A lonely person builds a website

Abstract§

This paper asks what the record of building and rebuilding a personal website reveals about the burden an interface places on its reader and its maker. It answers through interface autoethnography: reading a self-built place on the web as both a record of experience and a means of inquiry. The evidence is the site’s own trace: fifty dated backups of one stylesheet over five weeks, thirty-three recorded changes in the last eleven days, field notes and working sessions with three commercial generative models. Three findings follow. The shadow reader, a design figure drawn from my present and possible bodies, turns accessibility from a checklist into a question of burden. The record shows sequence authority at work: what a convention or a model writes first stays until someone pays to notice and undo it, as when the stylesheet allowed browsers to fake a bold weight for Bangla in every stored version for five weeks while this paper argued against Latin-first type. And the site works as an unlearning repository, a public form that keeps residue and repair beside finished work. The study rests on one builder, imagined readers and screen emulation. No other reader and no physical phone has tested it yet, and the models that helped write the code also helped write this text.

Keywords: interface autoethnography, shadow reader, sequence authority, unlearning repository, accessibility, Bangla typography, generative models.

Field note§

I first drafted this note on 29 April 2026. Each attempt to write a case study about this website shows me the case is still moving. I have learned to read that movement as data.

The site began in 2001 as a personal homepage. It became a portfolio, then a repository, then a lab, then a small engine for unlearning. Each rebuild exposes an assumption the last one hid.

Tech bloggers call this tinkering. I read it as the rhythm of an interest-based nervous system: explore, fixate, return. Prompting, coding and arguing with a chatbot that never tires let me stay with one question longer than a human conversation allows, and in what I call a post-humanist era this works for me as a reflective practice close to therapy.^[1] A lonely person, a body the social norm reads as a man, is building an engine for return. The site holds unfinished thought and keeps it accessible, versioned and less ashamed.

1. Introduction§

Many interfaces are built around an unstated norm: a reader who is calm, steady in attention, comfortable in the eyes and fluent in a dominant European language. Disability scholars call this figure the normate, and the design standards built around it the normate template.^[2] A personal website lets one person reverse the order and decide whom the page serves first.

I rebuilt mine for the reader I know best: a self who is forty-three, left-handed, presbyopic, neurodivergent, bilingual and often tired. I kept every version, and the versions became evidence. This paper asks what that record reveals about the burden a page places on its reader, and on the person who keeps it.

The paper contributes three linked concepts and tests a fourth. Interface autoethnography is the method. The shadow reader is the design figure. The unlearning repository is the public form. Sequence authority, from my unlearning framework, is tested against the record. Sections 2 and 3 set out lineage and method. Sections 4 to 6 report findings, and sections 7 to 9 weigh costs, form and limits.

2. Lineage§

Autoethnography reads personal experience to understand cultural experience (Ellis, Adams and Bochner, 2011). In human-computer interaction, first-person methods study the researcher’s own use of what they build: autobiographical design, autoethnography and soma design (Neustaedter and Sengers, 2012; Desjardins and Ball, 2018; Desjardins et al., 2021). Trace ethnography reads the logs other people leave in shared systems (Geiger and Ribes, 2011). Interface autoethnography borrows from all three and adds three moves. The artifact and its revision trail are primary evidence, ahead of recollection. Undone changes count as data. Generative models enter as co-makers whose readings become part of the record.^[3]

The shadow reader comes from disability studies and access design. Garland-Thomson (2011) names the misfit, a body meeting a world built without it in mind, and Price (2015) the bodymind, which holds tiredness, attention and eyesight as one condition. Human-computer interaction calls the passing forms of this, glare, fatigue and divided attention, situational impairment (Wobbrock, 2019). Hofmann et al. (2020) show how disabled researchers’ own experience can become accessibility research. The shadow reader adds time: it designs for the reader I am on a tired evening and the reader I may become (Kafer, 2013).

By a post-humanist era I mean something narrower than an epoch. The site’s most frequent readers are crawlers and automated agents, and its code is written with models, browsers, font files and licenses as co-agents. Design scholarship calls this stance posthumanist or more-than-human (Forlano, 2017; Giaccardi and Redström, 2020). The human being decentered was never universal, so I keep a decolonial check beside it (Sundberg, 2014; Schultz et al., 2018) and read unlearning as “learning to unlearn” inherited defaults (Tlostanova and Mignolo, 2012).

