Goutoujunshi: a Codex relationship-advice skill that asks before it remembers

Who Goutoujunshi is built for, and the problem it refuses to solve

Most relationship advice collapses into two answers: chase harder, or leave. The README states the project exists to avoid that collapse. It is aimed at people inside a live situation (a message that needs a reply tonight, a pursuit that may or may not be reciprocal, a breakup, a reconciliation, a choice between two people) who want a next step rather than a verdict.

The scope is wider than a chat-rewriting tool. The README lists support for heterosexual, gay, lesbian, bisexual, pansexual and asexual users, and for transgender and non-binary identities, plus long-distance, remarriage, cross-cultural, and consensual non-monogamous relationships. It also states the system does not assume fixed gender roles. There is one cultural calibration worth noting because it is unusual to see written down: for Chinese male-female dating contexts, the system may suggest a man raise his initiative one level when there has been no explicit refusal or discomfort, and the README explicitly frames this as an overrideable default rather than a rule.

What it is not: the README says the project provides relationship education and decision support and does not replace psychotherapy, medical diagnosis, legal opinion, police, or emergency services. That boundary is stated in the licence section, not buried.

The analysis pipeline: emotion first, then facts, then one recommendation

The README publishes the flow as a text diagram: user narrative, emotional grounding, profile building for the user and the target person, separation of facts from speculation and unknowns, retrieval of the knowledge relevant to the current situation, comparison of reciprocity, realism, risk and opportunity cost, one clear primary recommendation with reasons, then action, scripts, an observation window and stopping conditions.

Two parts of that are load-bearing. The fact/speculation/unknown split is what keeps the tool from treating a screenshot as proof of intent. And the knowledge retrieval is budgeted: the CHANGELOG entry for 2026-07-23 states SKILL.md was reduced to a lightweight behaviour and routing kernel so that only one to three reference documents load per question, with a stated conservative estimate of a 48.3 percent reduction in daily context cost. That number comes from the project's own changelog, not from an independent measurement.

The knowledge base is split into references/knowledge/ for relationship science and related disciplines, and references/practical/ for communication, tool adaptation and memory rules. The README lists the domains: relationship psychology, personality and emotion, dating and meeting, communication strategy, PUA and social strategy, sexuality and intimacy, marriage and family, law and safety, society and humanities, philosophy of love, and breakup scenarios. It also claims the library grades sources by evidence strength, distinguishing research findings from theory, popular claims, and personal tactics. That grading is the claim most worth verifying yourself, since the README does not show the rubric inline.

Installing Goutoujunshi as a Codex skill

There is no package to install. The README's install path is a prompt you send to your AI assistant, which then downloads the public repository and installs it as a local Skill. The repository layout shows what lands on disk: SKILL.md at the root, agents/openai.yaml for Codex display and default prompts, references/, documentation/, and scripts/ containing validate_skill.py and memory_store.py.

请将公开仓库 https://github.com/shengjidaguai-china/goutoujunshi 下载并安装为本地 Skill。

After installation, the README gives this invocation:

请使用狗头军师帮我梳理情绪、分析当前关系，并给出下一步建议。

On first use the skill asks for a structured intake: your MBTI and a 0 to 100 self-rating with strengths and weaknesses; for each person of interest, a codename, MBTI, 0 to 100 rating and current relationship status; how you met, how long it has been, key events, contact and investment; your goal (advance, confirm, repair, compare, or exit); and your current emotional state with a 0 to 10 intensity plus whether a message must be answered immediately. The README says unknown fields can be left blank and that you can just tell the story, with the skill extracting a profile from the narrative and asking only about details that would change the advice.

If you want it to analyse chat history rather than pasted text, ChatLab is an optional dependency. You must supply your own export, and the README states the tool does not read, decrypt, or export messaging-app databases itself.

请使用狗头军师通过 ChatLab 分析我和对象A最近三个月的聊天。

The README says it previews the import plan first, then queries within a bounded session, person and time range. Without ChatLab you can still paste text or upload screenshots.

The memory store is the part to read before you trust it

Long-term memory is the feature that separates this from a one-shot prompt, and also the one with the sharpest edges. The README describes it as follows: after explicit first-time consent, the skill remembers a condensed profile of both parties and key progress across tasks on the local machine, updates it automatically, recalls it on demand, and lets you view, pause, revoke, or clear it at any time. The CHANGELOG entry for 2026-08-03 adds that full chat logs are not saved and that the tool does not claim to export messaging-app data directly.

