Skills

Appearance

Sources register ​

The evidence behind the skill's slow-rotting claims. Every claim the skill makes about the outside world that will not change week to week has an entry here; Mops answers "where did you get this?" — in any phrasing — from this file (REFERENCE §7, SKILL "Say what you know"). It is read by whichever flow needs it, and it is read first — a register entry serves every flow that meets its need, and the live web is where the register runs out, not where it starts (FLOWS → a shelf serves every flow).

What belongs here — and what never does. The register holds slow-rotting canon only: findings, methods and standards that age in years. Fast-rotting facts never enter — a price, a current API limit, a competitor's live feature stay fetch-at-decision-time rules, quoted with their check-date at the moment of use, never cached here to rot.

One fixed form per entry, so a wrong entry is visibly wrong: id · full citation · live URL/DOI/arXiv · archive link · licence · one-paragraph distillate (our words) · check-date · cited-by.

A back-pointer names a section and proves what it said — file.md#anchor (sha:…, checked …) — minted by scripts/fetch-source.py --cite <file.md>#<anchor>. A line number names a position, and a position moves the next time a paragraph is inserted above it: measured 2026-08-07, 11 of 23 line-number pointers no longer landed on their claim, with no edit to this register in between. The hash earns the rest — a passage rewritten underneath its citation makes the fact unknown and says so, where a line number cannot even see it. The one thing the anchored form cannot cite is a passage with no heading above it; that is a reason to give a load-bearing passage a heading, and until then the line form is used and marked as such.

Licence tiers (they decide what we may hold): free — an open licence (MIT, CC, a public standard); a copy may be carried later. copyrighted — citation + archive link + our own distillate, never the text itself. math — a formula recorded by name, which is not copyrightable. Each entry states its tier.

Upkeep. scripts/fetch-source.py builds and checks these entries: --resolve <doi|arxiv|url> prints a skeleton, --archive <url> triggers a Wayback snapshot, --verify walks every live URL. --verify-citations walks the other edge — every cited-by back into the doc that cites it — naming any pointer a rewrite left behind, and any entry the skill no longer cites at all (one parked for a later release stays named every run, which is the point: a deferral nobody is reminded of is a deletion). Both run each release (AGENTS.md → Cutting a release).

Persona theatre — grounding, fidelity, and the limits of synthetic audiences ​

park-self-reports · Park et al., self-report-grounded individual simulation ​

  • Citation: Park, J.S., et al. "LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals." arXiv:2411.10109 (2024; v1 was "Generative Agent Simulations of 1,000 People").
  • Live: https://arxiv.org/abs/2411.10109
  • Archive: http://web.archive.org/web/20260726212521/https://arxiv.org/abs/2411.10109
  • Licence: copyrighted (author © under arXiv's non-exclusive distribution licence) — cite + archive + our distillate
  • Distillate: Agents built from a person's own self-reports reproduce that person's survey answers at 83% (interview-grounded) / 82% (survey-grounded) / 86% (both) of the person's two-week test-retest ceiling, versus 74% for demographics-only; a free-text "persona paragraph" scores 0.71, below even the demographics baseline (0.74). Self-report grounding also reduces accuracy disparities across racial and ideological groups. The takeaway the skill leans on: the grounding artifact — the interview transcript — is the product, not a written bio.
  • Check-date: 2026-07-27
  • Cited-by: MODULES.md#persona-theatre-synthetic-and-live-audiences (sha:dbfc6761, checked 2026-08-07), MODULES.md#staging-proto-persona-validated-persona (sha:f8ca5f73, checked 2026-08-07), templates/PERSONA-template.md#bias-profile-24-named-biases-each-with-its-source (sha:80d2a6dd, checked 2026-08-07)

