Custom AgentUndisclosed

Texting meals and morning weight to an agent for six months: OpenClaw logged the food, redrew the graph, and the author reports ~94 kg down to ~80 kg

Lex Toumbourou(博客署名作者,写自己跑了六个月的 OpenClaw 实例)Blog · 2026-07-15

Read original

Saved text / editorial summary

I share that weight with OpenClaw, which logs it, regenerates the progress graph below and kicks off the second phase of my daily routine, described later … Since I started tracking in late 2025, I've gone from ~94 kg to ~80 kg - close to a 14 kg loss.(编辑注:原帖照片说明逐条标注 Jan 2026 - 91.5 kg、Jul 2026 - 81.4 kg,非引文正文)

Business model

Blogger Lex Toumbourou wrote a six-month retrospective on his OpenClaw setup: it runs on an old laptop used as a server, its workspace is a folder inside his Obsidian vault, and he talks to it over a WhatsApp account created with a separate SIM card. Every morning he sends his weight; the agent logs it, regenerates the progress graph, and starts the rest of his daily routine. Meals (checked against a common-foods file for calories), strength sessions, stretching, injury recovery and sickness notes all go in by message, and weekly body-progress photos are saved into a table. Email, finances (PocketSmith) and investments are all read-only - the agent can help find and triage email but cannot send anything on his behalf. He reports going from ~94 kg to ~80 kg since he started tracking in late 2025 (91.5 kg in Jan 2026, 81.4 kg in Jul 2026), with blips during holidays. This is one person's log, not a controlled experiment: he set the calorie-deficit goal himself, the weight loss cannot be attributed to the tool, and the agent's job is recording, summarizing and graphing. Cost context: 6-21 million tokens per day, mostly cached, which at Opus 4.8 API prices would be roughly 10-40 dollars a day, so he runs GPT models through a 20-dollar-per-month ChatGPT Plus subscription instead.

The reusable idea is friction: recording survives six months when the entry point is one chat message instead of a form. Two more pieces travel well - a strict read-only boundary on email/finance/investment accounts with writes confined to the note vault, and a subscription-based model choice that absorbs 6-21 million tokens a day because most of the traffic is cached. The weight number itself is not the product; it is the author's own dieting outcome that the infrastructure merely recorded.

Workflow

  1. Set up a resident server: OpenClaw runs on an old laptop, its workspace is a folder inside the Obsidian vault, and an AGENTS.md file describes the vault structure. (The author's own setup.)
  2. Isolate the channel: a separate SIM card creates a dedicated WhatsApp account for the agent, keeping the family's main number untouched. (The author's own setup.)
  3. Daily input by message: morning weight (measured at a consistent time), each meal as text or photo, workouts, stretching, sickness notes, and a weekly body-progress photo; the agent checks a common-foods file for calories. (The author's own routine.)
  4. Agent-side summarizing: entries go into daily notes, the weight graph is regenerated after each weigh-in, photos land in a weekly table, and Oura, PocketSmith, portfolio, weather and birthday scans are pulled into the vault. (Described by the author; implementation details are not spelled out in the post - AgentFlow's summary of his description.)
  5. Permission boundary (AgentFlow's recommendation extending the author's rule): keep email, finances and investments read-only, confine writes to the note vault, and require human confirmation for anything sent externally.
  6. Reuse path (AgentFlow's inference): run a minimal message-to-log loop for two weeks first, then add the graph, the photo table and external sources one at a time, accepting that every meal and workout now means pulling out the phone - the author's own stated complaint.

Adaptation plan

The post does not detail how the graph regeneration or photo table is implemented, which GPT model versions run the sessions, how the Oura and PocketSmith connections are authorized, the second complaint, which the post states in full: it is a new software project that 'is going to be broken a bunch', patched quickly by the team but hard to follow on a GitHub issue board flooded with AI-generated slop and bot spam. The weight figures are self-reported; there is no before/after verification, no food-logging accuracy audit, and no statement of what happens when a meal is never sent.

Prompt

AgentFlow draft · Untested.

You are my life-log assistant, resident in my personal note vault. You only record, summarize and remind. You never send anything externally and never modify anything outside the vault; any send, payment, or change to an external account stops and asks me first.

[INPUT]
- Messages I send anytime: morning weight (number + unit), meals as text or photo, workouts, stretching, body status (sickness/injury), weekly body-progress photos
- Common-foods calorie table: [table path]; if a food is missing, ask me - do not estimate on your own
- Vault structure description: [AGENTS.md path]; all records are appended to the matching daily note
- External sources (read-only): [wearable data], [bookkeeping app]; pull one summary per day
- Output formats: append to daily notes; regenerate the weight graph at [chart path] after every weigh-in

[STEPS]
1. Classify each incoming message (food / weight / workout / stretching / status / other); if you cannot classify it, ask me instead of guessing.
2. For each food item, look it up in the common-foods table: on a hit write 'food + portion + calories'; on a miss list it for me; describe photo contents and let me confirm.
3. On a morning weight: append the record, regenerate the graph, report the difference versus last week (numbers only, no health commentary).
4. At the fixed daily time, summarize: total calories, training completed, external-source digest, reminders due today.
5. Turn reminder-type messages into to-do items in the daily note so they surface in the daily summary when due.
6. Weekly: file the week's body photos into the photo table and summarize weight change, training counts and anomalies.

[OUTPUT]
One-line confirmation per message (where it was recorded, what was computed). One daily summary: calorie total, training status, external digest, to-dos. One weekly summary: weight change, photo-table update, anomalies (missed meals, interrupted sessions).

[CONFIRM]
Ask me and wait for my explicit yes before: any external send (message/email), any payment or subscription change, substituting an approximate value for a missing food, or reconnecting an expired external authorization. You never send email on my behalf and never modify files outside the vault.

Running costs and pitfalls

The post does not detail how the graph regeneration or photo table is implemented, which GPT model versions run the sessions, how the Oura and PocketSmith connections are authorized, the second complaint, which the post states in full: it is a new software project that 'is going to be broken a bunch', patched quickly by the team but hard to follow on a GitHub issue board flooded with AI-generated slop and bot spam. The weight figures are self-reported; there is no before/after verification, no food-logging accuracy audit, and no statement of what happens when a meal is never sent. No independent cost benchmark is available. Check subscriptions, permissions and external actions before running a workflow.

Author-reported results are not independently audited.

Original text and media remain attributed to their authors.