Engineering Field Notes

Evidence before
the conclusion.

Two practical problems documented through constraints, decisions, available evidence, and honest limitations.

Private records and third-party information are omitted or replaced with clearly labeled reconstructions.

Personal workflow / February–June 2026

Student Operations Automation

Routing academic updates through a private destination while retaining review, privacy, and failure visibility.

Private prototype
Role
Workflow designer and reviewer
Tools
OpenClaw, Python, JavaScript, Discord
Scope
3 workspaces / 12 scheduled tasks
Evidence
5,894 recovered scheduler records

01 / Problem and constraint

Fewer places to check, without exposing classroom data.

Reviewing announcements across multiple Google Classroom subjects was cumbersome. I wanted one designated Discord destination for relevant updates, reminders, and personal study outputs.

The public explanation could not include class content, account identifiers, participants, or private server data.

02 / Decisions and implementation

Separate workspaces, reusable instructions, and purpose-based routing.

I configured three separate OpenClaw workspaces, wrote reusable instructions, set up 12 scheduled tasks, and routed selected updates to Discord.

I used AI assistance for parts of the Python and JavaScript integration, then reviewed the resulting behavior and records.

  1. 01CollectAccessible course updates
  2. 02ReviewInspect collected records
  3. 03RouteSeparate by purpose
  4. 04DeliverDesignated private channel

Sanitized workflow ledger

  1. 01
    Source boundary

    Accessible course updates were collected from the configured workspaces; classroom and account identifiers are omitted.

  2. 02
    Review gate

    Collected records were reviewed before selected updates were routed onward. This public note does not claim a verified deduplication guarantee.

  3. 03
    Purpose route

    Three separate OpenClaw workspaces and 12 scheduled tasks kept destinations separated by purpose.

  4. 04
    Private delivery

    Selected updates were delivered to a designated Discord destination; this note does not claim complete delivery or time saved.

Assistance & verification

AI assistance supported the integration; the decisions and checks remained mine.

AI assisted

Parts of the Python and JavaScript integration used AI assistance.

I decided

I separated the work into three OpenClaw workspaces, wrote the reusable instructions, defined the 12 scheduled tasks, and chose what routed to Discord.

I verified

I reviewed the resulting behavior and the recovered scheduler records, including successful, skipped, and failed executions.

Available evidence

This public note provides aggregate scheduler counts and a simplified workflow reconstruction. It does not expose classroom, account, or server data.

03 / What the records show

Total scheduler records
Successful
Skipped
Failed

These are execution records—not unique announcements or delivered messages. I did not measure time saved; the practical benefit was having fewer places to check.

04 / Validation and limitations

01

No complete automated test suite or CI pipeline was present in the recovered prototype.

02

The caption-processing component was a short personal pilot, while other scheduled workflows ran longer.

03

A public-safe rebuild should use synthetic fixtures, least-privilege access, retries, deduplication tests, and explicit retention controls.

Personal data-quality workflow

Task Archive Safety Review

A human-in-the-loop process for reviewing a large personal archive without allowing status messages to substitute for verification.

Read-only review
Role
Requirements and review
Focus
Data validation and recovery
Scale
More than 10,000 task records
Control
Backup and approval required

01 / Problem and risk

Manual review was impractical; unsupervised reorganization was unsafe.

The archive had grown too large to review casually, but allowing an agent to reorganize it without safeguards risked irreversible loss of history and context.

The system needed to make uncertainty visible and preserve a recoverable path before any edit was considered.

02 / Safety process

  1. 01
    Define boundaries.

    Identify protected fields, acceptable matches, and success criteria.

  2. 02
    Analyze without mutation.

    Produce candidates and reports before changing source data.

  3. 03
    Verify claims.

    Compare the reported result with the actual destination state.

  4. 04
    Build a recoverable copy.

    Require a backup before any irreversible operation.

  5. 05
    Validate structure.

    Confirm row totals and identifier integrity across more than 10,000 records.

  6. 06
    Gate destructive actions.

    Keep merge, archive, completion, and deletion behind human approval.

Assistance & verification

The agent assisted with review, but its status message was not accepted as proof.

AI assisted

An agent assisted with reviewing the large archive and reporting its result.

I decided

I defined the protected fields, acceptable matches, backup requirement, and human approval gates for destructive actions.

I verified

I compared the reported completion with the actual result and checked row totals and identifier integrity in a recoverable copy.

Available evidence

This public note provides the aggregate scale and a reconstruction of the review process. Personal task content is not published.

03 / Result and lesson

records in a structurally checked copy

The review caught a mismatch between an agent’s completion message and the actual result. The lesson was direct: reported completion still requires source-level verification.

Next

Return to the work,
or start a conversation.

View selected workEmail John Charles