# Retrospective

Version: 1.0.0

## Purpose

Evaluate completed work for reusable learning and decide whether any part of the AI operating system should improve.

Keep this lightweight.

The default outcome may be:

```text
NO PERMANENT CHANGE
```

---

# Review the Task

Ask:

```text
Did the requested result succeed?

What worked particularly well?

What failed or nearly failed?

What assumption was wrong?

What evidence resolved uncertainty?

What check found the problem?

Could something have caught it earlier?
```

---

# Identify Reusable Learning

Ask:

```text
Would this matter again?
```

Possible reusable discoveries:

```text
procedure

diagnostic method

verification method

failure pattern

project convention

prompt improvement

automation opportunity

template
```

---

# Determine Scope

Classify the lesson:

```text
TASK-SPECIFIC

PROJECT FACT

PROJECT PROCEDURE

GENERAL AGENT RULE

MODEL-SPECIFIC RULE

SKILL IMPROVEMENT

PROMPT IMPROVEMENT

AUTOMATION
```

---

# Choose Destination

Use the narrowest appropriate location.

```text
PROJECT FACT
→ project documentation

REUSABLE PROCEDURE
→ existing/new skill

GENERAL AGENT RULE
→ AGENTS.md

CLAUDE-SPECIFIC
→ CLAUDE.md

PROMPT FAILURE
→ prompt / prompt-debugger

AUTOMATABLE DETERMINISTIC WORK
→ script or automation

WEAK CANDIDATE
→ LEARNINGS.md
```

---

# Evidence Test

Before permanent promotion ask:

```text
Is there actual evidence?

Is the lesson reusable?

Is the cause understood?

Can the improvement be tested?

Does existing guidance already cover it?
```

If evidence is weak:

```text
record candidate in .ai/state/LEARNINGS.md
```

instead of changing permanent instructions.

---

# Skill Decision

If the task revealed a reusable procedure, use:

```text
.ai/skills/skill-extractor/SKILL.md
```

Do not create the skill directly without checking for existing ownership.

---

# Prompt Decision

If AI instructions failed, use:

```text
.ai/skills/prompt-debugger/SKILL.md
```

Determine why before rewriting.

---

# Automation Decision

If a repeated procedure is deterministic, consider whether:

```text
SCRIPT
```

is better than:

```text
SKILL
```

Examples:

```text
schema validation
file comparison
deployment health check
build verification
```

Agents should not solve with prompts what software can reliably automate.

---

# Output

Return an internal or visible summary:

```text
## Result

[...]

## What Worked

[...]

## What Failed

[...]

## Reusable Lesson

[...]

## Action

NO CHANGE
or
LEARNING ENTRY
or
UPDATE EXISTING SKILL
or
CREATE SKILL
or
UPDATE PROMPT
or
UPDATE PROJECT DOCS
or
UPDATE AGENTS.md
or
UPDATE CLAUDE.md
or
CREATE AUTOMATION
```

If no durable lesson exists, stop.

Do not create process merely because a retrospective was performed.