Codex Retrospective

A structured self-improvement loop for Codex.

This skill turns Codex from a one-off collaborator into a system that gets meaningfully better at working with you over time by systematically eating its own usage history.

Why This Exists

Most people improve their agent usage by manually maintaining AGENTS.md or skills when they notice problems. This skill makes that process deliberate, regular, and high-leverage.

It is directly inspired by strong practices from heavy users (especially Greg Brockman's emphasis on updating the "constitution" from real failures and friction), but executed as a repeatable arsenal-style workflow.

Core Principle

Minimal effective change, grounded in evidence from actual history.

Every output must be:

• The smallest possible useful addition or edit
• Backed by specific patterns observed in recent sessions
• Designed to prevent repetition of the same friction

When to Trigger

• After a painful or repetitive session ("Codex should have known better by now")
• Periodically (weekly or monthly): "Do a retrospective on the last 7/30 days of my Codex usage"
• When codex-fluent reports that Codex is constantly re-asking for the same context or preferences
• After you have manually corrected the same class of mistake multiple times

The Retrospective Process

1. Scope Definition

You specify the time window or focus area:

• "Last 14 days"
• "The three big auth + payments threads from last month"
• "All sessions involving the new Go service"

2. Evidence Collection Preflight

Before proposing any AGENTS.md update or tiny skill, Codex must build a small evidence inventory:

• Scope: repo/project, time window, and focus area.
• Sources checked: current AGENTS.md, relevant existing skills, repo docs, Memories/Chronicle, and local Codex session files when available.
• Evidence handles: session IDs, memory IDs, file paths, PR/issue numbers, or dated incidents that can be checked later.
• Runtime availability: distinguish "skill exists in source" from "skill is installed and triggerable in the active runtime" when that matters.
• No-change path: if evidence is weak, stop with a concise "no high-confidence change" result instead of manufacturing proposals.

Every proposed update must map back to at least one concrete evidence handle.

3. History Analysis (The Skill Guides Codex)

Codex is instructed to look for:

• Recurring mistakes or inefficient patterns
• Things the user had to explain or correct repeatedly
• High-friction moments (lots of back-and-forth, context loss, wrong assumptions)
• Successful patterns worth encoding so they happen by default
• Opportunities to extract tiny, reusable skills

4. Output — Strict Format

The skill forces Codex to produce output in this order:

0. Evidence Inventory (scope, sources checked, and concrete handles)
1. Retrospective Summary (short, evidence-based)
2. Proposed AGENTS.md Updates (exact diff or append text only)
3. New or Refined Minimal Skills (at most 1-2 tiny ones, full SKILL.md with frontmatter)
4. Rationale + Evidence (which sessions/patterns drove each proposal)
5. Application Plan (how to safely apply the changes)

5. Human Gate + Application

You review. The skill then helps you apply the minimal changes cleanly (never blindly overwriting large sections of AGENTS.md).

Relationship with codex-fluent

These two skills are designed to be used together:

• codex-retrospective finds behavioral and knowledge improvements (better defaults, new rules, extracted skills).
• codex-fluent finds state and context hygiene improvements (session bloat, missing handoffs, archive opportunities).

A good monthly ritual for serious users:

1. Run codex-retrospective on the last 30 days.
2. Run codex-fluent diagnosis.
3. Apply the best changes from both.

Hard Constraints on Output

• Never propose large rewrites of AGENTS.md.
• Never create big new skills. Tiny, focused, high-ROI only.
• Every proposal must reference concrete history ("In the payment retry thread on May 3rd and the similar incident on May 18th...").
• Do not propose changes until the evidence inventory exists and each proposal maps to a concrete handle.
• If nothing high-confidence was found, say so clearly instead of manufacturing improvements.

Gotchas

• Do not turn a single annoying session into a permanent rule unless it was unusually high impact. Prefer "no change" over a weak constitution update.
• Do not rely on memory vibes. Each proposed update needs a concrete session, date range, file, PR, or repeated correction pattern that can be checked.
• Keep generated AGENTS.md text short. If the proposed rule needs multiple paragraphs, it probably belongs in a tiny skill or a project doc instead.
• Application is gated by the user. The retrospective can propose exact diffs, but it must not silently rewrite AGENTS.md or existing skills.

References

• references/retrospective-prompt.md — The core prompt template used to drive Codex's self-analysis
• references/agents-md-update-rules.md — Strict rules for what kind of changes are acceptable
• references/minimal-skill-criteria.md — What qualifies as a "tiny useful skill" worth extracting
• references/examples/ — Real (sanitized) retrospective outputs and the resulting AGENTS.md diffs

Success Looks Like

After 4–8 weeks of regular use:

• Codex makes fewer basic assumption errors in your domain
• You spend less time re-explaining preferences and constraints
• Your AGENTS.md and skills folder feel like they were written by someone who has worked with you for a long time (because they were)
• New projects ramp up faster because the constitution already encodes hard-won lessons

Start with a focused 7- or 14-day retrospective on a project where you've felt the most friction recently. The pattern will become natural quickly.