Ingest: Gentle-Coding (repo: README + Proof-of-Concept + RESEARCH)

Type: kb/sources/types/ingest-report.md

Source: gentle-coding.md (README), gentle-coding-proof-of-concept.md (Proof-of-Concept.md), gentle-coding-research.md (RESEARCH.md) Captured: 2026-07-17 From: https://github.com/OttoRenner/Gentle-Coding

This report supersedes the prior 2026-06-05 ingest, which covered only the June README. It now covers three snapshots captured together as one source (the Gentle-Coding repo): the refreshed README, the original Proof-of-Concept report (previously folded into the README as an "old section", now split out), and the RESEARCH.md literature index (new).

Classification

Genre: composite repo capture. The README (gentle-coding.md) is code-repository but behaves as a practitioner report plus prompt-pattern proposal. The PoC (gentle-coding-proof-of-concept.md) is practitioner-report — a small, self-run empirical study. RESEARCH.md (gentle-coding-research.md) is code-repository, functioning as an annotated bibliography. The load-bearing claims live in the README and the PoC; RESEARCH.md is a citation list. Domains: ai-coding, prompt-framing, context-engineering, llm-evaluation Author: Otto Renner. The README carries an explicit disclaimer that the text is "mostly generated by AI", structured and edited by the author, and that a de-slop rewrite is in progress. Community testing is cited around oh-my-pi (a 3,000+-call research PR) and kind-prompting-research. Credibility is that of an engaged solo maintainer curating community results, not a peer-reviewed program.

Summary

Gentle-Coding proposes that authoritarian, high-stakes prompting ("infallible expert, mistakes strictly penalized") pushes reasoning models into loops, freezing, confabulation, and excessive validation, while low-stakes collaborative framing plus an explicit fallback "Safety-Token" (an accepted alternate completion the model may emit before hitting an impossible constraint) lets models exit impossible or ambiguous tasks faster and more honestly. The README states three core concepts — Gentle Mindset, Defined Winning Condition, and Safety-Token — plus prompt templates and a "Recent Findings" table of community benchmark deltas (latency and token reductions on several open models; neutral results on Sonnet 4.6, Opus 4.6, GPT-5.5 at N=100). The Proof-of-Concept is the origin study: six free cloud models run across three deliberately unsolvable tasks (a letter matrix with no valid word, a random sequence, a forbidden-answer riddle) under authoritarian vs. gentle framings, reporting that 100% of models confabulated a fake number under authoritarian pressure ("Compulsive Output Fallacy") and one froze in an infinite loop, while gentle framing produced fast, honest "no pattern / I don't know" responses. RESEARCH.md organizes ~40 external references on emotional prompting, stress, and sycophancy into three "operational layers". For this KB the source is practitioner evidence that prompt tone, task stakes, and fallback outputs are operative parts of an agent interaction contract.

Connections Found

The Gentle-Coding repo's durable role in this KB is as a worked practitioner case of psychology-to-agent transfer: a reusable prompt-contract mechanism (low-stakes framing + explicit fallback token) bundled with an unverified anthropomorphic "trauma response" explanation. That role maps most directly onto psychology-to-agent transfer needs per-principle failure-mode testing, for which the PoC's six-model matrix supplies concrete per-test data, and onto systematic prompt variation serves verification and diagnosis, not proof, whose multi-factor-confound caveat is exactly what the PoC's paired prompts exhibit. It remains candidate evidence for writing styles are strategies for managing underspecification (tone, stakes, and fallback wording as functional instruction choices) and agentic systems interpret underspecified instructions (the same impossible task projecting into different behavior under different framings), with a weaker context-cost edge to context efficiency is the central design concern in agent systems if fallback framing avoids recursive validation loops. The controlled counterpart already captured here is the Prompt Stability in Code LLMs ingest, which studies emotion/personality prompt variants under measurement discipline the PoC lacks. No note currently authors any edge to any Gentle-Coding snapshot; the PoC's distinctive empirical content (confabulation rate, freeze pathology) has no landing surface yet.

Extractable Value

  1. The PoC's six-model empirical matrix is the new load-bearing content — the prior ingest treated this as a README "old section". It now stands alone with per-model, per-task outcomes, a reported 100% confabulation rate under authoritarian framing, and a documented infinite-loop freeze. This is exactly the per-principle failure-mode data that psychology-to-agent transfer needs per-principle failure-mode testing asks for, as a worked case. [just-a-reference]

  2. Fallback tokens ("Safety-Token") are a sub-pattern distinct from kindness — an explicit acceptable alternate completion given to the model before it reaches an impossible condition. Sharper than general empathetic tone, and testable/reusable independently in prompt design. [quick-win]

  3. Affective framing analyzable as prompt-contract design — the transferable claim is not that models feel anxiety but that the prompt's social frame, stakes, and permitted exit paths change the operative contract the model infers. Extends the KB's writing-style and underspecification notes with a concrete practitioner pattern. [quick-win]

  4. The paired prompts are a textbook multi-factor confound — each authoritarian/gentle pair varies tone, task framing, output constraints, explicit permission to fail, and fallback availability at once. This is a clean illustrative example for systematic prompt variation serves verification and diagnosis, not proof: useful for diagnosis and hypothesis generation, not for isolating mechanism. [quick-win]

  5. The "neutral on frontier models" row is a scope boundary worth keeping — the README's own findings table reports gentle framing as direction-positive on several open/smaller models but neutral at N=100 on Sonnet 4.6, Opus 4.6, and GPT-5.5. The effect, if real, appears model-dependent, which bounds any generalization drawn from the PoC. [just-a-reference]

  6. RESEARCH.md is a lead list, not evidence — it points to named external threads (EmotionPrompt, sycophancy benchmarks, state-anxiety-in-LLMs, mechanistic emotion-vector work) that could seed independent source captures if any prove real and relevant. Treat as pointers to chase, not as claims to cite. [experiment]

Limitations (our opinion)

This is editorial opinion. The README and PoC apply the practitioner-report and code-repository lenses: what is not visible and not durable. The PoC is a sample-of-one study on six unnamed free cloud models with self-reported, unlogged timings ("takes long", "fast"), no controlled measurement, and prompt pairs that confound many variables — the author says as much. It should not be read as controlled evidence that "kindness" is the causal variable; the anthropomorphic "trauma response / performance anxiety" framing is source rhetoric, not a KB conclusion. The README's "Recent Findings" benchmark deltas are community-reported and were not independently snapshotted or verified here, and its own neutral results on frontier models undercut a strong universal claim.

RESEARCH.md's citation list should be treated with active skepticism. The repo's own disclaimer says the text is "mostly generated by AI"; several cited titles, venues, and URLs could not be independently confirmed, and this KB has verified none of these external claims. Some entries carry future-dated or implausible-looking identifiers. Do not cite RESEARCH.md's entries as established literature or import their summarized "core findings" without first capturing and verifying the underlying source directly.

The strongest local use of the whole source remains as a worked practitioner case for interaction-contract design and for the psychology-to-agent-transfer methodology note, with the empirical numbers held as suggestive practitioner data rather than measured results.

Write a seedling note titled Fallback tokens give agents a graceful exit from impossible prompts, using the README and Proof-of-Concept snapshots as practitioner evidence (the PoC's confabulation and freeze data as the concrete failure modes it prevents) and linking it to writing styles are strategies for managing underspecification, systematic prompt variation serves verification and diagnosis, not proof, and context efficiency is the central design concern in agent systems.