What happened
The Coollinks page presents several practices for using AI in day-to-day company work. For estimating time spent on a task, the author asks Codex to inspect project information. The suggested inputs include access to projects so Codex can review Git revision history, work chats, and the author's history of conversations with Codex, exported into a separate conversation index. The author says Codex estimated 180 hours for developing the latest product and 20 hours for preparing a training module. These figures refer only to the author's own hours; the author says the system cannot yet total the hours of the team.
A second practice is to connect a company knowledge base to work chats, email and bank exports when investigating accounting records and unresolved items. The author built a separate bridge service for local data. It runs on a Mac and provides read-only access to a local Telegram database and an email agent's database. The service also maintains a cache in a format designed for search and provides an interface using progressive disclosure. The account says this setup helped locate tax items, pay them, check the results, organize the records into folders and send a bug report to a bank about a problem with its export system.
The author also describes using the company knowledge base to identify opportunities for process improvement. The stated routine is to ask Fable 5.2 to act as an experienced startup adviser, review recent changes in the knowledge base, conduct research and suggest one small improvement. The author says this can surface gaps that seem obvious to others but have been missed internally, and prompt ideas for connecting additional data sources to existing processes.
The page does not describe an enterprise deployment or a general-purpose product launch. It is one practitioner's account of combining access to personal work records, local-data integration and an AI assistant. The author does not specify the underlying models beyond naming Codex and Fable 5.2, or provide an implementation guide for the bridge service. The practices are described as things the author does, not as tested instructions for other organizations.
What the archive adds
New Runtime's earlier coverage on bad company data and GTM agents concerned the quality of company data used by agents, not this author's workflow. This is a new account of personal operational practices, rather than a new fact in that earlier story.
Trend signal
This does not yet fit a pattern New Runtime tracks. The account combines searchable company context with human operational work, but it does not establish a governed write-and-correction loop or show how the approach performs beyond the author's examples.
Why now
The author frames these practices as recent discoveries and describes a continuing effort to improve company processes. The page gives no external launch or calendar-driven reason for the timing.
Connections to other events
The practices meet a broader operational question also raised by New Runtime's earlier coverage of company-data quality for GTM agents: what information an agent can use shapes the work it can support. Here, the emphasis is on the user's own records and access method, rather than on comparing commercial data providers.
Why it matters
For teams experimenting with AI at work, the account is useful less as proof that an agent can manage operations than as a reminder to scope the job and its data carefully. The time-estimation example distinguishes personal hours from team totals, an important boundary if estimates might influence staffing or project planning. Treating the number as a model-generated reconstruction—not a timesheet—avoids implying a level of precision the author does not establish.
The local-data example is more concrete about access: the bridge is described as read-only, with a searchable cache and progressive disclosure. Those are design choices teams can examine when connecting private records to an assistant. They do not, on their own, establish security or privacy; a team would still need to assess what is cached, who can access it and how records are retained. The improvement routine also keeps the ask small: review recent changes and suggest one next step. That may make recommendations easier to inspect than an open-ended mandate to redesign operations.
Limits and uncertainty
The account does not independently validate the 180-hour and 20-hour estimates, explain how Codex arrived at them, or report whether the figures were checked against contemporaneous time records. It explicitly says team-wide hours are not available from this approach. It also does not quantify the time saved or the accuracy of the accounting work. The bridge service's read-only access is described, but its authentication, encryption, retention, deployment safeguards and threat model are not. The page does not explain Fable 5.2's research process or whether its recommendations were adopted and produced measurable improvements.
Skill to build now
If you want to try project-time reconstruction, give an assistant a bounded task and a defined evidence set: for example, ask it to estimate your own hours for one completed project from Git history and relevant work conversations. Ask it to show the evidence behind the estimate and keep personal time distinct from team totals. For local records, begin with read-only access and a limited set of sources; expand only after reviewing what the assistant can retrieve and what its cache retains.
Reader takeaway
A useful company assistant depends on more than a capable model: it needs the right work context and carefully scoped access. This account offers practical starting points—historical evidence for personal estimates, read-only bridges for local records and small, reviewable improvement requests—but not validated performance or a ready-made blueprint.