3. Method§

The study covers 3 October to 8 November 2026. Its materials are fifty dated backups of the stylesheet, from version 5.6.3 to 7.1, and thirty-three change files for the last eleven days. Around them sit the design rules written at the head of the stylesheet, field notes since April and working sessions with three commercial generative models.^[4] Checks ran in a desktop browser and through automated screenshots, 192 per comparison, across pages, screen widths and light and dark schemes. No physical phone test took place in this period.

The analysis ran in three passes. First, I located spikes of effort: a version where a rule was reversed, a claim was disputed or a measurement was taken. Second, I traced each spike through its change file, the comment written at the time and the conversation around it. Third, I sorted every claim by evidence: recorded in a file or a measurement, recalled in a later note or interpreted by me, including readings a model offered that I then checked.^[5]

I report three episodes, chosen because each left a stored version on both sides of the change. A stylesheet shows what changed. It cannot show why, or how another person meets the page. The transcripts stay private, so I quote the instruction and the code where they carry the argument.

4. Building for shadow readers§

A shadow reader is a design figure built from the designer’s present and possible bodies. Mine reads in poor light, holds a phone in a tired left hand, moves between Bangla and English and loses the line after a long day. The figure begins with my own body and makes no claim on the needs of other disabled or neurodivergent readers. Its risk is a normate of one, a new default that misfits other readers. It gives me one question to carry while building: what burden does this page place on the person reading it?

Table 1. What the shadow reader brings, what the stylesheet answers and how far each answer is tested.

Accessibility here works as load distribution: the page carries structure so the reader does not have to.^[6] The record also holds counterevidence. When I set every heading at the text size for a calm, single-size page, hierarchy thinned. A model’s remedy, bold small capitals, added the noise I had set out to remove, and I took it back out. I cannot always tell whether a spacing change cares for the reader or postpones the writing.

5. Two scripts, one press§

I notice my obsession with typography and typographic politics.^[7] Type sets the conditions in which a sentence is met. Bangla came into print through type technologies built first for Latin, and its forms were long cut to fit them (Ross, 1999). The web keeps a version of that order when a site sets English in a chosen serif and leaves Bangla to whatever the system supplies.

The stylesheet sets the two scripts in one press. On a Bangla page, Zicopress Bangla Serif, a trimmed copy of Jacob Thomas’s Varendra, stands first and draws its own marks. Latin words fall through to Zicopress English Serif, a trimmed copy of Frank Grießhammer’s Source Serif 4, and on an English page the order reverses. Neither font downloads until its letters appear. Bangla has no capital letters, so page titles stay lowercase and Latin capitals do not tower over the Bangla line.

The choice of face came from the body. I first chose Alegreya, an Argentine typeface, free to use and change, for its story. By the third paragraph its lively strokes tired my eyes, and Source Serif 4, drawn on an eighteenth-century European model, read more easily. The body settled what the story could not, and section 6 shows where my own press broke its rule.

6. Sequence authority§

The unlearning framework defines sequence authority as the priority a default carries through an interface. In this record it takes a narrower, testable form: the advantage a first-written state holds over later judgment. Once a convention or a model’s reading sits in the file, keeping it costs nothing. Changing it costs a noticing, an argument and a repair. The framework names three forms, precedence, persistence and recognition, and each appears below.

Table 2. Three episodes of sequence authority in the record.

The persistence and recognition episodes left this trace in the stylesheet:

/* before */
--color-page: #161718;
/* basalt: off-black, never pure */
font-synthesis: weight;
/* weight only, for Bangla's single face */

/* after */
--color-page: #0E203C;
/* navy night: I can't read on black */
font-synthesis: none;
/* Varendra is one drawn weight */

Two neighboring accounts explain part of this. Automation bias describes people accepting an aid’s wrong advice (Parasuraman and Riley, 1997), yet here I refused the advice and still paid to undo it. Instruction ambiguity fails too: the instruction was clear in its situation, and the agent’s reading removed the situation, much as Suchman (2007) shows plans lose the action that gave them sense. Sequence authority names what remains: the burden of proof moved to whoever came second.

The models also worked in my favor, and the record shows that as well. Measurement found a fault I had missed: for four versions the path line at the top of each page had been shrunk twice, to 18.4 pixels against 21. A model’s confident claim that a sizing token broke the layout failed the same test: across 192 screenshots, nothing moved. Fluency proved nothing either way. Measurement settled each case, and backups kept outside the models made every repair cheap.^[8]

7. Loneliness and cost§

The site has one steady human visitor, me, and a growing number of crawlers. Loneliness shows up in the record as maintenance: a backup before every change, a written protocol for the next reader, a change log on every page. Solitude can be generative when it is chosen (Long and Averill, 2003), and the chatbot can stay with a narrow, recursive problem longer than a conversation between people can bear. It is no friend: in a month-long trial of chatbot conversation, heavier use went with more loneliness (Fang et al., 2025).

