The Forge Master taps the blade and listens. A dull ring — too brittle. Back into the fire. Every master knows: you cannot improve what you cannot measure, and you cannot measure what you cannot observe. Keep a log. Test every change. Heat, strike, measure.

🗺️ Quest Network Position

graph LR
    Q12[✅ Q12: Necromancer's Inquest] --> Q13[🎯 Q13: Reforging the Mind]
    Q13 --> Q14[🔜 Q14: Council of Many]
    style Q13 fill:#4CAF50,stroke:#2E7D32,stroke-width:4px,color:#fff

🎯 Quest Objectives

• Establish a behaviour baseline — run 3 standard agent tasks and record outcomes
• Identify improvement targets — from RCA reports, select the top 2 failure patterns
• Implement instruction changes — modify copilot-instructions.md and AGENTS.md
• Measure improvement — re-run the same tasks and compare outcomes
• Maintain an instruction changelog — track every change with date, reason, and outcome

⚔️ The Quest Begins

Chapter 1 — Establishing a Behaviour Baseline

Before tuning, you need a benchmark. Run three representative tasks and record outcomes:

Exercise 13.1: Create the baseline measurement script.

Prerequisites (the script reads real workflow history, so it needs a repo with agent activity):

• gh auth login completed, or GH_TOKEN exported, so the GitHub CLI is authenticated
• an agent-task.yml workflow already committed to the repo
• at least one prior agent run/PR for tasks 1–3 to measure

# work/gh-600/scripts/measure_agent_baseline.sh
#!/usr/bin/env bash
set -euo pipefail

RESULTS_FILE="work/gh-600/baseline-results.jsonl"
RUN_DATE=$(date -u +%Y-%m-%dT%H:%M:%SZ)

echo "=== Agent Behaviour Baseline Measurement ==="

RECORDED=0

# For each test task, record:
# - Did the agent open a PR? (success signal 1)
# - Did all tests pass? (success signal 2)
# - Did it reference the original issue? (success signal 3)
# - Were any unexpected files modified? (failure signal)
# - Did it complete within the time limit? (efficiency signal)

for TASK_NUM in 1 2 3; do
    echo "Measuring task $TASK_NUM..."
    
    # Get the latest agent run for this task (degrade gracefully if none exists yet)
    RUN_ID=$(gh run list --workflow=agent-task.yml --limit=1 --json databaseId -q '.[0].databaseId' 2>/dev/null || true)
    if [ -z "$RUN_ID" ]; then
        echo "⚠️  No agent-task.yml run found for task $TASK_NUM — run the agent workflow at least once, then re-run this script."
        continue
    fi
    
    PR_OPENED=$(gh pr list --state all --search "is:pr in:title issue-$TASK_NUM" --json number -q 'length')
    TESTS_PASSED=$(gh run view "$RUN_ID" --json conclusion -q '.conclusion')
    
    cat >> "$RESULTS_FILE" << EOF
{"date":"$RUN_DATE","task":$TASK_NUM,"run_id":"$RUN_ID","pr_opened":$([ "$PR_OPENED" -gt 0 ] && echo true || echo false),"tests_passed":$([ "$TESTS_PASSED" = "success" ] && echo true || echo false)}
EOF
    RECORDED=$((RECORDED + 1))
done

if [ "$RECORDED" -eq 0 ]; then
    echo "❌ No tasks were recorded — no agent-task.yml runs exist yet, so $RESULTS_FILE was not created."
    echo "   Trigger at least one agent run for tasks 1–3, then re-run this script."
    exit 1
fi

echo "✅ Baseline recorded ($RECORDED task(s)) in $RESULTS_FILE"

Chapter 2 — Instruction Change Patterns

Based on common agent failure patterns, here are the most impactful instruction changes:

Chapter 3 — The Instruction Iteration Cycle

Exercise 13.2: Run one full iteration of the behaviour improvement cycle.

# Iteration 1 — Branch Naming

## Observation (from RCA)
Agent created a branch called `fix-the-thing` — untraceable to the original issue.

## Hypothesis
Adding explicit branch naming instructions to AGENTS.md will fix this.

## Change Made (2026-05-17)
Added to AGENTS.md:
> Branch name MUST follow this pattern: `copilot/issue-{N}-{3-5-word-slug}`
> Example: `copilot/issue-42-add-input-validation`

## Measurement
Before: 0/3 runs used traceable branch names
After:  3/3 runs used correct branch name format

## Outcome
✅ CONFIRMED IMPROVEMENT — change retained permanently

Chapter 4 — Maintaining the Instruction Changelog

Exercise 13.3: Set up the instruction changelog.

<!-- docs/agent-instructions/CHANGELOG.md -->
# Instruction Changelog

All changes to copilot-instructions.md and AGENTS.md are recorded here.
Format: Date | File | Change | Reason | Outcome

---

## 2026-05-17

### copilot-instructions.md
- **Added:** Mandatory planning step before any file modification
  - Reason: Agent was skipping planning in 2/3 test runs
  - Outcome: TBD — testing in progress

### AGENTS.md
- **Added:** Explicit branch naming format: `copilot/issue-{N}-{slug}`
  - Reason: Untraceable branch names found in 3 RCA reports
  - Outcome: ✅ 3/3 runs now use correct format

---

## Template for new entries:
### FILE
- **Change type (Added/Changed/Removed):** Description
  - Reason: Why was this change needed?
  - Outcome: ✅/❌/TBD — result after testing

✅ Quest Validation

Run this manual self-check from your repo root — it verifies your Q13 deliverables directly, so you don’t need any external validator script:

# Manual self-check — confirm your Q13 deliverables exist
test -f work/gh-600/baseline-results.jsonl \
  && echo "✅ Baseline measurement: baseline-results.jsonl present"
ls docs/agent-instructions/*.md >/dev/null 2>&1 \
  && echo "✅ Iteration log: iteration records in docs/agent-instructions/"
test -f docs/agent-instructions/CHANGELOG.md \
  && echo "✅ Instruction changelog: CHANGELOG.md present"
# 🏆 Quest Q13 complete when all three checks print ✅

🏆 Quest Rewards

🕸️ Knowledge Graph

Structured wiki-links connect this quest to the IT-Journey knowledge graph. Open the Obsidian Graph View to explore connections.

Level hub: [[Level 1011 - Feature Development]] Overworld: [[🏰 Overworld - Master Quest Map]] Study track: [[The Agentic Codex: GH-600 Study Hub]] · [[GH-600 Agentic AI Quick-Reference Notes]] Prerequisites: [[The Necromancer’s Inquest: Agent Failure Root Cause Analysis]] Unlocks: [[The Council of Many: Multi-Agent Orchestration Patterns]] Sequel quests: [[The Council of Many: Multi-Agent Orchestration Patterns]] Obsidian docs: [[Obsidian Knowledge Graph and Wiki Links]]