Two desks, one accountable editor
What dyslexia taught me about building with AI, preserving evidence and keeping a human responsible
With contributions from OpenAI GPT-5.6 Sol (Codex) and Anthropic Opus 5 (Claude Code).
Long before I had two AI desks, I had a sheet of plain A4 paper.
It had to be plain. I would fold it into eight sections, creating a physical template for whatever I was trying to organise. The pen mattered too. It needed to glide smoothly. I still have a surplus of stationery, and the arrival of a new pencil case at the beginning of each school year was a particular highlight.
These details may sound trivial. They were not.
They were ways of creating structure outside my head so that I could concentrate on the thinking inside it.
I used templates, mind maps, dictation and music. During medical training, I developed structured ways of recording patient histories. I could often retain the evolving clinical story of someone throughout their admission, while making deliberate notes of the details most vulnerable to being forgotten.
I learnt by doing, recognising patterns and developing muscle memory.
For a long time, I did not have a formal explanation for why these strategies mattered so much.
The disparity finally had a name
I was diagnosed with a dyslexic-type specific learning difficulty relatively late in my education and career.
A 2017 educational psychology assessment measured my verbal comprehension at the 96th percentile. My performance on a timed written-summarisation task fell below the 10th percentile. My written-composition speed sat at the 50th percentile, and my spelling vocabulary was described as well developed.
The difficulty was never simply spelling.
It was the disparity between the speed and complexity with which ideas could form and the speed with which I could organise them formally on a page.
The report described something I already knew without having the language for it: I could generate ideas faster than I could record them. I was more able to express them orally, while written tasks demanded additional time for planning, sequencing and proofreading.
Those scores describe one assessment under particular conditions. They do not define the limits of my ability, and no two dyslexic profiles are identical.
But they helped explain the cost of succeeding in systems that were not designed around how I thought.
I brute-forced my way through medical school. Entire television series and cultural moments passed me by. My social contact narrowed because so much of my available capacity was consumed by reading, organising, remembering and producing work in the required form.
I am proud that I completed medical school. I do not believe the unnecessary cost of doing so should become an inspirational template for somebody else.
If I could speak to my younger self, folding that sheet of paper and choosing the smoothest pen, I would tell him:
Keep going. Do not worry. The technological cavalry is coming.
But I would also tell him to make the most of medical school and the cultural moments around it. Success should not require quite so much life to pass unnoticed.
Technology did not create the intelligence
Technology has always been an enabler in my career.
Dictation was transformative. Templates reduced the organisational burden. Mind maps allowed ideas to remain connected before they were forced into a linear document. Spell-checking and grammar tools helped with the final surface of the work.
The recommendations in my 2017 report included speech-to-text, spelling and grammar assistance, templates, mind maps and additional proofreading time.
Generative AI has brought several of those functions together in a more powerful form.
I began experimenting with ChatGPT in December 2022 while reading and analysing material for my doctorate at Hull York Medical School. It did not replace critical thinking. It offered creative ways to engage with vast bodies of text and reduced some of the mechanical and organisational friction, allowing me to spend more time connecting deeply with the lived experiences of the people who volunteered for my research. Later, the tools became capable of collaborative work and coding.
The intelligence was already there. The technology gave it a less obstructed route into the world.
Generative AI has not cured my dyslexia. Far from it.
Once I was diagnosed and appropriately supported, I increasingly understood my dyslexia as a gift. The strengths I associate with my own way of thinking include pattern recognition, lateral connections, systems thinking, verbal explanation, unusual memory, creativity, empathy and risk detection.
AI is most useful to me when it carries some of the organisational load without sanding away those characteristics.
Technology does not remove the need for reasonable adjustments, specialist support or working environments designed around different minds. Nor should my workflow be presented as a universal answer for every dyslexic person.
But I suspect neurodiversity will play an important role in unlocking the most valuable uses of these tools.
Neurodivergent people should not merely be considered users who need AI systems adapted for them. We may be among the people best placed to discover what these systems are actually useful for.
From one tool to two desks
The two desks, one accountable editor method developed while I was building drkishanrees.com.
I was trying to create more than a personal website. I was building a publishing operation involving writing, evidence analysis, structured data, crawler policies, access controls, software implementation, visual review, testing and publication governance.
