Impact log

impact-log

Turn the work in the current Claude Code session into a structured, local-only accomplishment record — then search it, total it up, and export it for resumes, reviews, and interviews. Nothing is ever sent anywhere; all data lives in a JSON file on this machine.

The skill is AI-first: you analyze the session and write the entry. The driver (impact_log.py, Python 3, stdlib only) just persists, searches, and reports. Do not make the user fill out a form.

Storage: ~/.impact-log/logs.json (override with the IMPACT_LOG_HOME env var). All commands below are run from this skill's directory; impact_log.py is here.

Capture a new accomplishment — /impact-log (the default)

This is the main flow. Do NOT interrogate the user. Instead:

1. Gather evidence from what's already available, in this order:
   • the current conversation (code written, bugs fixed, decisions made — you already have this context)
   • git diff, git status, git log --oneline -15 if this is a git repo
   • files recently changed in the session
2. Infer the fields: title, category, problem, solution, impact, tech (list), and — when you can estimate it — time_saved_hours. Set a confidence between 0 and 1.
3. Decide based on confidence:
   • ≥ 0.8 — present the drafted entry and ask only Save? [Y/n].
   • < 0.8 — ask ONLY for the specific missing piece (usually impact / time saved). One question, not five. Then save.
4. Write the entry to a temp JSON file and save it (see below).

category must be one of: automation, ai, performance, deployment, frontend, backend, security, bugfix, devops, database, architecture, other (anything unrecognized is stored as other).

Saving an entry (file-based — works on every shell)

Write the entry JSON to a temp file, then call add --file. Do not pipe JSON via stdin on Windows — PowerShell re-encodes it (UTF-16/BOM) and the parse fails.

PowerShell (the user's default shell):

$tmp = "$env:TEMP\impact-entry.json"
@'
{"title":"Task Assignment Automation","category":"automation","problem":"Manual task creation took 30+ min daily.","solution":"Built JWT-authenticated API automation.","impact":"Daily effort dropped to under 5 minutes.","tech":["Python","REST API","JWT"],"time_saved_hours":25,"confidence":0.94}
'@ | Out-File -FilePath $tmp -Encoding utf8
python impact_log.py add --file $tmp
Remove-Item $tmp

Bash:

cat > /tmp/impact-entry.json <<'EOF'
{"title":"Task Assignment Automation","category":"automation","problem":"Manual task creation took 30+ min daily.","solution":"Built JWT-authenticated API automation.","impact":"Daily effort dropped to under 5 minutes.","tech":["Python","REST API","JWT"],"time_saved_hours":25,"confidence":0.94}
EOF
python impact_log.py add --file /tmp/impact-entry.json

id and date are assigned automatically (id is YYYY-MM-DD-NNN, sequential per day). On success it prints saved <id> "<title>" [<category>].

Other commands

User says	Run	Does
/impact-log recent	python impact_log.py recent	last 10 entries (pass a number for more)
/impact-log search <q>	python impact_log.py search <q>	full-text search across all fields
/impact-log stats	python impact_log.py stats	totals, category breakdown, time saved
/impact-log export	python impact_log.py export	writes impact-report.md (use -o for a path)
—	python impact_log.py path	print the storage file location

After running a command, relay the output to the user in a readable way — don't just dump it.

Entry schema

{
  "id": "2026-06-20-001",        // auto-assigned
  "date": "2026-06-20",          // auto-assigned (today) if omitted
  "title": "Task Assignment Automation",
  "category": "automation",
  "problem": "Manual task creation took 30+ min daily.",
  "solution": "Built JWT-authenticated API automation.",
  "impact": "Daily effort dropped to under 5 minutes.",
  "tech": ["Python", "REST API", "JWT"],
  "time_saved_hours": 25,        // optional; only this feeds the stats total
  "confidence": 0.94             // optional
}

Only title is required. time_saved_hours is the only field that feeds the "Estimated Time Saved" total — include it whenever you can defensibly estimate it, otherwise leave it out (stats will say none is recorded rather than inventing one).

Gotchas

• Never pipe JSON to add via stdin on Windows. PowerShell pipes encode as UTF-16, so Python reads garbage. Always write a temp file and use --file. The driver reads files as utf-8-sig, so a BOM from Out-File -Encoding utf8 is fine.
• Privacy is the selling point. Data only ever touches ~/.impact-log/ (or IMPACT_LOG_HOME). Don't add any step that uploads or transmits entries.
• Don't fabricate impact. If you can't estimate time saved or business impact, ask one question or leave the field empty — don't make up numbers.

Troubleshooting

• error: input is not valid JSON — the temp file isn't valid JSON (often a stray trailing comma, or you piped via stdin on PowerShell). Re-write the file and use --file.
• error: ... logs.json is corrupt JSON — the store was hand-edited into an invalid state. Open python impact_log.py path to find it and fix/remove it.
• python not found — try python3 impact_log.py ....