Engineering Strategy Framework (ESF)

An operating contract for engineering strategy: reduce risk, compound the delivery machine, and maximize the probability that the right thing gets built under uncertainty. It is a portable agent skill — plain Markdown, no build step — so it runs in any harness that supports skill-style instructions.

Three composable modes. A greenfield initiative gets the full report structure; an existing-system audit (of code, a team, a business, or a career) leads with evidence and root causes; a single decision scales to the reversibility and cost of one call. The discipline they share is the point: gather evidence before reasoning, tag every claim with where it came from, rank then cut, red-team your own conclusion, and publish what you could not establish.

Installation

Skills CLI (any harness)

npx skills add vit-panchuk/esf --global --agent '*'

Omit --global for a project-local install a team can commit. Update with npx skills update esf --global.

Claude Code plugin

/plugin marketplace add vit-panchuk/esf
/plugin install esf@esf

Copilot CLI

The repository is its own marketplace, the same way it is for Claude Code — both read the same .claude-plugin/marketplace.json:

copilot plugin marketplace add vit-panchuk/esf
copilot plugin install esf@esf

(copilot plugin install vit-panchuk/esf also works today, but GitHub has deprecated installing straight from a repository; only plugin@marketplace will be supported in a future release.)

Codex

codex plugin marketplace add vit-panchuk/esf
codex plugin add esf@esf

Codex keeps its catalog at .agents/plugins/marketplace.json rather than the .claude-plugin/marketplace.json the other two read, so this repository ships both. The plugin itself is the same package for all three — plugin.json at the root, skills under skills/, per Agent Plugins 1.0.

Start a new session after installing: bundled skills only become available in a session opened afterwards. Invoke it by mentioning the skill ($engineering-strategy-framework) or by describing the task and letting Codex route to it.

Two other routes exist if that one does not fit. A workspace admin can import this repository for everyone under Admin → Plugins → Add → Import marketplace (public and private repositories, synced daily). And once the plugin is listed in the shared ChatGPT/Codex directory, /plugins in a session will find it without any marketplace being added.

Manual

The runtime artifact is SKILL.md, so any harness can use it directly — clone the repository wherever that harness looks for skills:

git clone https://github.com/vit-panchuk/esf.git ~/src/esf
ln -s ~/src/esf/skills/engineering-strategy-framework ~/.claude/skills/

Point any harness that takes a skills directory at ~/src/esf/skills. For claude.ai and the API, zip skills/engineering-strategy-framework and upload it.

The tooling

A report written against this contract can be authored and published by hand. It can also be compiled: esf-dsl (npm) is the typed vocabulary those reports are written in, and the compiler that turns one source into the web page, the markdown, the deck, the JSON-LD and the rest — so the evidence meter is counted from the claims rather than written beside them. DSL.md inside the skill is the framework's view of that language; the package's own esf dict is the authority.

The two published reports below were compiled that way: one source each, out to a page, a deck and a Ukrainian edition.

The two repositories are independent. Follow the framework with no tooling at all, or use the language for a report that was never written against this contract.

What is in here

The plugin carries one skill, at skills/engineering-strategy-framework/:

plugin.json at the root is the Agent Plugins 1.0 manifest that Codex and Copilot CLI read; .claude-plugin/ holds the Claude Code manifests. All three discover the skill from skills/.

Worked examples

Two of the three are live, and go on changing as the engagements do — read them published at vit-panchuk.com/reports, or open the copies carried inside the skill.

• oss-audit-solidus.md — an existing-system audit run from outside, on a public open-source project, unanonymised because all of it is public. Twenty- eight revisions under adversarial review, a 43-entry decision log, and the policy layer compiled from a zero-mandate position. Its authoring source (.mdx) and its WIP graph (.graph.cypher) ship beside it. Published: Solidus — rev. 28, 132 claims, 99 minutes — with a 38-slide deck, a Ukrainian edition, and an essay drawn from it, The Moneypot Nobody Spends vs the Work Nobody Does.
• client-audit-saas-erp.md — the same mode from the inside: a private client, anonymised for publication, where the evidence base is a 27-month chat export because the project keeps no tracker. Carries a risk recorded as fired between revisions. Authoring source ships beside it. Published: SaaS ERP — rev. 4, 112 claims, 61 minutes — with a 21-slide deck, a Ukrainian edition, and an essay drawn from it, Founder, Start Adopting Your Product Now!.
• greenfield-telemetry-portal.md — a greenfield initiative, fully synthetic. The decisive reframe is that the "greenfield" is the producer side of an integration contract already running in production. Unpublished.

Version history

• 0.7.0 — the policy layer: a Policy register with operations as fields, the policies-first report opening in place of an executive summary, the Easy Wins fast lane, the Watchlist, and refinement fields on the top bets.

License

MIT. See LICENSE.