In developmentWorking titleNot for sale
AI workflows that check their own work.
The AI Workflow Toolkit is three complete, step-by-step workflows for work you have to stand behind: research, planning and professional writing. Each one builds in a fresh-chat challenge step and a point where you check the sources yourself, so more mistakes surface before anyone else sees them. Not another prompt pack.
No price set, no checkout, no release date. The free sample needs no sign-up.
Why it exists
AI is fluent, fast, and wrong more often than it sounds. The fix isn’t a better prompt. It’s a better procedure.
When Columbia’s Tow Center asked eight AI search tools to identify the source of news excerpts, they answered incorrectly in more than 60% of 1,600 queries, usually with no sign of doubt (Tow Center study via Nieman Lab, March 2025 (opens in a new tab)). Prompt collections sell volume. They rarely cover the part that matters: knowing which parts of an answer to trust. Every workflow here makes the AI show its evidence, hands the checking to a separate chat, and tells you exactly which few sources to open yourself.
Three workflows, done properly
Three complete workflows rather than thirty shallow ones. Each is 5 or 6 linked steps, with a point where your own judgement comes in.
Workflow 01
Research and fact-check
Go from a vague question to a one-page brief where every claim has a source, a quote and a confidence rating.
- Solves
- Confident answers that turn out to be wrong, out of date or unsourced.
- For
- Anyone who has to stand behind an answer: a recommendation, a policy check, a client question.
- Time
- 30–60 min
Steps
- R1Frame the question
- R2Gather, or research from a source pack
- R3Claim ledger
- R4Challenge passnew chat
- ✓You open the key sources (done by you, without AI)
- R5Write the brief
You get: Scope card · Claim ledger (CSV) · Fresh-chat challenge table · One-page brief with confidence label
Tested on Claude. Run end to end on a real licensing question using Meta's Llama 3.1 licence. A fresh-chat challenge pass, on the same assistant, flagged claims the first pass had wrongly approved; we then confirmed those against the licence text by hand. Research with live web access is not yet tested.
Workflow 02
Idea to project
Turn a page of messy notes into a one-page charter, an honest pre-mortem and a first week that fits your real hours.
- Solves
- Ideas that stay ideas, and plans that ignore how much time you actually have.
- For
- Side projects, small-business launches, internal initiatives, freelance offers.
- Time
- 45–60 min, then 10 min a week
Steps
- P1Unpack the idea
- P2Project charter
- P3Pre-mortem
- P4Milestones and first week
- P5Weekly review
You get: Project one-pager · Ranked pre-mortem with a cheap test for each risk · First-week task list · Decision log
Tested on Claude. Run end to end with a fictional user and a deliberately messy first week. The weekly review caught the drift the pre-mortem had predicted.
Workflow 03
Content draft and review
Write something that sounds like you, uses only facts you supplied, and has been through a fresh-chat editor's review, plus your own fact check, before you send it.
- Solves
- Generic AI writing, invented examples, and drafts that review themselves kindly.
- For
- Anyone whose name goes on the writing: consultants, owners, freelancers, team leads.
- Time
- 40–60 min per piece
Steps
- C1Style card (once)
- C2Brief and outline
- C3Draft
- C4Editor reviewnew chat
- C5Revise
You get: Reusable style card · Approved outline · Draft plus 'went beyond your brief' list · Editor's review and final version
Tested on Claude. Run end to end on a blog post about Making Tax Digital, using facts from GOV.UK. The editor step caught an overstated claim in the writer's own wording. One weakness was found and fixed, then re-tested.
See what a fresh-chat challenge catches
In our test, an AI researched whether an agency could build a client tool on Meta’s Llama 3.1 model. The same chat then marked its own claims. A fresh chat on the same assistant, given the same source documents, disagreed with three of these four. We then checked those points against the licence text ourselves. A fresh chat is a second look, not independent verification: opening the primary sources by hand is still essential.
Real rows from the recorded test, lightly shortened.
Claim ledger · from our recorded test of Workflow 1
The licence grants broad rights to use, copy, distribute and modify the model
The licence calls itself a “limited license”, conditional on other sections.
NEEDS QUALIFIER
It is ambiguous whether the agency can hand its licence over to the client
Not ambiguous: the licence is “non-transferable”. The client is bound by its own use.
