AI became genuinely useful to me when I stopped treating it as a clever answer box. The unit of work is not a prompt. It is a loop: establish reality, define the outcome, route the work, inspect what changed, prove the result, and record enough state for the next run.
The human remains accountable for direction and consequences. The agents provide leverage: they can search, compare, write, test, browse, render, challenge and repeat without tiring. The system below keeps those strengths connected without pretending the machines are infallible.
The six-loop operating system
- 01OrientWhat is true now?
- 02DirectWhat outcome, scope and agent?
- 03BuildWhat is the smallest coherent change?
- 04VerifyWhat independent evidence proves it?
- 05RecordWhat must the next run remember?
- 06LearnWhat should change next time?
§1Make intent cheap, then make it precise
I dictate most prompts. Voice lowers the cost of giving context, explaining why something matters, and correcting direction. That matters because the bottleneck quickly moves from execution to articulation: the richer the intent, the less the agent has to guess.
Cheap input is not vague input. Every assignment still needs an outcome, a boundary, evidence and a stop condition. “Improve this page” is not an assignment. “Make the mobile first viewport easier to scan, preserve the existing information, verify at two widths, and stop before deployment” is.
| Work type | Route | Why |
|---|---|---|
| Ambiguous direction | One lead agent | Keep judgement and trade-offs in one context. |
| Independent investigations | Parallel agents | Fan out only when lanes can be verified and merged cleanly. |
| Mechanical edits | Fast execution profile | Use deliberation where decisions remain, not where the decision is already made. |
| Private or repeatable work | Local model when suitable | Keep sensitive context local and control marginal cost. |
| Sign-off | Fresh reviewer | The builder carries intent bias; the reviewer should inspect the artifact. |
§2Reality outranks the file that happens to be open
Agentic work becomes dangerous when speed outruns orientation. A clean local directory can be older than production. A generated file can be newer than its source. A green response can still contain a truncated answer. The system begins by identifying which surface is authoritative for the decision in front of us.
Verification then climbs an evidence ladder: inspect the source, run the focused check, scan the blast radius, render the result in a real browser, and compare the public outcome. The last step matters because users experience the deployed artifact—not our intention, local diff, or passing console output.
§3Leave the next run a better starting point
A useful run produces more than an artifact. It leaves decisions, proof, unresolved risks and a sensible next step. That state belongs in the smallest durable surface that can carry it: a commit, a handoff packet, a decision ledger, a test, a queue item or a concise memory.
The same principle drives this public notebook. Completed work is harvested for reusable lessons, checked against what already exists, and turned into one useful page at a time. After publication, search behaviour, reader paths and corrections decide whether a page should be expanded, retitled, connected, merged or left alone.
§4Different strengths, explicit ownership
| The human owns | The agents accelerate | Evidence decides |
|---|---|---|
| Direction and priority | Exploration and option generation | Whether the chosen route works |
| Risk and public consequences | Implementation and repetitive checks | Whether constraints were preserved |
| Taste and final judgement | Variants, critique and comparison | Whether the result serves the reader |
| Release permission | Preview, testing and handoff | Whether production matches the approved artifact |
Delegate execution widely. Delegate judgement deliberately. Never delegate accountability.