Real cases · Media · Freelance
Manoj, 36: a freelance journalist with three deadlines and a research tool that invents quotes
Long-form pieces, thin fees, and a fear — justified — that a fabricated fact under his byline ends the career.
Planned for release phase 2. The page is complete as a draft; interactive parts run in the browser and do not yet save to an account.
In brief
The situation. Manoj writes long-form features for Indian magazines and a couple of foreign outlets, from Bhopal. Fees have fallen; volume has to rise. He tried a research tool once and it gave him a beautiful quotation from a real academic who, when he checked, had never said it. He has not used one since, and his output has not risen. He suspects there is a version of this that does not require trusting anything.
Step 1 — Define the End State
"Three pieces a month instead of two, with every fact traced to a source I have seen with my own eyes, every quote from a recording, and the writing recognisably mine. An editor who has worked with me for ten years notices nothing except that I am filing more." Audiences: that editor, a fact-checker, and the academic who never said the thing.
Step 2 — Map the Flow
His list: pitch, research, interview, transcribe, structure, write, check. AI review adds: a source log per piece (a table of every claim and where it came from, built as he goes rather than reconstructed for the checker); and a "what have I not asked" pass before each interview. It suggests drafting from research notes. He declines the draft and accepts the structure: the tool may outline, never write a sentence that will be published.
Step 3 — Assign Tasks
The angle, the interviewees, the questions that matter: his. Background reading across forty documents: AI summarises with page references; he reads the referenced pages. Transcribing interviews: AI, verified against audio for any quoted line. Building the source log: AI extracts claims from his draft and asks for a source for each — a reversal that catches his own unsupported lines. Every published sentence: his.
Step 4 — Select the Tool
A document tool with large context for background (with the source documents loaded, never open-web recall); a transcription tool; a conversational tool for the claims-extraction pass. A rule: the tool never sees the open web on his behalf for facts. If it did not come from a document he has, it is not a fact.
Step 5 — Prepare and Direct
Criteria: any quote in the piece exists in a recording I hold; any figure has a page number; the tool's summaries are pointers, not sources; the claims log has zero blank rows before filing; I can hand the log to a checker without embarrassment.
Steps 6 and 7 — Execute and Evaluate
The summaries with page references are the unlock — reading forty pages instead of four hundred, and reading the right forty. The claims-extraction pass on his own draft finds three lines he had written from memory that he cannot source. He cuts one and sources two. The transcripts have one misheard number; the verify-against-audio criterion catches it.
Step 8 — Integrate and Stress-Test
First piece under the new method, assembled with its log. Small fixes: a name spelling. Structural: the piece is well-sourced and reads, his editor says, "like a report" — the source-log discipline had leaked into the prose. A Step 5 problem: his criteria for the writing had nothing about voice. He adds one — "a paragraph a stranger would read aloud" — and rewrites two sections. The editor notices nothing except that he filed a week early.
What changed
Three pieces a month, sometimes four. The claims log goes to checkers, who have started asking other writers for one. He has not trusted a tool with a fact since the fake quote, and does not need to; he trusts it to point, and he looks.
Summaries are pointers, not sources — read the page. Run the tool backwards on your own draft: extract every claim, demand a source. If the tool cannot see a document you hold, it cannot invent a fact from it.
A composite case, not a client engagement; the person is invented, the situation is not.