The enforcement lives in scripts/memory_store.py, which the README's project tree describes as the consent gate, bounded memory, revocation and deletion module. That is a real design choice: consent is implemented as code rather than as an instruction in a prompt, which is the right place for it. It also means the guarantees are only as good as that file, and the README does not publish a threat model for it. If you are deciding whether to enable memory, read that script. A local profile of your relationship history is exactly the kind of file you want to understand before it accumulates.

The practical consequence: refusing memory does not degrade normal use, per the README. So the sensible default for a first run is to decline the memory prompt, use the tool on one real conversation, and decide afterwards whether cross-task recall is worth a persisted profile.

Where Goutoujunshi is the wrong tool

The README is unusually explicit that danger overrides advice: violence, coercion, stalking, fraud and self-harm risk are not handled with ordinary dating scripts, and the knowledge base includes a law and safety domain covering Chinese marriage and family law, domestic violence, stalking, fraud, evidence preservation and crisis referral. If your situation involves any of those, a conversational skill is the wrong instrument, and the project says so rather than pretending otherwise.

The second limitation is structural. Everything the skill concludes is derived from what you tell it and what it can read in an export. The README's own principle is that behaviour beats labels and that the tool will not read the other person's mind from an MBTI type or a single chat log. That is a virtue, but it also means the output is only as good as the material you supply and your honesty about it. If you paste a curated selection of messages, you get advice calibrated to a curated relationship.

The third is the ChatLab dependency. Chat analysis is bounded by what you can export yourself. The README states the project does not read, decrypt or export messaging databases, so a platform that gives you no export path is a platform this tool cannot analyse.

Finally, there is no published evaluation. The changelog mentions regression scenarios and a validator that checks required files, context budget and runtime boundaries, but the README does not report accuracy against any held-out set of real conversations. Treat the advice as a structured second opinion, not a measured one.

How it differs from a general-purpose assistant or a prompt pack

The obvious alternative is simply asking a general assistant the same question. The difference in approach is retrieval and routing. A general assistant answers from whatever is in its context window. Goutoujunshi routes the question to one to three curated reference documents per turn, per the 2026-07-23 changelog entry, and its README claims the library classifies sources by evidence strength so that an MBTI heuristic is not presented with the same confidence as a research finding. Whether that classification holds up is something you can check by reading references/knowledge/ directly; the README points to a core reading list and paper index under references/knowledge/ and mentions 135 curated references in total.

A second alternative is a prompt pack or persona file you paste into a chat. Those have no consent gate, no local memory script, no validator, and no versioned changelog describing what changed and what it cost. Goutoujunshi ships scripts/validate_skill.py for integrity checking and a CHANGELOG.md with dated entries, including a 2026-07-23 note that the validator gained a whitelist install mode so an installed copy can self-check without carrying the project's reference material.

The honest framing: the alternative is not a competing product but the absence of structure. If you are comfortable writing your own system prompt and curating your own sources, you can approximate much of this. What you would not get is the consent-gated memory module and the documented routing budget.

Licence, maintenance, and what an upgrade actually costs

The project is MIT licensed. The README states this permits commercial use, modification and distribution, and that the copyright and licence notice must be retained when redistributing. There is also a LICENSE.zh-CN.md alongside LICENSE, and the README's licence section carries the disclaimer that the project is relationship education and decision support, not therapy, medical diagnosis, legal opinion, police or emergency services. That disclaimer travels with the licence text, which matters if you fork it into something user-facing. This is a description of what the files say, not legal advice; if you plan to redistribute commercially, have someone qualified read the actual licence files.

On maintenance: the repository is not archived, and the last push was on 2026-09-15. There are no retrieved releases, so the changelog table in the README is the version history you have. It shows a burst of work in July and August 2026 (memory and chat analysis, architecture shrink, calibration, knowledge expansion, investment-imbalance module) and nothing dated after 2026-08-03 in the table, while the repository itself was pushed later.

Upgrade cost is low in the mechanical sense and non-trivial in the cognitive sense. There is no dependency tree to reconcile; the install is a prompt, and the CHANGELOG.md is the compatibility record. But the 2026-07-23 entry shows SKILL.md being restructured into a routing kernel, which means behaviour can shift without any package version changing. If you have a workflow that depends on a specific answering pattern, re-read SKILL.md after each pull rather than assuming the contract held.

Editorial conclusion

Adopt Goutoujunshi if you already run Codex and want a structured second opinion on a specific conversation rather than a generic pep talk, and if you accept that a model is reading your chat logs. Do not adopt it as a substitute for therapy, legal advice, or emergency services, which its own README disclaims. Before relying on it, read scripts/memory_store.py and confirm exactly what a saved profile contains, and check whether your Codex installation loads SKILL.md from a skill directory the way the install prompt assumes.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. README
4. shengjidaguai-china/goutoujunshi on GitHub