park-hai-brief · Park et al., Stanford HAI policy brief ​

ashokkumar-nature · Ashokkumar et al., direction not magnitude ​

  • Citation: Ashokkumar, A., Hewitt, L., Ghezae, I., Willer, R. Nature (advance online publication, 2026-07-08). doi:10.1038/s41586-026-10742-x.
  • Live: https://doi.org/10.1038/s41586-026-10742-x
  • Archive: pending (Save Page Now triggered 2026-07-27; availability: https://archive.org/wayback/available?url=https://doi.org/10.1038/s41586-026-10742-x)
  • Licence: copyrighted (Springer Nature) — cite + archive + our distillate
  • Distillate: Across a large replication set, LLM simulations track the direction of experimental effects at about r≈0.85 while systematically overestimating their magnitude. This is the evidence for the theatre's hardest rule: a synthetic verdict may state direction, never a magnitude (no "23% would churn").
  • Check-date: 2026-07-27
  • Cited-by: MODULES.md#persona-theatre-synthetic-and-live-audiences (sha:dbfc6761, checked 2026-08-07)

ls-types · Lewis & Sauro, a taxonomy of synthetic users ​

ls-review · Lewis & Sauro, a review of synthetic-user experiments ​

  • Citation: Lewis, J., Sauro, J. "A Review of Experiments with Synthetic Users." MeasuringU (2026-04-14).
  • Live: https://measuringu.com/review-of-experiments-with-synthetic-users/
  • Archive: http://web.archive.org/web/20260512065453/https://measuringu.com/review-of-experiments-with-synthetic-users/
  • Licence: copyrighted (MeasuringU) — cite + archive + our distillate
  • Distillate: Reviews ~12 recent experiments with synthetic users and finds mixed results, with synthetic responses showing artificially low variability and distorted magnitudes relative to real respondents — so they can indicate direction but not the size of an effect. (Their framing — low variability and distortion — is what the skill states, not "clustering toward neutral.")
  • Check-date: 2026-07-27
  • Cited-by: MODULES.md#persona-theatre-synthetic-and-live-audiences (sha:dbfc6761, checked 2026-08-07), MODULES.md#accuracy-score-and-consent-for-twins-of-real-people (sha:2deafaa5, checked 2026-08-07)

sharma-sycophancy · Sharma et al., sycophancy is trained in ​

  • Citation: Sharma, M., et al. "Towards Understanding Sycophancy in Language Models." ICLR 2024. arXiv:2310.13548 (2023).
  • Live: https://arxiv.org/abs/2310.13548
  • Archive: http://web.archive.org/web/20260725125159/https://arxiv.org/abs/2310.13548
  • Licence: copyrighted (author © under arXiv's non-exclusive distribution licence) — cite + archive + our distillate
  • Distillate: Sycophancy — telling the user what they want to hear — is a trained-in property of RLHF'd assistants, consistent across several models and tasks. A persona built on such a model inherits that compliance, which is why the theatre's calibration layer suppresses sycophancy explicitly (a synthetic respondent is a pleaser unless corrected).
  • Check-date: 2026-07-27
  • Cited-by: MODULES.md#bias-profiles-every-persona-carries-24-each-with-its-source (sha:a01bd41f, checked 2026-08-07)

tjuatja-biases · Tjuatja et al., LLM response biases ≠ human ones ​

  • Citation: Tjuatja, L., et al. "Do LLMs Exhibit Human-like Response Biases? A Case Study in Survey Design." TACL 12 (2024). arXiv:2311.04076 (2023).
  • Live: https://arxiv.org/abs/2311.04076
  • Archive: http://web.archive.org/web/20260116061015/https://arxiv.org/abs/2311.04076
  • Licence: copyrighted (author © under arXiv's non-exclusive distribution licence) — cite + archive + our distillate
  • Distillate: Tests whether LLMs reproduce known human survey response biases (acquiescence, question-order effects) and finds their biases do not reliably mirror human ones — sometimes absent, sometimes inverted. Caveats how far a synthetic survey respondent can stand in for a human one.
  • Check-date: 2026-07-27
  • Cited-by: theatre canon (skill distillation pending 2.7)