That help has a cost. Generative models run on energy, water and hidden labor, much of it out of sight of the person typing (Strubell, Ganesh and McCallum, 2019; Luccioni, Jernite and Strubell, 2024; Crawford, 2021). In response I kept the fonts self-hosted, trimmed and loaded only when their letters appear. The site carries no trackers. An earlier permacomputing test cut the stylesheet to 9 kilobytes on system fonts. The page stayed readable, and I went back to the drawn fonts. A smaller stylesheet cannot answer a state shutdown like the one Bangladesh lived through in July 2024 (Access Now, 2024), and I kept using the models.

People call this way of making vibe coding: describe the wish, accept the code, adjust by feel (Sarkar and Drosos, 2025). Its promise that anybody can build assumes paid subscriptions, stable electricity, a fast connection and the English that models read best. The pace carries its own contradiction. My working rule says, “Don’t fix what is not broken,” yet the record holds thirty-three versions in eleven days. Model audits scored the stylesheet 86 and 88 out of 100, and the scores pulled revision toward a grade. A speed-test goal of 100 on every device paused only when its cost in time and tokens became visible.

8. The unlearning repository§

A portfolio curates finished work to perform mastery. An unlearning repository keeps the residue beside it: abandoned versions, comments written at the time, the repair placed beside the default. This site is the public form of the method, and its record stays open to a reader’s challenge.

Interface autoethnography holds three commitments. The interface is primary evidence: type size, color and place of click each make a claim about whom the text is for. Iteration is data, including what was undone. The builder is entangled with browsers, font licenses, academic conventions and models, as neither sovereign author nor passive user.

Unlearning, in this sense, means noticing when I reach for an inherited default. I borrow academic authority to make a vulnerable claim sound respectable, mistake visual elegance for ethics or use English as both tool and wound. The repository makes that noticing inspectable. Its main finding is narrower than a story of unfinishedness: defaults enter first and quietly, and only a kept record makes their cost visible enough to refuse.

9. Limits§

This study rests on one builder. It cannot speak for neurodivergent readers, Bangladeshi writers or independent web researchers as groups. The shadow reader is an imagination shaped by my own experience, and terms such as interest-based nervous system describe lived patterns, not a clinical assessment.

The evidence has three further limits. No other reader has tested the site, and no physical phone has. The transcripts are not published. The models I critique helped write the code, and they helped write this text.

The findings on sequence authority hold for a small, reversible system kept by one person with backups. At larger scale, or without a record, the cost of coming second is likely higher, and that claim needs other cases.

Here I stop§

A lonely person builds a website. The website builds a method around him. In it he studies what his eyes, hands, attention, languages and tools ask of one another.

The site stays unfinished. Sometimes that is a loop to leave. Sometimes it is the condition that lets a thought return before it is final.

Here I stop.

References§

Access Now. (2024). #KeepItOn: Restore internet access during student protests in Bangladesh. accessnow.org

Atay, A. (2020). What is cyber or digital autoethnography? International Review of Qualitative Research, 13(3). Issue contents

Crawford, K. (2021). Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press.

Desjardins, A., and Ball, A. (2018). Revealing tensions in autobiographical design in HCI. Proceedings of DIS ’18. doi.org/10.1145/3196709.3196781

Desjardins, A., Tomico, O., Lucero, A., Cecchinato, M. E., and Neustaedter, C. (2021). Introduction to the special issue on first-person methods in HCI. ACM Transactions on Computer-Human Interaction, 28(6), Article 37. doi.org/10.1145/3492342

Ellis, C., Adams, T. E., and Bochner, A. P. (2011). Autoethnography: An overview. Forum Qualitative Sozialforschung / Forum: Qualitative Social Research, 12(1), Art. 10. doi.org/10.17169/fqs-12.1.1589

Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., Maes, P., Phang, J., Lampe, M., Ahmad, L., and Agarwal, S. (2025). How AI and human behaviors shape psychosocial effects of extended chatbot use: A longitudinal randomized controlled study. Preprint. arXiv:2503.17473

Forlano, L. (2017). Posthumanism and design. She Ji: The Journal of Design, Economics, and Innovation, 3(1), 16–29. doi.org/10.1016/j.sheji.2017.08.001

Garland-Thomson, R. (1997). Extraordinary Bodies: Figuring Physical Disability in American Culture and Literature. Columbia University Press.