CONTRADICTED
If the licence ends, the model must then be deleted
The duty covers the “Llama Materials”, whose definition doesn’t include fine-tuned versions.
NEEDS QUALIFIER
The licence is governed by California law
§7: “governed and construed under the laws of the State of California”.
CONFIRMED
The challenge ran in a new chat on the same assistant, with no memory of the research. That is a second look, not independent verification. Amber marks a verdict that changed.
What’s in the download
Plain files that work with any AI assistant and on any device. No extension, no account, no subscription.
Folder
- 00-START-HEREGetting started, with a 20-minute first run · HTML, PDF, Markdown
- 01-research-and-fact-check/Playbook, 6 prompts, 3 templates, printable checklist, full test transcript
- 02-idea-to-project/Playbook, 5 prompts, 3 templates, printable checklist, full test transcript
- 03-content-draft-and-review/Playbook, 5 prompts, 2 templates, printable checklist, full test transcript
- limits-and-failure-modesEvery failure we saw in testing, how you’d notice it, and what to do
- testing-notesWhat we tested, on which model, and what we haven’t
Turn the findings above (or the text below) into a claim ledger. Do not add new research and do not fix anything yet. 1. List every separate factual claim. Split compound sentences: "Sales rose 12% in 2024, the fastest in a decade" is two claims. 2. For each claim, fill in this table: | # | Claim (exact wording) | Type | Source given | Supporting quote | Risk | Status | … Be strict. If a quote says "around 60%" and the claim says "60%", mark it PARTLY.
Playbooks
Why each step exists, what to check before moving on, a worked example, and a troubleshooting table.
Prompts
16 plain-text prompts with fill-in fields. Copy from the files or use the copy buttons in the playbooks.
Templates and checklists
A claim ledger (CSV), research brief, project one-pager, decision log, style card and three checklists.
Test transcripts
Every input and output from our test runs, unedited, including the failures.
How we tested it
Every workflow was run end to end on 10 November 2026, and every transcript ships with the toolkit, failures included.
We used Claude through Anthropic’s own command-line tool, with realistic scenarios and real sources: Meta’s licence text and GOV.UK guidance. Where a step fell short, we changed the prompt and ran it again. One change worked, one only half worked, and the testing notes say which.
Not tested yet
- Any assistant other than Claude (ChatGPT, Gemini, Copilot and others)
- Independent verification: every challenge pass ran on the same Claude model family that did the work
- Workflow 1's research step with working web access
- Real users on their own projects
- Repeatability beyond the Claim Check and the claim-ledger step, which we re-ran on the same inputs
Free · no sign-up
The Claim Check
One prompt and a five-minute routine for checking any AI answer, article or draft before you rely on it. It's two steps of Workflow 1, combined so they work on their own.
We gave Claude a 117-word paragraph with five planted errors and ran the Claim Check five times: once with web search, then four times without. It flagged all five errors every time. Our environment blocked web pages, so every run worked from the model's memory with no sources opened, and all runs used the same paragraph and the same model. It's a demonstration that the prompt behaves consistently, not a measured accuracy rate: you still need to open the sources yourself.
It works on its own. It isn’t a cut-down teaser: it’s the same checking method the full toolkit uses.
Who it’s for
People who already use AI most weeks and need results they can put their name to.
Consultants, freelancers, small-business owners, specialists and team leads who send AI-assisted work to clients, managers or the public.
Probably not for you if…
You’re brand new to AI (start with our free prompting guide), you want hundreds of one-line prompts, or you need developer tooling.
Free, and what the toolkit adds
Free on this site
- Guides that teach the habits: splitting answers into claims, going to the original source, keeping a human check
- The Claim Check: one prompt and a five-minute routine for checking any text
- Worksheets for briefs, choosing a model and building a workflow
What the toolkit adds
- Three complete workflows with fifteen prompts that hand work from step to step
- Templates and printable checklists that hold each workflow together
- Troubleshooting and a limits page drawn from real test runs
- Unedited transcripts of every test, so you can see how the prompts behave before you trust them
If the free material is enough for what you do, use it. The toolkit is for people who want the complete, tested method in one place.
Details
- Format
- Digital download: HTML, PDF, Markdown and plain-text prompts (proposed)
- Works with
- Any capable AI chat assistant. Tested on Claude only so far.
- Price
- Not set
- Release date
- Not set
- How to buy
- Not available yet. There is no checkout on this site, and nothing is for sale.