reflect

Files (1)

• SKILL.md

SKILL.md (first lines)

# Reflect — Human AI-Usage Self-Assessment

This skill is for **you** (the human) to reflect on how well you're leveraging AI as a thought amplifier — not a review of agent performance.

**Data source**: The `reflect` hook (session-end) automatically collects observations, analyzes patterns, and updates `metrics.json`. This skill consumes that hook-produced data to produce a human-readable self-assessment.

```
Hook (auto)                    Skill (on-demand /reflect)
─────────────                  ──────────────────────────
observe → obs/*.jsonl  ──→     epic reflect --context 30
evolve → metrics.json  ──→     5-dimension scorecard
seed → evolved skills  ──→     Action items for the human
ingest → memory graph  ──→     Trend analysis
```

## Iron Law

No score without evidence. Every rating must directly cite at least one of: obs stats, evolution patterns, memory nodes, or session summaries.
Block self-serving bias: "doing well" conclusions require concrete metrics.

## Process

### Step 0 — Collect Context

```bash
# Uses Rust subcommand — works on all platforms (Linux, macOS, Windows)
epic reflect --context 30 > /tmp/reflect_ctx.json
```

Fallback if subcommand fails:
```bash
echo "obs_files: $(ls "$HARNESS_DIR/obs/" | wc -l)"
python3 -c "import json; m=json.load(open('$HARNESS_DIR/metrics.json')); print('total_sessions:', m.get('total_sessions',0))"
```

Query memory (if active):
```bash
epic mem recall "AI usage patterns decisions metacognition" --limit 8
epic mem list --type decision --limit 5
epic mem list --type pattern --limit 5
```

### Step 1 — 5-Dimension Reflection

Score each dimension independently: 1–10 + evidence citation + one-line diagnosis.

#### Dimension 1: Thought Amplification
**Question**: Is AI a mere executor (code typist) or a genuine thought partner?

Metrics:
- Agent tool call ratio (`Agent / total_obs` — higher = delegated thinking)
- Skill invocation frequency (meta-layer usage)
- council/discover/spec execution history
- Memory decision node count

Scoring:
| Score | Signal |
|-------|--------|
| 8–10  | Agent delegation ≥ 5%, diverse skill usage, council execution recorded |
| 5–7   | Mostly Bash/Read/Edit, occasional Agent, complex decisions made solo |
| 1–4   | Bash+Edit 90%+, AI at autocomplete level |

#### Dimension 2: Self-Improvement
**Question**: Learning from mistakes, or repeating the same patterns?

Metrics:
- `evolution_stats.pattern_frequency` — recurring patterns?
- `evolution_stats.stagnation_count` — stagnant sessions
- `evolution_stats.trend_last10` — improving/stable/declining distribution
- Evolved skills count vs timespan

Scoring:
| Score | Signal |
|-------|--------|
| 8–10  | Trend improving ≥ 60%, no pattern recurrence, evolved skills growing |
| 5–7   | Mostly stable trend, some pattern repeats, evolved skills plateau |
| 1–4   | Trend declining or frequent stagnation, same mistakes repeated |

#### Dimension 3: Metacognitive Expansion
**Question**: Are conversations with AI helping recognize and upgrade your own thinking?

Metrics:
- Memory concept/pattern node count and recency
- Decision/ADR recording frequency
- Session snapshots mentioning "learnings"
- `/discover` `/spec` execution history (problem-framing practice)

Scoring:
| Score | Signal |
|-------|--------|
| 8–10  | Regular decision nodes, /discover /spec used, ADRs exist |
| 5–7   | Intermittent recording, decision rationale only in code, no explicit notes |
| 1–4   | Almost no memory nodes, context breaks between sessions |

#### Dimension 4: Prompt Engineering
**Question**: Is prompt quality improving over time?

Metrics:
- `output_quality` dimension average trend (metrics score_history)
- `tool_success` rate trend
- Evolved skills with prompt improvement patterns
- Session average score trajectory (early vs recent)

Scoring:
| Score | Signal |
|-------|--------|
| 8–10  | output_quality ≥ 0.80, score trend upward, evolved prompts increasing |
| 5–7   | output_quality 0.65–0.80, improvement plateau |
| 1–4   | output_quality < 0.65, downward trend, frequent re-edits |

#### Dimension 5: Execution Efficiency
**Question**: Achieving the same goals faster and cheaper through AI?

Metrics:
- `execution_cost` dimension average (1.0 = optimal)
- Bash dominance (Bash > 50% suggests excessive low-level repetition)
- Context compaction frequency (too frequent = context waste)
- Agent sub-agent parallel usage

From epicsagas/epic-harness, Apache-2.0.

More from epicsagas/epic-harness