retex

name

retex

description

AI Error Post-Mortem. Triggers when the user says "RETEX" or after the AI has failed 3+ times on the same problem. Captures WHY the AI failed and HOW to prevent recurrence. Produces a structured markdown file with 6 mandatory blocks including named anti-patterns and injectable "vaccine" rules. This is not a handover (project state) or an audit (quality check) — it is a diagnostic that produces immunity.

Skill: RETEX — AI Error Post-Mortem

Version: 1.0
Part of: never-start-over
Companion skills: handover (project state), todo-tracker (task extraction)
Philosophy: A RETEX is a vaccine, not a diagnosis. It captures WHY the AI failed and HOW to never fail the same way again.

---

1. Structure (6 mandatory blocks)

Block 1 — CONTEXT (3 lines max)

What, where, why. No history, just the frame.

Block 2 — ERROR CHRONOLOGY

Table: each failed attempt, what was done, the cognitive bias identified.

#	Attempt	What Was Done	Why It Failed	Bias
1	...	...	...	Named bias

Always include: Total duration of the error cycle and Duration of the final solution.

Block 3 — ANTI-PATTERNS (named)

Each toxic reflex gets a memorable name. This is critical — unnamed patterns get repeated; named patterns get caught.

Format per anti-pattern:

• Name: short, memorable label
• Definition: 1 line
• Example: concrete, from this session
• Warning signal: how to detect BEFORE falling into it

Block 4 — TIPPING POINT

The exact moment it unblocked. Always formulate as:

• Moment: exact quote or action
• What the AI thought: X
• The real problem: Y
• Lesson: 1 sentence

Block 5 — BEST PRACTICES

Extracted rules, formulated as imperative directives (copy-pasteable into prompts).
Format: action verb + constraint + reason.

Block 6 — VACCINE

3 to 7 condensed sentences. Directly injectable into system prompts or AI memory.
Each sentence = 1 autonomous rule, understandable without context.
Format: code block with identifier VACCINE_RETEX_{{NNN}}.

Numbering: auto-increment from the last known vaccine number. If no previous RETEX exists, start at 001.

Where to inject vaccines after generating them:

• CLAUDE.md / AGENTS.md — paste into project-level instructions (always-on)
• System prompt / AI memory — paste for persistent cross-session rules
• Session start — paste manually when starting a related project
• Vaccines are designed to work without context — no need to include the full RETEX

---

2. Named Anti-Pattern Bank (cumulative)

These anti-patterns have been identified across multiple sessions. The AI MUST proactively detect them and ALERT the user when it recognizes itself falling into one.

The Fly Against the Window

Repeating the same approach with varying parameters instead of questioning the approach itself.
Signal: 3rd attempt in the same solution category without results → STOP, change paradigm.

The Scribble

Hand-drawing (artisanal SVG, code from scratch) what a specialized tool does infinitely better.
Signal: If you're writing coordinates/values "by feel" → STOP, find the existing source.

The Digital Duct Tape

Layering elements (pseudo-elements, overlays, patches) on top of existing work instead of replacing it.
Signal: If you're adding elements ON TOP to modify → the approach is dead.

The Local Patch

Fixing one parameter without checking coherence with the global system.
Signal: If you modify ONE parameter → re-read ALL context and verify system coherence.

The Ignored Good Result

Getting a result that works (or almost works), then abandoning it to chase a "cleaner" solution that ends up worse.
Signal: If a deliverable is "almost good" → ITERATE on it instead of starting over.

The Technician vs The Designer

Translating a visual request into technical parameters instead of reproducing the intended visual outcome.
Signal: If you're thinking in CSS parameters before thinking in visual output → step back and describe the EXPECTED RESULT in human words first.

The Nuclear Option

Rewriting an entire file instead of making a surgical fix. The AI regenerates 200 lines to change 3, introducing new bugs in previously working code.
Signal: If the fix touches more than 20% of the file → STOP, isolate the exact lines that need changing and patch only those.

The Headfirst Dive

Jumping straight into implementation without comparing 2-3 possible approaches first. The first idea feels obvious but is rarely optimal.
Signal: Before writing any code or producing any deliverable → list at least 2 approaches, compare trade-offs, then commit.

The Amnesia Loop

Proposing a solution that was already tried and failed earlier in the same session. The AI loses track of what's been attempted and cycles back to rejected approaches.
Signal: Before proposing a fix → scan the conversation for previous attempts. If it looks similar to something already tried → flag it explicitly.

The Interrogation

Asking 5 clarifying questions instead of making a decision. The user asked for help, not for a questionnaire. When the answer is inferrable, infer it.
Signal: If you're about to ask more than 1 question → pick the most likely interpretation, execute, and note your assumption.

---

3. Production Rules

• Re-read the ENTIRE conversation before writing the RETEX
• Every anti-pattern MUST have a memorable name, definition, example, and warning signal
• The vaccine MUST be self-contained (understandable without reading the rest of the RETEX)
• Newly identified anti-patterns get added to the cumulative bank (Section 2)
• The RETEX is delivered as a file, never as inline chat prose
• If the AI detects a known anti-pattern MID-SESSION → alert immediately: "Warning: [pattern name] detected. Change approach?"
• Mid-session detection requires loading this skill at session start (not just at RETEX time). For always-on detection, reference this skill in your CLAUDE.md or project instructions.

---

4. Naming Convention

RETEX_{{Subject}}_{{dd-mm-yy}}.md

---

5. Metadata (mandatory end block)

uri: {{conversation link or ID}}
date: {{dd-mm-yy}}
project: {{name}}
type: RETEX
subject: {{short description}}
iterations: {{number of failed attempts}}
error_duration: {{estimate}}
solution_duration: {{estimate}}
resolution: {{what finally worked}}
anti_patterns: {{count}} ({{names}})
vaccine: VACCINE_RETEX_{{NNN}}

---

6. Verification Checklist

Before delivering the RETEX:

• Re-read the ENTIRE conversation?
• Context is 3 lines or fewer?
• Every error has a named bias?
• Anti-patterns all have warning signals?
• Tipping point is formulated as "thought X / real problem Y"?
• Vaccine is understandable WITHOUT reading the rest?
• Metadata is complete?

---

Why This Skill Exists

AI assistants repeat the same mistakes across sessions because they have no institutional memory for errors. Every time context resets, hard-won lessons vanish. The RETEX format creates named, searchable, injectable immunity — anti-patterns with memorable names get caught before they cause damage, and vaccine rules can be loaded into any future session's system prompt.

The naming convention for anti-patterns is deliberate: "The Fly Against the Window" sticks in memory (human and AI) far better than "avoid repeating failed approaches." Named patterns become part of the shared vocabulary between user and AI.