The first move is configuring the model itself, the universal settings that govern every conversation. This is the next layer down: how to make an AI assistant behave with operator-grade discipline inside a single project.

Real work does not run on one-off prompts. It runs on projects. A deal you are screening. A board pack you are building. A portfolio company you are diagnosing. These span days, files, multiple sources, and a sequence of decisions. The unit of useful AI work for a leader is the project, not the chat.

AI fails inside projects in three structural ways. Calibration is the most obvious: it asks five clarifying questions on a one-line formatting task, then accepts a multi-stakeholder board memo with no checking at all. Sourcing is the more dangerous failure: a confident citation to a single weak post, with the occasional invented specific filling gaps the real sources could not cover. The handoff failure wastes time after the work is done: a 3,000-word file lands in your inbox with no signal about what to read first, what blocks your next decision, or what to skip.

Each failure mode has a fix. The fix is a named protocol you write once, save as a file in your project's knowledge, and reuse across every project after. These are the three I run on everything.

THREE FAILURES, THREE PROTOCOLS THE FAILURE THE PROTOCOL WHAT IT FORCES Calibration. Five clarifying questions on a formatting task, then no checking at all on a board memo. The Socratic Protocol Question regime set by task size. Trivial, substantial, or complex, named before the work starts. Sourcing. A confident citation to one weak post, with invented specifics filling the gaps. The Research Protocol Source tiers and triangulation. Two independent Tier A or B sources on any critical claim. Handoff. A long file lands with no signal about what to read first, or what blocks the next decision. The File Delivery Protocol Four elements on every output. Purpose, read priority, blocking decisions, status and assumptions. Each protocol is a file in the project's knowledge base, so the discipline loads on every project after. Exhibit | TeakCharge
Three structural failure modes inside a project, and the protocol that closes each one.

Protocol 1: calibrate questioning to task size

Default AI behaviour is uncalibrated. The Socratic Protocol forces the model to pick a regime based on task size, name the regime it picked, and act accordingly.

  • Trivial work gets zero questions. Reformat this transcript, summarise this prospectus, deduplicate this list. Execute without asking.
  • Substantial work gets one or two grounded options, not open-ended interrogation. "I see two ways to size this market: bottom-up from observed deal volume or top-down from total spend. Which fits, or should I assume top-down?"
  • Complex work gets a full read-back before any execution: understanding, approach, expected output, key assumptions, then a check. Right on this, or redirect?

The grounded-options pattern is the senior-advisor signal. A framed choice reads as a peer thinking. "What are your priorities?" reads as an assistant fishing. Have the model state the regime it chose before it starts, so you can see how it is calibrating before you read a word of content.

Protocol 2: source credibility before assertion

The worst failure in AI research is not the wrong answer. It is the confidently wrong answer, sourced from one weak link and presented as established. The Research Protocol forces source-tier discipline and triangulation before any critical claim ships.

  • Tier A: primary and official. Filings, regulatory documents, peer-reviewed work, direct quotes from named officials.
  • Tier B: established secondary. Reuters, Bloomberg, the FT, the WSJ, Nikkei Asia, major published research.
  • Tier C: trade press and named-expert commentary. Useful, but weaker.
  • Tier D: forums, anonymous posts, undated content, AI-generated summaries cited elsewhere. Discount or exclude.

Two rules do the heavy lifting. Triangulation: critical numbers, dates, and named attributions need two or more independent Tier A or B sources, and two copies of the same wire story count as one. Anti-fabrication: if a specific is not in the sources, the model says "not in sources" rather than inventing one. "A Tier 1 PE firm" beats a named firm the sources never mentioned. A generic placeholder beats a fabricated specific every time.

Protocol 3: outputs that get read

The handoff is the highest-leverage moment in any AI session, and the default is to drop a file and wait. An executive needs to know in five seconds what is in the file, what blocks the next decision, and what to read first. The File Delivery Protocol requires four elements on every output.

  • File and purpose, one sentence each: what it is, how long, how many exhibits.
  • Read priority and time: read this first, five minutes, skim section three, skip the appendix.
  • Blocking decisions with next step: what needs your input, and what happens on each branch of your answer.
  • Status flag with open assumptions: draft v1, three assumptions listed, ready for review but not for the board.

No file ships without those four. Multi-file deliveries get a one-paragraph index at the top.

Making them yours

Save each protocol as its own file in the project's knowledge base, then reference it by name in the project instructions. The instructions box is the trigger; the knowledge files carry the doctrine. Spend twenty minutes tailoring the thresholds and the source tiers to how you actually work, and the discipline loads automatically on every project after. Next comes the layer that makes corrections persist between sessions rather than vanishing when you close the chat.

Written by Alex Szabo. Alex Szabo is the founder of TeakCharge Commercial Strategy and Implementation.