whatidid

name	whatidid
description	Generate a daily activity report from Claude Code session data. Harvests human/assistant turns from ~/.claude/projects/*/*.jsonl, analyzes them with claude-haiku-4-5 via the Anthropic API, and renders an HTML report with KPI cards, goals table, and effort/leverage breakdown.
argument-hint	[YYYY-MM-DD]

What Did I Do — Daily Activity Report

Overview

Produces a structured "what did I do today" report by reading Claude Code session JSONL files, extracting user instructions and tool summaries, calling the Anthropic API for goal/task analysis, and rendering an HTML report with KPI cards, goals table, and token cost breakdown.

The report answers: what goals were pursued, what tasks were completed, how much time was spent, and what the ROI was relative to the seat cost.

When to Use

• At the end of a workday to generate a standup/status report
• When you need an objective log of what was worked on for a given date
• When asked to "report on today", "what did I do", or given a date argument

Not for: Real-time session introspection, cross-repo analysis beyond git metadata, or project planning.

Helpers & References

File	Purpose
helpers/whatidid.rs	Entrypoint — orchestrates harvest → analyze → report
helpers/harvest.rs	Scans ~/.claude/projects/*/*.jsonl for today's sessions
helpers/analyze.rs	Calls Anthropic API; caches result to cache/YYYY-MM-DD.json
helpers/report.rs	Renders HTML with KPI cards and goals table; opens in browser
references/architecture.whatidid.md	Data flow, session format, token cost model
references/effort-estimation.whatidid.md	Effort estimation heuristics

---

Steps

1. Resolve the target date

Use $ARGUMENTS if provided (format: YYYY-MM-DD). Otherwise default to today.

2. Harvest session data

Run helpers/harvest.rs [YYYY-MM-DD] via rust-script.

Scans all ~/.claude/projects/<project-slug>/*.jsonl files. Each JSONL line has a timestamp field (ISO 8601). Events with a matching date prefix are included.

Key event types extracted:

• type: "human" → message.content[].text (user instructions; noise-filtered)
• type: "assistant" → message.content[] text blocks + tool_use names
• Metadata from every line: cwd, gitBranch, sessionId, slug

Noise filtered from human turns: single-word confirmations ("ok", "yes", "act"), <system-reminder> injections, <current_datetime> tags.

Output: JSON array of session records with messages[], tool_calls, read_calls, edit_calls, bash_calls.

3. Analyze with Anthropic API

Run helpers/analyze.rs <sessions.json> [YYYY-MM-DD] via rust-script.

Requires ANTHROPIC_API_KEY in environment. Inject via:

op run --env-file=$HOME/dev/.env -- rust-script helpers/whatidid.rs

Builds a transcript from harvested sessions (cwd, branch, SIGNALS block with tool counts, then interleaved human/assistant turns). Sends to claude-haiku-4-5-20251001.

Model prompt lives in prompts/analysis.txt. Returns structured JSON digest: goals[] each with tasks[], human_hours, domain_skills, tech_skills.

Caches result to cache/YYYY-MM-DD.json — re-runs skip the API call.

4. Compute KPIs

total_human_hours  = sum(goal.human_hours for goal in goals)
human_value        = total_human_hours × 72          # $72/hr blended rate
seat_cost_per_mo   = 39                              # Claude Code Pro seat
leverage           = human_value / seat_cost_per_mo

Token cost model: prefix-matched against report.rs → MODEL_PRICING. Fallback: $3.00 input / $15.00 output per 1M tokens.

5. Generate HTML report

Run helpers/report.rs <digest.json> [YYYY-MM-DD] via rust-script.

Layout:

Header bar   — date, primary_focus, session count
Narrative    — day_narrative (2 sentences from digest)
KPI cards    — Hours | Human value | Leverage
Goals table  — Goal | Tasks (bullet list) | Hours | Skills

Written to /tmp/whatidid-YYYY-MM-DD.html. Opened automatically via open.

6. Report to user

[1/3] harvesting sessions for 2026-05-17...
[2/3] analyzing with Claude...
[3/3] rendering report...
Date: 2026-05-17  |  Goals: 5  |  Hours: 6.5h
Human value: $468  |  Leverage: 12x
Report: /tmp/whatidid-2026-05-17.html

---

Token Cost Model

Provider	Model prefix	Input $/1M	Output $/1M
Anthropic	claude-opus-4	$15.00	$75.00
Anthropic	claude-sonnet-4	$3.00	$15.00
Anthropic	claude-haiku	$0.80	$4.00
OpenAI	gpt-5, gpt-4o	$2.50	$10.00
OpenAI	gpt-4.1	$2.00	$8.00
OpenAI	o3	$10.00	$40.00
Google	gemini-2.5-pro	$1.25	$10.00
Google	gemini-2.5-flash	$0.15	$0.60

Fallback (unknown model): $3.00 input / $15.00 output.

Update report.py → _MODEL_PRICING when rates change.

Common Mistakes

• Including injected context as user instructions — filter <current_datetime>, <system-reminder>, and single-word approval messages ("ok", "yes", "go") before building the transcript.
• Double-counting sessions — a single working session may span midnight; use created_at (session start) not updated_at as the date key.
• Calling the API without a cached result check — always check cache/YYYY-MM-DD.json before hitting GitHub Models; the API has rate limits.
• Hardcoding model names — use prefix matching in _MODEL_PRICING so new model variants don't silently fall through to zero cost.