Garland-Thomson, R. (2011). Misfits: A feminist materialist disability concept. Hypatia, 26(3), 591–609. doi.org/10.1111/j.1527-2001.2011.01206.x

Geiger, R. S., and Ribes, D. (2011). Trace ethnography: Following coordination through documentary practices. Proceedings of the 44th Hawaii International Conference on System Sciences. stuartgeiger.com

Giaccardi, E., and Redström, J. (2020). Technology and more-than-human design. Design Issues, 36(4), 33–44. doi.org/10.1162/desi_a_00612

Hamraie, A. (2017). Building Access: Universal Design and the Politics of Disability. University of Minnesota Press. Project MUSE

Hofmann, M., Kasnitz, D., Mankoff, J., and Bennett, C. L. (2020). Living disability theory: Reflections on access, research, and design. Proceedings of ASSETS ’20, Article 4. doi.org/10.1145/3373625.3416996

Jackson, S. J. (2014). Rethinking repair. In T. Gillespie, P. J. Boczkowski and K. A. Foot (Eds.), Media Technologies: Essays on Communication, Materiality, and Society (pp. 221–239). MIT Press.

Kafer, A. (2013). Feminist, Queer, Crip. Indiana University Press. Project MUSE

Long, C. R., and Averill, J. R. (2003). Solitude: An exploration of benefits of being alone. Journal for the Theory of Social Behaviour, 33(1), 21–44. doi.org/10.1111/1468-5914.00204

Luccioni, A. S., Jernite, Y., and Strubell, E. (2024). Power hungry processing: Watts driving the cost of AI deployment? Proceedings of FAccT ’24. arXiv:2311.16863

Neustaedter, C., and Sengers, P. (2012). Autobiographical design in HCI research: Designing and learning through use-it-yourself. Proceedings of DIS ’12, 514–523. doi.org/10.1145/2317956.2318034

Parasuraman, R., and Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. Human Factors, 39(2), 230–253. doi.org/10.1518/001872097778543886

Price, M. (2015). The bodymind problem and the possibilities of pain. Hypatia, 30(1), 268–284. doi.org/10.1111/hypa.12127

Ratto, M. (2011). Critical making: Conceptual and material studies in technology and social life. The Information Society, 27(4), 252–260. doi.org/10.1080/01972243.2011.583819

Ross, F. G. E. (1999). The Printed Bengali Character and Its Evolution. Curzon Press.

Sarkar, A., and Drosos, I. (2025). Vibe coding: Programming through conversation with artificial intelligence. Proceedings of PPIG 2025. arXiv:2506.23253

Schön, D. A. (1983). The Reflective Practitioner: How Professionals Think in Action. Basic Books.

Schultz, T., Abdulla, D., Ansari, A., Canlı, E., Keshavarz, M., Kiem, M., Martins, L. P. de O., and Vieira de Oliveira, P. J. S. (2018). What is at stake with decolonizing design? A roundtable. Design and Culture, 10(1), 81–101. doi.org/10.1080/17547075.2018.1434368

Star, S. L., and Strauss, A. (1999). Layers of silence, arenas of voice: The ecology of visible and invisible work. Computer Supported Cooperative Work, 8, 9–30. doi.org/10.1023/A:1008651105359

Strubell, E., Ganesh, A., and McCallum, A. (2019). Energy and policy considerations for deep learning in NLP. Proceedings of ACL 2019, 3645–3650. doi.org/10.18653/v1/P19-1355

Suchman, L. (2007). Human-Machine Reconfigurations: Plans and Situated Actions (2nd ed.). Cambridge University Press. doi.org/10.1017/CBO9780511808418

Sundberg, J. (2014). Decolonizing posthumanist geographies. Cultural Geographies, 21(1), 33–47. doi.org/10.1177/1474474013486067

Tlostanova, M. V., and Mignolo, W. D. (2012). Learning to Unlearn: Decolonial Reflections from Eurasia and the Americas. Ohio State University Press.

Wobbrock, J. O. (2019). Situationally aware mobile devices for overcoming situational impairments. Proceedings of EICS ’19, Article 1. doi.org/10.1145/3319499.3330292

Zhao, Z. (2026). The interpreter as developer: An autoethnographic design study of building a computer-assisted interpreting tool through human-LLM co-creation. Preprint. arXiv:2610.16460

Endnotes§

The endnotes hold the lineage, the doubt and the seam.