Anchor – Disciplinas Comportamentais para Qualidade de Código
johnpatrickwarren-oss/anchor
Disciplinas comportamentais que os modelos base não fazem por padrão
Ver no GitHubFramework de seis disciplinas comportamentais que complementa ferramentas determinísticas (como sprag) para elevar a qualidade de código além dos padrões base de modelos de IA. Inclui revisão independente, contratos spec-first, halt-on-contradiction, V/Q debugging, medição honesta e trilha durável de projeto — cada uma testada para agregar valor que não pode ser automatizado.
Instalação
npx convez-ai johnpatrickwarren-oss/anchor1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Elevar qualidade de projetos multi-agente sem orquestração complexa
- Aplicar spec-first contracts e revisão independente em fluxos de desenvolvimento com Claude
- Estabelecer trilha durável de decisões e estado do projeto (STATE.md + ADRs)
Veredito

Anchor é uma metodologia enxuta e bem fundamentada, focada em disciplinas que modelos base realmente não fazem por padrão. Destaca-se pela honestidade (aposentadoria de práticas superadas com upgrades de modelo) e pela sinergia com sprag, formando um stack mínimo e eficaz. Recomendado para equipes que querem qualidade sem overhead orquestral.
README
A tradução para o português deste README ainda está na fila.
anchor
The behavioral half of code quality. sprag is the deterministic half.
Anchor is a small, deliberately short set of behavioral disciplines that still beat a strong base model's defaults — the judgment calls you can't compile into a gate. It pairs with sprag, which enforces the deterministic floor (structure, test-efficacy, scope) mechanically, with no model in the loop.
sprag is the gate. Anchor is the prompt. Together they are the whole quality stack — a mechanical floor plus a handful of disciplines — with no orchestration harness.
The disciplines
The substance is one file: DISCIPLINES.md. Six disciplines, two
load-bearing:
- Independent cold-eye review — a fresh-context reviewer, prompted to find what's wrong, that never saw the build reasoning. (load-bearing)
- Spec-first contract — acceptance criteria + anti-scope before code; every conjunct and every prescription gets a check. (load-bearing)
- Halt-on-contradiction — verify the spec's premises; when spec ≠ reality, stop and ask a bounded question instead of coding around it.
- V/Q debugging — enumerate hypothesis variants before going deep; frame the falsifying question; re-enumerate after ~3 dead ends.
- Honest measurement — instrument the gap, don't footnote it; a headline that needs a caveat is misreporting.
- Durable project trail — leave a cold-readable record: overwrite a short
STATE.md(the "now"), append one ADR per decision (the "forever"). So a project survives a months-long gap or a handoff.
Each entry gives the trigger, the discipline, why it beats the base model's default, and how to apply it in a dynamic session.
Why it's short
A discipline earns a place only if it passes both sieves: (1) it can't be made deterministic —
else it's a sprag invariant, not prose; and (2) it still beats the base model's default — else
it's ceremony. Most of the original Anchor fails one sieve; the full mapping is in What used to be
here in DISCIPLINES.md. The list also has a half-life: re-test each entry on every model
upgrade and retire what the model now does unprompted — that ritual is what keeps the pack from
becoming the overhead it replaced.
How to use it
Pair it with sprag. Reference DISCIPLINES.md from your project's CLAUDE.md
(e.g. @DISCIPLINES.md) so the agent applies the disciplines in-context, and let sprag's gate
enforce the deterministic floor. No roles, no pipeline, no runtime.
Discipline 2 covers eliciting the project spec from an operator who may not code (lead with
must-never). For behavioral properties of a single function — invariants that feed sprag's
arch property — use the question-driven PRD-INTERVIEW.md: plain-language
answers, drafted impl-blind, accepted by the deterministic gate.
History
Anchor began as a five/six-role build-orchestration methodology distilled from running
DeploySignal as a multi-agent project
(250+ coordination files; ~94% autonomous execution — self-measured, one project). With a
1M-context model the orchestration largely inverted into overhead. The design trail is in
design/adr/: ADR-0001 (orchestration → overhead) and ADR-0002 (the
deterministic / behavioral split — sprag and this repo — that resulted). The original methodology
and the experimental @anchor/* tool are preserved, unmaintained, in legacy/.
License
Apache-2.0 — see LICENSE. © 2026 John Warren ·
john.patrick.warren@gmail.com