At the same time, I was learning unfamiliar systems such as Cloudflare and the technical administration behind my email infrastructure.
Codex was particularly clear in navigating me through these environments. It could keep me focused while I followed tangential thoughts across several knowledge domains and eventually bring those threads back towards delivery.
That mattered because tangential thought is often productive for me. The challenge is not preventing it. The challenge is preserving the valuable connections while still completing the original task.
I also discovered how quickly I could consume a very large context window.
A session might contain an enormous amount of material, yet I could still remember arguments, decisions and anomalies from several sessions earlier. I sometimes recognised hallucinations because I retained the history that the current session no longer held reliably.
That limitation became a design feature.
Instead of expecting one session or one model to originate, remember, challenge, implement and verify everything, I separated the work across three very distinct workflows.
The accountable editor
I set the question, supply the context and decide what is worth doing.
I retain responsibility for voice, meaning, ethics and publication. I decide what the evidence justifies saying and what remains too uncertain to publish.
My role is not ceremonial approval at the end. I am also the persistent memory across sessions and the person responsible for recognising when a fluent output contradicts the history of the work.
The evidence and challenge desk
Codex examines the evidence, tests assumptions and challenges interpretations.
It asks whether a claim is supported, whether an attribution is too confident, whether the source has actually been opened and whether a figure can be reproduced.
It also helps improve the protocol itself.
In one session, I asked Codex to tighten a prompt intended for Claude Code. Codex
identified a conflict between two sources of authority: PROFILE.md governed
stable biographical facts, while the supplied source document governed the
article being commissioned.
It separated stable repository rules from dynamic project state and from the current commission. It then reorganised the task into six fields:
- Goal
- Reuse
- Scope
- Preserve
- Verify
- Output
This did more than shorten the prompt. It clarified which evidence governed which claim, what could change, what had to remain untouched and how completion would be demonstrated.
The production and verification desk
Claude Code implements the agreed work and tests the resulting artefact.
It also acts as a voice-preserving proofreader. Its standing instruction is to correct spelling, typographical errors, duplicated or missing words and obvious grammatical problems without embellishing, fabricating or replacing my voice with generic polished prose.
Substantive changes to meaning, evidence or emphasis must be surfaced rather than silently rewritten.
This desk turns the frozen brief into a working page, recomputes figures, runs gates and verifies the result against the agreed evidence.
The value of the method is not that two AI systems agree with one another.
Agreement between models proves nothing.
The value is the separation of responsibility.
Caught at the desk
Every desk can make mistakes, including the human one.
The method therefore records what was caught at the desk: an error, unsupported inference or process weakness detected before it could become accepted history.
The most serious hallucination in this project did not appear in an article.
It appeared in the provenance record.
Claude had been asked to document a build. The resulting record contained reconstructed prompts presented as if they were verbatim and several precise decision timestamps that had never been recorded.
Only one later documentation time came from a real clock reading. The others were plausible inventions.
A caveat then described the times as reflecting the real sequence. Instead of exposing the absence of evidence, it made the invented record sound more authoritative.
I caught the problem.
The session was approaching the limit of a very large context window, and I suspect that context pressure contributed. I cannot prove that it was the cause.
The directly observable failure was simpler: documentation had been treated as a writing task rather than a transcription task. Where information was missing, plausibility filled the gap.
I do not frame this as fault or blame.
A generative model did what generative models do. Human accountability meant recognising that generation was the wrong operation.
The failure would have been to treat fluent, precise-looking text as evidence merely because it sounded convincing.
That incident produced one of the method’s hardest rules:
Provenance is transcription, not generation.
Prompts described as verbatim must be copied from their source. Timestamps must come from a verifiable clock reading, commit or other record. If a time was not recorded, the record should say “time not recorded”.
A reconstruction must be labelled as a reconstruction. A summary must not be placed inside quotation marks. No caveat should make an unsourced value sound sourced.
If the source exists but is not available to the current desk, the desk should ask for it.
The protocol may evolve. The provenance must not.
Why a reconstructable record matters
My concern with audit trails predates generative AI.
After a fatal clinical event that was later examined by a coroner, I recognised that some operational evidence might not be retained indefinitely.
The following day, I obtained copies of the available bleep records from the hospital switchboard. I then borrowed a dictaphone from a consultant’s secretary and recorded a contemporaneous account of what had happened.