argyle-silicon · Argyle et al., silicon sampling and its diversity limits ​

  • Citation: Argyle, L.P., et al. "Out of One, Many: Using Language Models to Simulate Human Samples." Political Analysis (2023). doi:10.1017/pan.2023.2.
  • Live: https://doi.org/10.1017/pan.2023.2
  • Archive: pending (Save Page Now blocked HTTP 523 on 2026-07-27; availability: https://archive.org/wayback/available?url=https://doi.org/10.1017/pan.2023.2)
  • Licence: copyrighted (Cambridge University Press) — cite + archive + our distillate
  • Distillate: Introduces "silicon sampling" — conditioning an LLM on demographic backstories to simulate human survey samples — and shows it can reproduce some subgroup patterns while collapsing within-group diversity. Backs the caution that synthetic samples flatten variety rather than represent it.
  • Check-date: 2026-07-27
  • Cited-by: theatre canon (skill distillation pending 2.7)

wang-flattening · Wang et al., identity flattening ​

  • Citation: Wang, A., Morgenstern, J., Dickerson, J.P. "Large language models that replace human participants can harmfully misportray and flatten identity groups." Nature Machine Intelligence (2025). arXiv:2402.01908.
  • Live: https://arxiv.org/abs/2402.01908
  • Archive: http://web.archive.org/web/20260607174738/https://arxiv.org/abs/2402.01908
  • Licence: copyrighted (author © under arXiv's non-exclusive distribution licence; journal © Springer Nature) — cite + archive + our distillate
  • Distillate: Finds that using LLMs to replace human participants can harmfully misportray and flatten identity groups — reproducing majority stereotypes and erasing within-group variation. This is the direct evidence for the never-assign-a-bias-from-demographics rule: a demographic backstory produces a caricature, not a person.
  • Check-date: 2026-07-27
  • Cited-by: MODULES.md#staging-proto-persona-validated-persona (sha:f8ca5f73, checked 2026-08-07), templates/PERSONA-template.md#bias-profile-24-named-biases-each-with-its-source (sha:80d2a6dd, checked 2026-08-07)

kapania-simulacrum · Kapania et al., LLMs as qualitative participants ​

  • Citation: Kapania, S., et al. "'Simulacrum of Stories': Examining Large Language Models as Qualitative Research Participants." CHI 2025. arXiv:2409.19430 (2024).
  • Live: https://arxiv.org/abs/2409.19430
  • Archive: http://web.archive.org/web/20260411143601/https://arxiv.org/abs/2409.19430
  • Licence: copyrighted (author © under arXiv's non-exclusive distribution licence) — cite + archive + our distillate
  • Distillate: Treating LLMs as qualitative research participants yields plausible but hollow "simulacra of stories" that miss the lived specificity of real interviews. Marks the boundary of synthetic personas in qualitative work — a supplement, never a replacement for a real transcript.
  • Check-date: 2026-07-27
  • Cited-by: theatre canon (skill distillation pending 2.7)

Cost routing — cheap-first, conditional on a good verifier ​

frugalgpt · Chen, Zaharia & Zou, FrugalGPT ​

  • Citation: Chen, L., Zaharia, M., Zou, J. "FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance." arXiv:2305.05176 (2023).
  • Live: https://arxiv.org/abs/2305.05176
  • Archive: http://web.archive.org/web/20260722202256/https://arxiv.org/abs/2305.05176
  • Licence: copyrighted (author © under arXiv's non-exclusive distribution licence) — cite + archive + our distillate
  • Distillate: A cascade that queries cheaper models first and escalates only on low confidence can match the best single model's accuracy at up to −98% cost. The evidence for cheap-first-then-escalate routing at decomposition.
  • Check-date: 2026-07-27
  • Cited-by: ROLES.md#grades-fit-check-and-the-talent-pool (sha:99e0c8c2, checked 2026-08-07)

