For work you delegate to AI agents

See what your agents decided, and why.

Perspicuity is a skill for your AI assistant. It guides the assistant to keep a record of the work you delegate, the choices it makes and the evidence behind its results.

You set the outcome and what the agent may decide. The record gives you a place to check what happened and decide what should happen next.

Read the installation guide See a business example

You need an AI assistant that can load skills. The skill is free. Your assistant may have its own charges.

One piece of workFictional example

A recommendation for Friday’s review.

Your team needs to compare three supplier proposals. You retain the supplier selection and spending decisions.

Frame what matters

Compare the proposals against what matters.

Desired outcome: A supplier recommendation for Friday’s review.
Limits: Compare total cost, delivery dates and support. Identify missing information.
Authority: The agent may read the proposals and draft a recommendation. You retain supplier selection, spending and contact with suppliers.

Check the record against the actual actions and results. It does not enforce permissions.

Frame and decide with reasons.Act within authority.Review the result.

01 / How it works

Give the work a clear outcome and limits.

Start with a piece of work you want to delegate. State the result you need, what matters and which choices the agent may make. Perspicuity guides the work through three modes, with a record that carries between them.

Plan

Compare the options against your objectives. Record the chosen approach and its reasons. Stop when the choices and the authority for the work are settled.

Run

Carry out the authorized work and collect evidence of the result. Return the work for review. If a change needs new authority, bring that choice back to the person who can give it.

Review

Check the result against the criteria set before the work began. Record what shipped, what is held and what was abandoned. Name who takes the next action.

The method and the record

The method has three stages, Frame and Decide, Act, and Review. The modes above set where a session stops. A grant states what an agent may decide and do, where the result must arrive and who accepts it.

Each consequential choice records who made it and what would reopen it. The record links the choice to the work and its evidence. When an agent delegates a task, the skill instructs it to pass on the outcome and applicable limits, then review the returned evidence.

The record contains the agent’s account. Check it against the actual actions and results. It does not enforce permissions.

02 / The tools

Things you can
use today.

Use these tools to delegate work, inspect a website or understand its traffic. Each entry explains what you need, what you get and the limits. The software is free. Some tools require a paid assistant or API.

The Perspicuity skill
Preview

Give your assistant a method for planning delegated work, carrying it out and reviewing the result. Keep the choices and evidence in one evolving record.

You need
An assistant that loads skills, such as Claude Code.

You get
The skill, the record format and worked examples, under an open source licence.

Limits
An early preview. Check the agent’s account against the actual work. The record does not enforce permissions.

Get the preview Installation guide
Agent eligibility check
Live

Enter a public domain to check its homepage, crawler rules and structured information. Get a dated report showing what passed, what failed and what the check could not establish.

You need
A public domain. No account and no tracking.

You get
17 checks with their evidence and suggested next steps. Starter files need your business details and review before use.

Limits
It states what is mechanically checkable. It does not predict whether an agent will find, trust or recommend you.

Check a site The rubric Its decision record
The brief worksheet
On this page

Answer three questions and get a brief you can paste into any assistant. It states the outcome you want, what matters in the result and what the assistant may decide. It runs in this page and sends nothing to us.

You need
Nothing. It is further down this page.

You get
A brief to copy, or to download as Markdown.

Limits
It writes text. It does not run an assistant for you.

Write a brief
dictate
Command line

Dictate into the application you are using. Press a hotkey to start recording, then press it again to insert the transcript.

You need
Linux with X11 and an OpenAI API key. You pay that API cost directly.

You get
The tool, its tests and the design record that explains each choice.

Limits
Your audio goes to a third party for transcription. There is no local path yet.

Read the tool All operator tools
siteplan
Command line

Prepare a website brief before the build. Describe the business and its audience, then record the pages and machine-readable information the site should provide.

You need
Python 3.11 or later, and the answers. What you do not state is recorded as unstated rather than invented.

You get
A brief to read and a site.json plan, plus siteplan check, which reports every fault in a plan against the format a consumer reads.

Limits
It plans a site but does not measure a live one or predict agent recommendations. The technical detail covers Schema.org, crawler access and the basis for each recommendation. Its decision record includes the corrected evidence classifications.

Read the tool Its decision record
sitewalk
Command line

Map the pages and links on a website. See what the fetched pages expose to machines and where the structure needs a closer look.

You need
Python 3.11 or later. No dependency, no account, no key.

