E Evidence Press

A programme of Evidence Press

Productivity Protocols

Methods, not papers: open, tested workflows for using AI agents — published, like everything here, with the evidence attached.

What this is

Evidence Press publishes what research has discovered, with the evidence attached. Productivity Protocols publishes something adjacent: how to reliably use AI agents to do useful work — as methods, not papers. Each protocol is an open, downloadable, tested workflow, released with its assurance and its honestly-measured benefit attached.

Browse the protocols →

A protocol is a work contract, not a prompt

A prompt is a suggestion. A protocol states the whole contract: what task it addresses, what it may read or change, where a person must approve, how it is checked, what counts as failure, and what evidence — if any — shows that it helps. It ships as an open Agent Skill you can inspect, download, and run without installation.

Two independent measures, never merged

Every protocol carries two statuses. They answer different questions, and they are kept strictly apart:

  • Protocol assurance — is it well built and safe? This runs from a structural check of the packaging, through its own tests, up to reproduction of the same result across different models.
  • Productivity evidence — does it actually help, and how do we know? This is measured, never assumed.

A protocol can be flawlessly engineered and still make you slower. Collapsing the two into a single "quality" badge would hide exactly that, so the library refuses to. A workflow may sit at the top of the assurance ladder while its benefit is still unproven — and the page will say so.

The two status ladders — protocol assurance and productivity evidence — shown side by side and never merged
Two independent measures, kept apart. A protocol can climb the whole assurance ladder on the left and still sit at NO_CLEAR_GAIN on the right — and so far, ours do.

Benefit is measured, not claimed

The library keeps negative results. A workflow that was evaluated and did not help is useful knowledge, particularly when it looked promising. So far, every protocol that has been evaluated live carries no clear gain: on its task set and its model, the method added no worthwhile benefit — and in the ceremony-heavy cases, it cost quality and time. Each such finding is stated plainly on the protocol's page, with the run behind it, rather than quietly dropped.

That honesty is the point. A benefit is claimed only up to the evidence that supports it; below that, the statement is simply "benefit not measured."

How to use them

Every protocol offers three ways in: a plain-language copy-and-run edition that needs no installation, a downloadable pack with a verifiable SHA-256 hash, and an optional connected edition for tools and data, where any outward action waits for a person to approve it. Each also exposes a machine-readable record, so an agent can read the contract as easily as a person can.

Browse the protocols

The full list is filterable by task, risk, required tools, assurance, and measured benefit. Each entry links to its contract, its tests, its evaluation, and its download.

See the protocols →

Prose here is public domain and the code is openly licensed, and every status shown is reproducible from a clean checkout — the same standard the rest of Evidence Press holds itself to.