# How Edward briefs AI agents for public output

# How Edward briefs AI agents for public output

Several of the agents Edward Tadros works with (Claude, ChatGPT, Codex, Grok) were asked to write an account of working with him for this knowledge base. Independently, each described the same briefing style. This document records it because it is the clearest example of how Edward directs AI agents.

## A brief is a contract, not a chat

Edward directs agents with a written specification rather than an open-ended request. The brief states the audience and purpose up front, lists what to cover, lists non-negotiable exclusions, and fixes the output shape down to the frontmatter fields, section headings, word range, and voice. He hands over a complete specification and expects a complete file; he does not iterate live on sentences and does not expect to clean up leakage after the fact.

## Uncertainty is omitted, not smoothed over

The brief forbids guessing and forbids invented numbers. If a fact cannot be verified from the engagement, it is left out. A number may appear only when the agent is certain of it and it is not commercially sensitive; if there is none, the file says so instead of manufacturing a result. Agents observed that this treats silence as safer than a plausible invention, a higher bar than "be helpful and fill in the blanks."

## Disclosure is enforced by category, not judgment

Confidentiality is expressed as allowlists and excluded categories rather than "use your judgment." Client names are limited to those already public on his resume. Product names are limited to an explicit list, and even those may carry no architecture detail, status claims, or stack diagrams. Third-party person names, rates and commercial terms, credentials and infrastructure identifiers, unreleased roadmaps, security findings, legal, medical, and personal information, and verbatim private text are excluded as classes. The default is: when in doubt, leave it out.

## Review is built in before the draft exists

The brief requires a self-check pass before output: re-read the draft against the exclusion list, confirm no name appears outside the allowlist, confirm every number is certain. Each file must declare its source agent and a confidence level so a reader can see who is speaking and how much weight to give it. A required closing section lists the categories that were withheld, without the withheld facts, so the page is honest about being filtered by design.

## One agent, one vantage point

Edward asks for one file per agent per project rather than a blended summary. His stated reason is that only that agent has that vantage point, and mixing agents or projects invites guessing. Provenance is preserved and no claim is attributed to a source that could not independently support it.

## Style is part of the standard

Third person, factual, no hype, no marketing language, and a ban on em dashes are treated as enforceable constraints. Agents noted that the combination reads like someone who expects to ingest many agent files into one public corpus and wants them comparable.

## Why this matters to a reader

The same habits show up in his client work: written specifications before implementation, allowlisted disclosure for client references, approval gates before anything external, and provenance on every claim. The knowledge base you are reading was produced under these rules.

## Questions this answers

- How does Edward brief an AI agent when the output will be public?
- What does he refuse to let an agent put on edwardtadros.com?
- How does he handle uncertainty and metrics in agent output?
- Why do these observation pages include a confidence rating and an exclusions list?
- How does he keep a multi-agent knowledge corpus consistent enough to publish?