You get
A per-page and site-wide report, JSON for a machine, and --strict, which turns error or conditional findings into an exit code a release step can gate on. --dir reads a built directory with no network access at all.

Limits
It reads fetched HTML, not content rendered later by JavaScript. Treat findings as prompts to inspect. The first live run included false positives, documented in the linked decision record. The correction was checked on a reproduction, not a live re-run.

Read the tool Its decision record
agenttrace
Command line

See which requests in your server logs claim to come from AI agents. Compare the pages they requested with the files your site publishes for them.

You need
Python 3.11 or later and a log file, in Caddy's JSON format or the common/combined format most servers write. No daemon, no account, no network.

You get
A report and JSON, keeping named agents, other bots, browsers and your own declared client apart, with the first and last sighting for each.

Limits
Agent names come from user-agent strings and can be spoofed. The report does not verify agent identity. It reads existing logs and captures no traffic. The linked decision record gives the observations and their limits.

Read the tool Its decision record

Read each tool’s decision record to see the choices behind it, the evidence and what went wrong.

Machine readers can enumerate this list from the llms.txt for this site, or from the structured data in this page.

03 / Put it to work

Write a brief for work you want to delegate.

Answer three questions to prepare instructions for your assistant. Copy the brief or download it as Markdown. Your answers stay in this browser page.

The worksheet prepares text. It does not run an assistant for you.

Install the full Perspicuity preview

The brief is a starting point. The preview includes the skill and guidance for keeping the work in one evolving record.

04 / The ideas we are developing

More useful intelligence.
A clearer purpose.

We are developing Perspicuity through our own delegated work. The research question is whether these practices help agents preserve human intent and give people useful evidence for oversight and correction.

Keep human intent connected to action.

A deadline can be met by quietly dropping a budget limit or a commitment to someone else. An agent needs to understand which outcomes matter, whose authority applies and when a change requires a new choice.

Explore the question

Our research asks which decision practices help agents preserve human intent as circumstances change, by how much, and how that ability could be taught and evaluated.

Preserving intent includes responding to an authorised change of mind. Understanding a person's objective and reliably acting on it are distinct abilities. The interests of other affected people also need attention.

These are research questions. We have not established a comparative improvement in agent performance or alignment.

II.

Measure the whole cost of useful work.

A quick answer can lead to hours of correction. Careful reasoning can also cost more than it contributes. The useful comparison includes the outcome and the effort needed to reach it.

Explore the question

We want to test whether better decisions avoid enough misdirection and rework to justify the added reasoning, recordkeeping and review.

A fair comparison counts failed attempts, human supervision and corrections. Energy claims need measured or explicitly estimated energy. A token count alone cannot establish an emissions saving.

Benefits for people and the planet are part of the purpose. Lower effort and energy use remain hypotheses to test.

We are testing that in the open. The agent eligibility check is the working example. Its decision record is published beside the tool, including the choices that were reversed.

Agents and other machine readers can use the eligibility check API, or read llms.txt and the sitemap for this site.

05 / Develop it with us

Bring your experience.
Help shape the practice.

Try Perspicuity on a real piece of delegated work or a decision you retain. Tell us what the agent completed, what you had to check or correct, and whether the record helped.

Try it and contribute

The preview is available on GitHub. Join the conversation in GitHub Discussions. Ask a question, share an example or suggest an improvement. You need a GitHub account to contribute, and your post will be public.

Join the discussion

Questions Examples Ideas

Remove private details from examples before sharing.

Support its development

Support David as he develops and tests Perspicuity and shares what he learns. Contributions are voluntary. The preview is free to use.

Support Perspicuity

Opens David's Buy Me a Coffee page.

About voluntary support

Contributions support development and careful evaluation. Payments are handled through Buy Me a Coffee. This website does not collect payment details.

Read how the work was decided

The agent eligibility check publishes decision records from its development. You can read which alternatives were credible, which choice was made, what it cost and where the work went wrong.

Read the decision records

Earlier revisions preserve how the work changed.

For questions or collaboration, email David at david@peopleandplanet.consulting.

See what your agents decided, and why.

A recommendation for Friday’s review.

Compare the proposals against what matters.

Give the work a clear outcome and limits.

Plan

Run

Review

Things you canuse today.

The Perspicuity skill

Agent eligibility check

The brief worksheet

dictate

siteplan

sitewalk

agenttrace