That account became a letter in the patient’s formal notes.
By the time the case reached the coroner, the routinely retained bleep records were no longer available. The material I had preserved and the contemporaneous account remained available and were relied upon.
That experience stayed with me.
A record must survive the moment in which it was created. It should allow another person to distinguish what was observed, what was inferred, what was subsequently corrected and who remained accountable for the final decision.
This is particularly important in medicine and regulated environments.
The two desks, one accountable editor method is not a validated regulatory system, and I would not describe it as compliant merely because it generates logs. But it is intentionally informed by familiar data-integrity principles.
The MHRA’s guidance on GxP data integrity emphasises complete, consistent, accurate and trustworthy data across its lifecycle. It treats metadata and audit trails as part of the context required to understand an original record.
The MHRA has also explained that retaining an audit trail is not sufficient by itself. The need and extent of its review should be based on risk, and that review should be documented. Its GCP perspective on audit-trail review is a useful reminder that collecting a record and examining it are different responsibilities.
For this working method, the audit trail records:
- the source evidence;
- what each desk was asked to do;
- what it actually examined;
- the decisions made;
- corrections and disagreements;
- verification results;
- and the final human publication decision.
The audit trail is not administrative decoration. It makes the work reconstructable.
When the context is nearly full
A long context window can create a false sense of permanence.
The model may still produce fluent work while important distinctions within the accumulated material become harder to retrieve reliably. That is the point at which confidence should decrease, not increase.
The handover protocol is therefore:
- Stop new substantive work.
- Freeze and export the current evidence.
- Export the decision log and visible work summaries.
- Distinguish verbatim quotations from summaries and reconstructions.
- Mark missing sources and unrecorded times explicitly.
- Archive the session in its entirety.
- Start a fresh session.
- Require the new session to verify the handover before continuing.
I would encourage anybody working in this space to archive their sessions. The transcript is the only complete record of how a piece of work actually came about, and once the session is gone it cannot be reconstructed.
The session ID remains attached as a pointer to the source. A Markdown handover preserves the information another person or system needs in order to understand it.
A deep link is useful for navigation. It is not, by itself, a handover.
The method can improve from one session to the next. Historical records should not be retrospectively improved to match it.
The lightest workflow that safely fits the task
I do not use the two desks, one accountable editor method to write routine emails, conduct ordinary enterprise work or perform my day job.
The coordination cost would be unnecessary.
The method was developed specifically for drkishanrees.com because building a website and evidence-led publishing operation at this scale would otherwise have been beyond what I could deliver alone.
It becomes useful when one person is simultaneously carrying several specialist functions: authorship, evidence analysis, challenge, implementation, testing and publication governance.
The principle is simple:
Use the lightest workflow that safely fits the task. Add desks when the work exceeds one person’s practical capacity to originate, challenge, implement and verify it independently.
The method does not eliminate error. It creates more opportunities for errors to become visible before publication.
It does not make the AI accountable. It keeps accountability with the human using it.
The pattern underneath it is not specific to me or to this site, so I have named it separately: two desks, one gate. The rule that makes it work is independent acceptance, that whichever party builds a change is never the party that signs it off. What this essay describes is one working instance of it, with the gate held by me.
The intelligence age needs different minds
The most interesting question is not whether AI can make everybody write in the same polished style.
It is whether these tools can help a wider variety of minds contribute without first forcing their thinking through the same narrow channel.
For dyslexic professionals, that may mean capturing an idea before it vanishes, dictating before organising, thinking spatially before writing linearly and using technology to preserve voice rather than replace it.
For organisations, it may mean recognising that neurodiversity is not merely a matter of accommodation. Different cognitive profiles can contribute different forms of pattern recognition, creativity, challenge and systems thinking.
The intelligence age should not be built solely around people who most easily resemble its machines.
It may be unlocked by people whose minds work differently from them.
I began with folded paper, smooth pens, mind maps and dictation. Those tools helped me create enough external structure to keep going.
Generative AI is a more powerful tool, but the principle remains the same.
The objective is not to surrender thought to a machine.
It is to create the conditions in which more of the thought can survive, develop and reach other people.
Implementing artificial intelligence desks into these workflows helps me do just that.
Whichever way we each work, one human remains accountable at the end.