The honest build log.

A web app that accepts a résumé and job description, then identifies missing important keywords. The core emerged in three sessions; accounts, recovery, privacy decisions, and reliable parsing took much longer.

Claude and Claude Code helped build it. An always-on agent handled the operating environment, deployments, logs, backups, and changes.

Scope, brand direction, privacy decisions, error hunting, release judgment, and the decision not to charge before the product earned trust.

Stacked build sessions made the first version fast. About 100 real résumés and more than 20 fix-test rounds returned messy-file parsing to strong performance.

Clean test files hid the real problem. Performance fell sharply on two-column, scanned, old, and unusually formatted résumés. Security and code reviews found additional holes.

A demo can happen in a day. A trustworthy product requires weeks of human-led testing, regression coverage, and duty-of-care decisions.

A reviews-and-news site about agentic AI, with distinct editorial formats and living resources that can be updated when tools and prices change.

Research, drafting, source collection, and repeatable publishing labor.

Editorial direction, voice, public approval, source standards, and the rule to ask when information is missing instead of guessing.

Separate skills preserved the intended format and tone for reviews, guides, takes, and daily briefs.

Early systems exposed the familiar risks: stale information, generic voice, and the temptation to let confident output outrun evidence.

Honesty is not a prompt. It is a rule supported by sources, review, and refusal behavior.

A multi-station workflow supporting agent-assisted online business operations, from trend research and design variants to draft listings, captions, and email work.

Bulk research, option generation, metadata, drafts, and repetitive production steps.

Every design is inspected. Brand voice, claims, platform compliance, approval, and which products deserve further effort remain human decisions.

Agents changed the content of the workday, reducing repetitive production and expanding the number of ideas that could be tested.

Automated queues collided, posts silently vanished, and systems sometimes reported success when the intended outcome never appeared.

Outcome monitoring matters more than a green status. Automate one station, keep approvals, and alert on absence.

Always-on operating agents with durable tasks, private access, logs, backups, scheduled work, and written operating habits.

Recurring research, operations, content, monitoring, and deployment work.

New jobs are introduced manually, permissions stay bounded, live state outranks memory, and a person owns the kill switch.

Moving from an unstable first framework to Hermes reduced maintenance and made durable task recipes reusable.

Files wandered when stale memory overruled the machine’s live state. Other jobs quietly overwrote queues. Credentials appeared in scripts and retired code.

Memory is a convenience, not truth. Check live state first, keep secrets out of code, and make silent failure loud.