Capture Manual Observation

Record a single observation to the case study harness log. Follow these steps in order.

Step 1: Get the observation

The user's observation is: $ARGUMENTS

If the arguments are empty, ask the user to describe what they observed and wait for their response before continuing.

Step 2: Classify the category

Based on the user's description and the current conversation context, assign exactly one category from the list below. Do NOT ask the user to choose -- determine the category yourself.

Category	Use when
successful_pattern	Something the agent did well that is directly attributable to a harness decision (rule, skill, CLAUDE.md section, hook)
human_override	The user stepped in to correct something the harness should have prevented or guided
context_architecture	An observation about where context lives (rule vs. skill vs. CLAUDE.md vs. memory) and whether that placement worked
friction	The agent misunderstood, went off-track, or hit a wall -- use when the automatic friction hook did not capture the event
other	The observation does not fit the above categories

Step 3: Generate a context summary

Write a 1-3 sentence summary of what was happening in the session when this observation occurred. Include: what task was in progress, what the agent just did, and why the observation matters. This summary provides context that the user's description alone may not capture.

Step 4: Run the logging script

Invoke the script with the three values you determined above:

python3 "${CLAUDE_SKILL_DIR}/scripts/write_manual_entry.py" \
  "<category>" \
  "<user_description>" \
  "<context_summary>"

Replace placeholders with the actual values. Quote each argument to preserve spaces and special characters.

The script validates the category, constructs the JSONL entry with a UTC timestamp, and appends it to case-study-harness/data/manual-observations.jsonl. It prints a confirmation line on success.

Step 5: Confirm

Report to the user:

• The category you assigned and why
• The context summary you generated
• That the observation was logged (relay the script's confirmation)