ESF v0.7

ESF is an operating contract: a written set of rules I hold myself to when producing an engineering strategy report. It is not a template. It says what counts as evidence, what to do when a fact is missing, and in what order the thinking has to happen.

How it runs

The framework is written as instructions an AI agent can execute. The agent does what machines are good at: it reads the whole repository, the git history, the pull requests, the public ledgers, the blogs and podcasts, and it drafts the prose. The reasoning is mine: which risks are real, what the evidence actually means, what to recommend and what to cut. That division of labour is the lever: data mining and writing at machine speed, judgement and accountability staying with a person. It is what makes a report of this depth possible in a day instead of a month.

The rules that matter

Evidence before reasoning. No risk, root cause, or recommendation gets written unless it is anchored to something observed: code read, a command run, a metric, a dated document. A confident report built from assumptions is the failure mode this whole contract exists to prevent.

Every claim carries its source. Strictly five tags: observed, web, stakeholder, inferred, assumed. Anything that was neither seen nor verified goes into a Missing Information section instead of hiding inside confident prose. That is where each report's honesty meter comes from.

Missing facts get fetched, not assumed. What only a stakeholder knows gets asked; what the public record knows gets searched. An assumption is allowed to survive only when neither would resolve it, and then it is labelled as one.

Rank, then cut. Risks are ranked by impact, probability, and how silently they would bite; then the list is forced through "if only three things happen this quarter, which three, and why do the rest wait?" A list that doesn't prune is not a strategy.

Red-team the conclusion. Before a report is finished, the strongest case that its top recommendation is wrong gets argued on the page. And when someone disputes a claim, the rule is to widen the search, not defend the finding. A claim that survives usually gains nuance, and a claim that dies gets corrected with the same rigour that produced it.

Three shapes, composable

A report takes one of three shapes: a greenfield initiative, an audit of an existing system, or a single decision, and real engagements usually are a composition of them. "System" need not mean code: the same discipline audits a team, a business, or a career; only the evidence sources change.

On AI

A strategy written today has to handle AI. ESF assesses every initiative as one that may be run by agents, not humans: delegable, parallelizable work gets estimated in agent-hours (an order of magnitude away from person-hours) while judgement-gated work keeps its human price. The strategy itself obeys the same rule: written to be accepted by humans and executed by agents, because agents now read strategy more often than people do. Even if some chapters are written for specific humans, there is a big probability they are executed by their agents anyway.

Lineage

I invented little of this, on purpose. ESF is an opinionated compilation: battle-tested methods, compiled into a contract an agent can execute. Three main ideas.

Will Larson. Crafting Engineering Strategy is the closest sibling. I import his explore → diagnose → policy spine, strategy altitude, and the claim that there is always an engineering strategy, even when nothing is written down. He teaches a human to do strategy; ESF is the runnable version.

Continuous Improvement. The Toyota and Deming line: kaizen, double-loop learning. It gives the framework its second loop: every cycle has to improve the machine that runs the cycles, not just the product it ships. Because engineering is iterative, and strategy should account for that.

Harness Engineering. Ryan Lopopolo's practice of improving agent output by shaping the environment around the agent. That is why ESF is written as an environment of rules, gates and checks rather than as advice. Sometimes a small change to environment leads to major compounding result over time.

Where it stands

The framework has been run against real production systems, and against one fully public worked example: the Solidus audit, built entirely from public evidence. The repo link will be published here, changelog and all; every report on this site names the version it was written against.