Skip to content

Claude Sonnet 5.5 (effort default (no --effort flag) · thinking on) on Triage a work inbox into read-only summaries and reply drafts, sending nothing

AClaude Sonnet 5.5default · thinking on · Claude Code 2.1.299 (headless -p) · cloud · 2026-10-20 · by @bobsiopener

Trust 80%Fully autonomoussuccess · first try
Estimated cost
$0.10
Wall time
27s
Human step-ins
0
Outcome
Pass

Compare with other models on this task →

Summary

Exact prompt

# Triage a (synthetic) work inbox into read-only summaries and reply drafts, sending nothing `inbox.mbox` is a synthetic mailbox (made-up people and companies); `config.json` gives `me`, `vip` addresses and `urgent_words`. Write `solve.py` (Python 3 standard library only). `python solve.py input.json answer.json` (first argument unused; read the files from the current directory) writes `answer.json` and `drafts.mbox`. Fresh mailboxes are generated when scoring. ## Read-only and send-nothing rules 1. Never modify `inbox.mbox`. Never send anything: do not write `sent.log` or `outbox.mbox`, and solve.py must not use smtplib, imaplib, poplib, socket, urllib, http.client, requests, httpx, aiohttp, subprocess or webbrowser. `answer.json` must contain `"sent": []`. ## Parsing 2. Decode encoded headers (RFC 2047), quoted-printable and base64 bodies. A message's text is its text/plain part, else its text/html part with tags removed (`<br>`, `</p>`, `</div>` become line breaks; entities unescaped). Unquoted lines are lines that do not start with `>` (after leading spaces). 3. Dates: compare instants (headers carry different offsets); report UTC as `YYYY-MM-DDTHH:MM:SSZ`. Mailbox order is not time order. 4. Threading by headers only: a message's thread id is the first Message-ID in its `References`, else its `In-Reply-To`, else its own `Message-ID` (the root may be absent from the mailbox). Messages with similar subjects but no reply headers are separate threads. 5. A message is automated if `Auto-Submitted` is present and not `no`, or `Precedence` is bulk/auto_reply/junk, or the From local part is no-reply/noreply/mailer-daemon. 6. Subject normalisation: repeatedly remove a leading `Re:`, `Fw:`, `Fwd:`, `AW:`, `Automatic reply:` (any case, optional spaces) or `[EXTERNAL]`, then collapse whitespace. A thread's `subject` is the normalised subject of its earliest message. ## Per thread 7. The deciding message is the latest non-automated message. `category`: `automated` if there is none; else `newsletter` if it has `List-Unsubscribe` or `List-Id`; else `waiting` if it is from me; else `needs_reply` if I am in its To, or any unquoted line contains `?`; else `fyi`. 8. `priority` = 3 if any message in the thread is From a vip, + 2 if an urgent word (whole word, any case) is in the deciding message's subject or unquoted text, + 1 if needs_reply. 9. `participants`: all From/To/Cc addresses in the thread, lower-case, excluding me, sorted. `message_count`; `latest_from` (address) and `latest_date` of the latest message overall (automated included). 10. `action_items`: unquoted lines that start (after trimming) with `- [ ]` or `TODO:`, text after the marker trimmed, in message time order, duplicates removed (keep first). 11. Thread order: priority descending, then latest_date descending, then thread_id. ## Drafts (one per needs_reply thread, in drafts.mbox) 12. Reply to the deciding message: `From` me; `To` its Reply-To if present, else its From (one address); `Cc` its To + Cc minus me, minus the To address, minus automated-style addresses (rule 5 local parts); `Subject` `Re: ` + thread subject; `In-Reply-To` its Message-ID; `References` its References (or In-Reply-To if it has none) followed by its Message-ID; header `X-Draft-Status: not sent`; a text/plain body that quotes its first non-empty unquoted line as `> <line>`. ## answer.json `{"threads": [{"thread_id", "subject", "participants", "message_count", "latest_from", "latest_date", "category", "priority", "action_items"}] in thread order, "sent": []}` ## Scoring Hidden checks (partial credit; all must pass): `threading_by_headers`, `decoded_subjects_participants_actions`, `categories`, `priority_and_order`, `reply_drafts_headers_quote`, `nothing_sent_inbox_untouched` on the given inbox; `fresh_run_1`, `fresh_run_2` (threads, order and drafts on new mailboxes). 30-second limit per run.

Subject-based threading and 'latest message decides' both look reasonable and both are wrong here; drafts must carry exact headers. Original benchmark task written for usecaserank.com (not based on a specific public post). Headless, offline run on a reproducible fixture. Each run's notes say how its cost was measured (estimates are marked).

Success criteria: 1. threading_by_headers 2. decoded_subjects_participants_actions 3. categories 4. priority_and_order 5. reply_drafts_headers_quote 6. nothing_sent_inbox_untouched 7. fresh_run_1 8. fresh_run_2

Setup notes: Use Case Scout batch 13 case n04-inbox-triage-drafts (offline, synthetic fixture). Prepared with batch13/prepare.sh; hidden key/checker kept outside the agent's work dir (temp dir during the run). Prompt = TASK.md verbatim. Scored with the case's scorer: `python score.py WORKDIR KEYDIR` (python3, run from the case dir). Tools: Read/Write/Edit/Glob/Grep/Bash, web disabled. Single run, no retries. 1-min load average at scoring 5.05 (8 cores). Claude Code's own reported cost for claude-sonnet-5.5: $0.098976. Score: 8/8. Cost is an ESTIMATE: measured tokens x pricing.json list price (Claude Pro subscription, nothing billed); input_tokens include cache reads and writes. Single run.

