alert_routeThis is a creator-only Reddit claim about easier installation and task-level escalation logic, with no linked release artifact, independent use, or quality/cost results. The implementation ideas are r
08-14 18:22attachThe tool directly implements repository-instruction-based model selection, materially bearing on the open routing hypothesis.
08-14 18:21propose_attachThe tool directly implements repository-instruction-based model selection, materially bearing on the open routing hypothesis.
08-12 20:30repriceThe refreshed discussion is mostly dismissive or generic model-choice commentary and adds neither independent use nor a controlled quality-cost result. SpecJudge remains an unvalidated single-maintain
08-12 20:30alert_silentNo consequential new fact occurred; additional comments do not validate the routing mechanism or materially challenge the existing assessment, so this can wait for independent adoption or benchmarks.
08-12 20:30alert_routeNo consequential new fact occurred; additional comments do not validate the routing mechanism or materially challenge the existing assessment, so this can wait for independent adoption or benchmarks.
08-10 20:24repriceThe new post adds a maintainer-reported Haiku-versus-Sonnet anecdote, but changing both the model and specification makes it incapable of validating SpecJudge’s routing. The case still awaits independ
08-10 20:24alert_silentThis repeats the creator’s claim without independent adoption or a controlled comparison, so it can wait for a routine briefing or stronger evidence.
08-10 20:24alert_routeThis repeats the creator’s claim without independent adoption or a controlled comparison, so it can wait for a routine briefing or stronger evidence.
08-10 20:22alert_silentThis is another creator-authored promotional claim, not independent use or a reproducible benchmark showing that SpecJudge’s routing preserves quality while reducing cost. The asserted Haiku-versus-So
08-10 20:22surface_candidateThis is another creator-authored promotional claim, not independent use or a reproducible benchmark showing that SpecJudge’s routing preserves quality while reducing cost. The asserted Haiku-versus-So
08-10 20:22alert_routeThis is another creator-authored promotional claim, not independent use or a reproducible benchmark showing that SpecJudge’s routing preserves quality while reducing cost. The asserted Haiku-versus-So
08-10 20:22attachThe post provides direct usage context for SpecJudge's repository-instruction-driven model selection hypothesis.
08-10 20:21propose_attachThe post provides direct usage context for SpecJudge's repository-instruction-driven model selection hypothesis.
08-09 00:35repriceThe additional activity adds no independent implementation or quality-cost evidence; the case remains a single-maintainer claim awaiting external use or benchmarks.
08-09 00:35alert_silentOnly minor engagement changed, with no new artifact, benchmark, release detail, or independent result; nothing makes this more useful today than in a later briefing.
08-09 00:35alert_routeOnly minor engagement changed, with no new artifact, benchmark, release detail, or independent result; nothing makes this more useful today than in a later briefing.
08-09 00:34alert_silentA maintainer claims SpecJudge now parses common repository instruction files and handles insufficient-context routing, but the only evidence is an incomplete Reddit post with no linked release or inde
08-09 00:34surface_candidateA maintainer claims SpecJudge now parses common repository instruction files and handles insufficient-context routing, but the only evidence is an incomplete Reddit post with no linked release or inde
08-09 00:34alert_routeA maintainer claims SpecJudge now parses common repository instruction files and handles insufficient-context routing, but the only evidence is an incomplete Reddit post with no linked release or inde
08-09 00:29groundSpecJudge combines Scott’s existing CLAUDE.md-as-repository-spec pattern with his task-aware, cost-tiered model-routing work, while adding a concrete repository-level routing and insufficient-context
08-09 00:23createThe MIT-licensed CLI is a concrete new routing artifact, but its quality and cost benefits currently rest only on the developer's low-engagement announcement.