Codex Reasoning Rigor

TL;DR

Lock task contract (goal/constraints/non-goals) → explore 2+ options → ground in repo artifacts → attach monitoring signals → validate output specificity. Never settle for generic "best practices" without concrete evidence.

Codex Reasoning Rigor

Overview

Use this skill to turn vague or generic model behavior into explicit decision-making, repo-grounded evidence gathering, and output contracts that are concrete enough to verify. Activate it when the task is important enough that "sounds right" is not good enough.

Activation

Activate on explicit $codex-reasoning-rigor or $rigor.

Activate when the user says or implies:

• "don't be generic"
• "think harder" or "go deeper"
• "make it specific"
• "I need better workflow / monitoring / output quality"
• "give me tradeoffs, not vague advice"
• "show evidence"
• "use the repo, not generic training data"

Also activate for:

• architecture or refactor decisions with hidden risk
• reviews that need actionable findings instead of summaries
• operational plans that need monitoring signals and stop/ship criteria
• repeated tasks where model output has drifted or become bland

Core Protocol

Run this protocol before answering or implementing:

1. Task Contract

Lock these fields explicitly:

• goal: what outcome must change
• constraints: what cannot change
• non_goals: what should stay out of scope
• evidence_required: what proof would make the answer credible
• quality_bar: what would make the output clearly non-generic

If the task is complex, generate a brief with scripts/build_reasoning_brief.py. By default the brief generator now requires those reasoning fields; use --allow-placeholders only when you intentionally want a scaffold.

2. Decision Surface

Do not jump straight to one answer. Frame at least 2 options whenever the task has hidden tradeoffs.

For each option, compare:

• why it could work here
• cost or blast radius
• failure modes
• what evidence would confirm or reject it

3. Repo Grounding

Anchor the answer in concrete artifacts whenever possible:

• real file paths
• real commands
• actual counts, versions, or dates
• existing scripts, references, or templates
• diffable outputs instead of abstract advice

Use references/evidence-ladder.md to keep evidence quality high.

4. Monitoring Lens

Every serious workflow or recommendation should state:

• what to watch
• what "healthy" looks like
• what drift or failure looks like
• what to do next if the signal goes red

Use references/monitoring-loops.md when the task affects delivery, release quality, or repeated execution.

5. Output Contract

Before finishing, make the output earn its credibility. Use references/output-contracts.md and check that the result includes:

• the chosen path
• evidence or commands that support it
• risks or open questions
• next steps or exit criteria

Anti-Generic Rules

Never end at "best practices", "ensure", "optimize", "improve quality", or similar filler unless you immediately attach:

• a named artifact
• a concrete command
• a specific file or boundary
• a measurable condition
• a reason this matters in this repo or workflow

Use references/anti-generic-patterns.md when the answer starts sounding interchangeable with any other project.

Document Clarity and Token Economy

When the user wants clearer writing, better Vietnamese output, or lower token overhead:

• remove repeated framing and prompt restatement
• prefer explicit labels over connective filler
• keep each sentence semantically complete, even when the sentence is long
• use concrete verbs, exact artifacts, and stable section ordering
• treat mojibake or mixed-encoding output as a quality failure
• stack codex-document-writer when the deliverable is a report, memo, guide, handoff, proposal, or formal document

Use references/document-clarity.md for writing rules, Vietnamese labeling, and rendering-safe output patterns.

Working Pattern

For medium or high-stakes work:

1. Build a reasoning brief.
2. Name the decision surface.
3. Ground the answer in repo artifacts.
4. State the monitoring loop.
5. Run an output-quality pass with $output-guard before declaring the work complete.
6. For plans, reviews, and handoffs, add an editorial pass with $editorial-review so the result reads like a human-made engineering deliverable instead of safe model prose.

For plans, reviews, and handoffs, treat strict output validation as the default quality bar, not an optional extra.

Resources

• scripts/build_reasoning_brief.py: generate a structured execution brief for complex, high-signal work.
• references/anti-generic-patterns.md: common signs of generic output and concrete repair moves.
• references/document-clarity.md: token economy, Vietnamese document structure, and mojibake-safe writing rules.
• references/evidence-ladder.md: how to rank evidence from weak intuition to strong verification.
• references/monitoring-loops.md: how to attach health signals and follow-up actions to workflows.
• references/output-contracts.md: response shapes that push outputs toward specificity and accountability.
• assets/reasoning-brief-template.md: markdown template used by the brief generator.
• assets/output-review-template.md: template for reviewing whether a deliverable is specific, accountable, and editorially strong enough.
• assets/monitoring-checklist-template.md: template for tracking healthy vs drifting workflow signals.