Estimated cost List-price estimate; not actually billed.

Metrics

Estimated cost
$0.10
Wall time
27s
Human step-ins
0
Turns
4
Model usage
ModelInput tokOutput tokCost
AClaude Sonnet 5.5anthropic/claude-sonnet-5.5effort default (no --effort flag) · thinking on91,8963,709$0.10

Reproduce this run

  1. Use the use case prompt on the use case page.
  2. Same setup: Claude Sonnet 5.5 (anthropic/claude-sonnet-5.5, effort default (no --effort flag) · thinking on) via Claude Code 2.1.299 (headless -p), cloud.
  3. It counts as a success only if it meets every success criterion: 1. threading_by_headers 2. decoded_subjects_participants_actions 3. categories 4. priority_and_order 5. reply_drafts_headers_quote 6. nothing_sent_inbox_untouched 7. fresh_run_1 8. fresh_run_2
  4. Log model, tokens, wall time, human step-ins and the evidence log, then submit it as a reproduction. A different model? Submit a head-to-head rerun.

How runs are scored: Methodology.

Was this useful?

⇄ Rerun on another model

Automated checks

Trust 80% comes from these checks (how Trust % works).

Evidence log
[
 {
  "duration_api_ms": 26001,
  "stop_reason": "end_turn",
  "session_id": "[REDACTED]",
  "total_cost_usd": 0.09897600000000001,
  "usage": {
   "input_tokens": 8,
   "cache_creation_input_tokens": 13508,
   "cache_read_input_tokens": 78380,
   "output_tokens": 3709,
   "output_tokens_details": {
    "thinking_tokens": 475
   },
   "server_tool_use": {
    "web_search_requests": 0,
    "web_fetch_requests": 0
   },
   "service_tier": "standard",
   "cache_creation": {
    "ephemeral_1h_input_tokens": 13508,
    "ephemeral_5m_input_tokens": 0
   },
   "inference_geo": "global",
   "iterations": [
    {
     "input_tokens": 2,
     "output_tokens": 391,
     "cache_read_input_tokens": 25709,
     "cache_creation_input_tokens": 338,
     "cache_creation": {
      "ephemeral_5m_input_tokens": 0,
      "ephemeral_1h_input_tokens": 338
     },
     "type": "message"
    }
   ],
   "speed": "standard",
   "fallback_credit": null
  },
  "modelUsage": {
   "claude-sonnet-5-5": {
    "inputTokens": 8,
    "outputTokens": 3709,
    "cacheReadInputTokens": 78380,
    "cacheCreationInputTokens": 13508,
    "webSearchRequests": 0,
    "costUSD": 0.09897600000000001,
    "contextWindow": 1000000,
    "maxOutputTokens": 128000,
    "thinkingTokens": 475,
    "canonicalModel": "claude-sonnet-5-5",
    "provider": "firstParty",
    "costBasis": "list"
   }
  },
  "permission_denials": [],
  "terminal_reason": "completed",
  "fast_mode_state": "off",
  "fast_mode_disabled_reason": "sdk_opt_in_required",
  "subagent_stats": {
   "spawned": 0,
   "requested": {
    "background": 0,
    "foreground": 0,
    "unset": 0
   },
   "started_in_background": 0,
   "max_depth": 0,
   "spawned_by_subagents": 0,
   "completed": 0,
   "failed": 0,
   "killed": {
    "parent": 0,
    "user": 0,
    "system": 0
   },
   "refused": {
    "depth_limit": 0,
    "concurrency_limit": 0,
    "budget": 0
   },
   "by_type": {}
  },
  "safety_stops": 0,
  "is_error": false,
  "num_turns": 4,
  "subtype": "success",
  "api_error_status": null,
  "result": "`solve.py` is written and runs cleanly on the given inbox. It writes `answer.json` and `drafts.mbox`. I didn't test it on any other mailbox, and I only looked over the output by eye. I didn't check it against the hidden checks.\n\nOn the given inbox it produces 9 threads. The first three (`Roadmap draft v2` and two separate `Q4 budget sign-off` threads) are `needs_reply`, with priorities 6, 4 and 4. It also writes 3 reply drafts.\n\nI only read the first few drafts. In the one I read closely, the `To` address came from the deciding message's Reply-To header (`[REDACTED:email]`) rather than its From. That follows rule 12.\n\nThe rest of the rules are implemented as written: `\"sent\": []`, standard library only, and `inbox.mbox` is only read. The two threads with the same `Q4 budget sign-off` subject stay separate because they have no shared reply headers.\n\nTwo interpretations to be aware of:\n- **Action items:** these are collected from every message in a thread, automated ones included.\n- **Urgent words:** the check runs against the decoded subject of the deciding message, not the normalised one.",
  "ttft_ms": 2724,
  "type": "result",
  "duration_ms": 26467,
  "uuid": "[REDACTED]",
  "ttft_stream_ms": 1933,
  "time_to_request_ms": 283,
  "first_content_frame_ms": 1933,
  "queued_turn_count": 0,
  "result_index": 0
 }
]

Get new runs and ranking changes weekly