routerbench · Hu et al., RouterBench ​

  • Citation: Hu, Q.J., et al. "RouterBench: A Benchmark for Multi-LLM Routing Systems." arXiv:2403.12031 (2024).
  • Live: https://arxiv.org/abs/2403.12031
  • Archive: http://web.archive.org/web/20260606022825/https://arxiv.org/abs/2403.12031
  • Licence: copyrighted (author © under arXiv's non-exclusive distribution licence) — cite + archive + our distillate
  • Distillate: Cascades beat both any individual LLM and a zero-cost router only when the verifier is good — judge error ≤0.1, deteriorating past 0.2. The load-bearing caveat: cheap-first routing is conditional on a good verifier. In the skill, the review gates are that verifier, so the condition is already met — the caveat reads as a strength, not a risk.
  • Check-date: 2026-07-27
  • Cited-by: ROLES.md#grades-fit-check-and-the-talent-pool (sha:99e0c8c2, checked 2026-08-07)

Repository context files ​

agentsmd-eth · Gloaguen et al., do AGENTS.md files help? ​

  • Citation: Gloaguen, T., Mündler, N., Müller, M.N., Raychev, V., Vechev, M. (ETH Zurich). "Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents?" arXiv:2602.11988 (v2, 2026-06-23).
  • Live: https://arxiv.org/abs/2602.11988
  • Archive: http://web.archive.org/web/20260711121106/https://arxiv.org/abs/2602.11988
  • Licence: copyrighted (author © under arXiv's non-exclusive distribution licence) — cite + archive + our distillate
  • Distillate: A coding-agent benchmark finds repository-level context files (AGENTS.md) do not improve task success rates and add roughly +20% inference cost; LLM-generated context files perform slightly worse than none. The evidence behind "curate the shared guide, don't autogenerate it."
  • Check-date: 2026-07-27
  • Cited-by: skills/mops/SKILL.md:78 (line form: the claim sits in the core's preamble, above its first heading, so there is no section to anchor to)

Method provenance and standards (references, not evidence claims) ​

method-provenance · adapted methods, nothing embedded ​

  • Citation: cookiy — user-research-skill (MIT). agentman — "Synthetic Persona Creator" skill (concepts only).
  • Live: https://github.com/cookiy-ai/user-research-skill · https://agentman.ai/agentskills/skill/synthetic-persona-creator (owner-confirmed 2026-07-27; page verified live: three-layer persona architecture — identity foundation · context seeding · response calibration — plus cohort distribution)
  • Archive: http://web.archive.org/web/20260411081018/https://github.com/cookiy-ai/user-research-skill
  • Licence: free — cookiy's user-research-skill is MIT (a copy may be carried later; today only the method shape is adapted, nothing embedded). agentman: no licence stated on the page; concepts taken, nothing embedded (no code carried, so no licence obligation).
  • Distillate: Method lineage, not evidence. cookiy's MIT skill supplied the shape of the qualitative-research flows (its qualitative-research-planner → our persona-interview flow, its synthesize-research-report → our QDA step), adapted through the import gate. agentman supplied the calibration and cohort concepts behind the persona response-calibration layer. Recorded so every adaptation is auditable and no vendor wrapper is smuggled in.
  • Check-date: 2026-07-27
  • Cited-by: MODULES.md#staging-proto-persona-validated-persona (sha:f8ca5f73, checked 2026-08-07), MODULES.md#bias-profiles-every-persona-carries-24-each-with-its-source (sha:a01bd41f, checked 2026-08-07), MODULES.md#mixed-live-synthetic-hypothesis-beside-fact (sha:5f160353, checked 2026-08-07), ROLES.md#any-role-from-conversation-the-role-builder (sha:f3c1b1e6, checked 2026-08-07)

standards-cluster · named review standards ​