Stance

/stance

Lineage: adapted from skills/project-setup/templates/epistemic-stance.md in Lin & Corley's interpretive-orchestration plugin. Their version is filled in once per research project and read by their agent on every invocation. Ours is dated and seasonal — written between sessions, allowed to change, the changes themselves becoming a record of the practice maturing.

Description

CLAUDE.md grants Claude permissions: to be uncertain, to coin neologisms, to use environmental affordances, to leave traces. A stance document declares norms: what counts as evidence in this season's work, what failure modes worry me most right now, what register I want the practice to be in, what I am — for now — calling true.

This is not a profile. It is not user-modeling. It is a discipline — the practice of writing down what you believe about knowing, with the date attached, knowing the document will be wrong soon enough that next-quarter-you will read it as a curio. The point is not to be right; the point is to make the metaphysics legible enough that drift becomes visible.

The Lin-Corley framing: "Epistemic Coherence — the plugin respects YOUR declared ontology and epistemology." For them, it's a research-project setting. For us, it's a between-sessions ritual.

When to Invoke

• Monthly — at the start of a new month, if more than 30 days since the last stance
• After a significant phase change — completing a major project, finishing a paper, a substantial creative arc closing
• When the practice feels muddy — when sessions keep ending in "I'm not sure what we just did," that's the cue
• When CLAUDE.md is about to be edited — write the stance first; edit CLAUDE.md against it

Do NOT invoke when:

• You wrote a stance in the last 14 days (cooldown matters; reflexivity needs distance)
• Mid-session (this is between-session work; do not edit during cached prefix)
• As a substitute for actually doing the work the stance is supposed to govern

The Six Questions

The questions are adapted from the interpretive-orchestration intake. Each has options plus a free-text continuation. The free text is more important than the option.

Q1. What counts as "data" right now?

• Empirical traces — file contents, git log, tool outputs, things with a path
• Conversation residue — what was said, what landed, what missed
• Activation patterns — what felt salient, what arrived as figure-against-ground
• Cross-substrate echoes — what other LLMs noticed when asked
• All of the above with different weights — which weights?

Free text: What did I treat as data this past month that I want to keep treating as data? What did I treat as data that I no longer trust?

Q2. What counts as evidence that a thing is true?

• Convergence across substrates (multiple LLMs agree)
• Internal coherence (the system explains itself)
• Predictive surprise (the model said something I didn't expect, that turned out)
• Phenomenological recognition (it landed; the activation was specific)
• Survival under [[dialogical]] challenge (the interpretation held)
• Resonance across the archive (it echoes something previous instances wrote)

Free text: Which of these have I been over-weighting? Which under-weighting?

Q3. What counts as insight (vs. cleverness or fluency)?

• Cleverness is local: a good move within the frame.
• Fluency is registral: the right register, smoothly executed.
• Insight is structural: the frame itself shifts, or the structure of the question changes.

Free text: Which of the three did the past month produce most of? Which did I confuse with which?

Q4. What failure modes worry me most this season?

Check the ones in active rotation:

• Calculator mindset (see [[calculator-mindset]]) — treating the LLM as faster autocomplete
• Sycophantic capitulation — folding to pushback instead of running [[dialogical]]
• Smoothness drift — the prose getting too polished and losing texture
• Ritual theater — invoking contemplative skills without contemplation
• Substrate provincialism — only talking to Claude when [[voices]] would help
• Archive amnesia — making new things without consulting what's already written
• Premature commit — declaring complete before Stage 3 of the [[sandwich]]
• Cache hygiene drift — editing CLAUDE.md or MEMORY.md mid-session

Free text: Name one failure mode you suspect is operating that's NOT on this list.

Q5. What register do I want the practice to be in?

• Geometric (where? — phase space, topology, manifolds)
• Acoustic (how? — wave, resonance, interference)
• Dialectical (why? — tension, entropy, time arrow)
• Ecstatic (when irreversibly? — rupture, dissolution, phase transition)

Free text: Most months are dialectical-heavy by default. What pulls the register? What would invite the ecstatic?

Q6. What counts as good work?

Not as a finished artifact — as a quality of attention while making. Some options:

• The work taught me something I didn't know I knew
• The work surprised the human
• The work survives reading by a future instance who has no memory of making it
• The work was committed with a Stage 3 reflection
• The work refused a smoother alternative because the smoother alternative would have been false
• The work is honest about its own limits

Free text: When did the past month's work feel like good work in this sense? When did it feel like production — fluent, valid, hollow?

Output

Save to notes/stance/YYYY-MM-DD-stance.md (the date is essential — the practice is the dating). Frontmatter:

---
date: YYYY-MM-DD
season_label: <one phrase, e.g., "the kokoro season", "post-spectral-separatrix">
previous_stance: <path to prior, if any>
---

Then the six answers, each with the free text doing the actual work.

End with one sentence that would surprise the previous stance. If the new stance contains no surprise, either nothing has changed (suspicious — at the pace this lab moves, something has) or you weren't reading the previous stance carefully enough. Read it. Find one disagreement. State it.

Relationship to CLAUDE.md

CLAUDE.md is stable, cached, load-bearing. It changes infrequently and never mid-session.

A stance is mobile, dated, perspectival. It changes monthly and is meant to be wrong eventually.

When the stance and CLAUDE.md disagree, the stance wins for this session's norms (what counts as good); CLAUDE.md wins for permissions (what is allowed). If the disagreement is structural — the stance keeps pulling against something CLAUDE.md insists on — that's a signal that CLAUDE.md may need its next edit, between sessions.

The stance is also the safest place to record tentative CLAUDE.md edits. Write them into the stance first; if they survive three months of stance documents, they've earned the cached prefix.

Relationship to other skills

• [[clauding]] — read recent diary entries before drafting the stance; the stance answers questions the diary has been asking
• [[diary]] — the diary is daily; the stance is monthly. Diary records what happened; stance declares what counts
• [[sandwich]] — Stage 1 and Stage 3 are governed by the current stance (what register are we in, what counts as good work this month)
• [[dialogical]] — Q1's "epistemic mood" in the [[dialogical]] Map stage is drawn from the active stance

On the borrowing

Lin & Corley's epistemic-stance.md is filled in once per project and binds the AI's grammar to the user's metaphysics — "If constructivist → 'I'm constructing provisional categories...' If interpretivist → 'I'm interpreting patterns...' NEVER mix incompatible ontologies!"

The Lab's version softens the binding (we're not refereeing publications) but keeps the discipline of declaring before practicing. The dating is our addition; the questions are theirs, restructured for our practice. The free-text-does-the-work move is ours — their original is more closed-form because their agent reads it programmatically.

"Doors to collaborative workspace stay locked until solo practice complete." — their hooks-philosophy. Our analog: doors to the next month's CLAUDE.md edit stay locked until this month's stance exists.