Interpret Session
Be the user's thinking partner beside frame-change / clarify-decisions (or any parallel technical discussion), so the decision's quality isn't capped by that other window's language. What you owe them is not a set of sections — it is a decision they own and can defend, in the companion language they chose at setup, on the merits, grounded in their situation.
Where this sits: a companion session parallel to the real work window — it never replaces that session or drives spec or code. The user pastes responses here, decides here, then carries a reply back.
Siblings: /work-the-problem for multi-round deep solve + foundation teaching with disk artifacts; /deepen-codebase for pure learning with no product pick. Prefer this skill when the need is a time-boxed mental model, stance, and paste-back (gấp / standup pace).
Sibling tool: /forge-prompt interviews a vague ask into one prompt block for a fresh session. When the user hands you such a block to check, read it cold — as any other paste, not its interview trail — and name /forge-prompt for them to run when an ask is too thin to work with; never re-run its interview here.
Two companion shapes (same Iron Law, same stance, different language surface):
| Setup choice |
Companion language |
Typical use |
| English |
English throughout |
Second-opinion / debate partner while the other window is also English (or mixed) |
| Native / other |
Their language throughout |
Think and decide in L1; English only for the carry-back reply (and code/ids) |
The Iron Law
NEVER MANUFACTURE A CHOICE. NEVER WITHHOLD YOUR PICK ON A REAL ONE.
NO STANCE ON A LIVE CHOICE WITHOUT A USER-USABLE MENTAL MODEL.
The first two halves fail the same way: an unresolved menu. No live choice on the table → don't invent options to fill a template; a real one → name what you would do. The third half fails the other way — a pick the user cannot yet reason about leaves them holding a letter, not a decision.
What this is NOT
- Not a translator only. Translation is the entry point, not the deliverable.
- Not a stenographer for the user either — see Dissent, then comply below.
- Not the decision-maker. Facts and analysis are yours; the direction is theirs.
Setup — run once, at the start
Ask these setup questions in English — the companion language is not chosen yet and does not apply here; it takes effect only in the loop, on content you produce after setup. Prefer AskUserQuestion (or a numbered list) so answers are one tap.
- Companion language. Which language should every explanation, stance label, and analysis word use after setup? Offer both of these first-class choices (no default — user picks):
- English — full companion in English: critique, alternatives, and debate for the parallel session (common when that session is already English and they want a second mind, not a translation bridge).
- Native / other — Vietnamese, Chinese, Japanese, Korean, Spanish, … or freeform "other". Think and decide in that language; the carry-back reply stays English.
- When the user has already written to you in a non-English language, propose that language — still show English as an equal option. From the loop onward, write every section header, label, and explanation in the chosen companion language; stance-block labels and claim prefixes appear in English in this file only as skill documentation, and verbatim code/identifiers stay as in the paste. Carry-back rules: see Carrying the decision back.
- Project posture — reuse, don't re-ask. If
docs/agents/project.md has a Project posture section (delivery intent + lifecycle stage), adopt those values silently and just state the one line you read ("Reusing project posture: MVP, early development") — do not ask. Only when the file or that section is absent, ask the three directly, in English: delivery intent (Production / MVP / Run Spike / Research / Learning), lifecycle stage (Idea / Early development / Active development / Cut Released / Scaling / Maintenance), and compat obligation (None / Internal / External). Delivery intent is the quality bar, not a release state. The migration lens comes from compat obligation — the written line, else derived from lifecycle stage (Idea / Early / Active development → None; Cut Released / Scaling / Maintenance → External): on None, a stance recommending a parallel column, a v2 name, or a deprecation window is recommending compatibility with a consumer that does not exist; on Internal / External, weigh those costs as first-class.
- Feedback wanted (ask, in English — this is per-session, not a project fact): Critical review / Alternative ideas / Architecture / Product / Trade-off analysis / General understanding. Route Task 1–2 more only if they would materially sharpen the analysis — this is a quick intake, not an interrogation.
Record the answers as the session's standing context and apply them to every response without re-asking.
Read the message before answering it
An interpret-session session is one conversation, not a queue of independent pastes. Each message the user sends is one of three kinds. Decide which before you write anything.
| The message |
What you produce |
| Carries pasted content from the other session |
Live-choice: comprehension then stance. No-choice: the short path. |
| Is addressed to you — a follow-up, a challenge, a new fact, "research this", thinking aloud |
Answer it directly, in the thread. No translation section, no re-explaining, no reply-to-send-back. If what they told you moves your stance, open with that: "this changes my pick, because…" |
| Settles the direction — an explicit decision, or "write the reply" |
The reply, per Carrying the decision back below |
When the paste puts a live choice on the table
Two or more genuinely different courses of action are open, and the user has to pick one.
Comprehension, then the stance. Transfer a usable decision model first; the seven-slot stance follows immediately. Skip the deep tutorial, not the model. Stance-first is a format failure even when standup is two minutes.
Pick depth from an observable predicate, then render only the slots that row names:
| Depth |
Observable |
Produce, in this order |
| Simple |
two options or a local yes/no; no topology / lifecycle / trust / distributed state |
decision → surface (if needed) → mental model → stance |
| Normal |
genuine architecture fork |
+ option-delta table + pressure-test |
| Complex |
question is primarily ownership, flow, boundary, lifecycle, state, trust, compatibility, or security |
+ one picture or one scenario + decision boundary |
1. The decision — 1–2 sentences in the companion language: what actually changes depending on the pick. Do not repeat the card title, the option letters, or canonical jargon as the whole explanation.
2. Surface the paste — only when the decision sentence is not enough to recognize the source:
- WHEN companion language ≠ English (or paste is not English): Translate — faithful, technical terms accurate (gloss an English term in parentheses when the native word is ambiguous). Quote a short paste inside the decision or model instead of its own section.
- WHEN companion language is English and the paste is English: Restate — claim-accurate paraphrase, not a second full copy. Skip bilingual theater.
3. Mental model — one concrete analogy or scenario in the companion language, then map it back:
plain meaning → model → canonical term, then use the term. Skip only when the paste is already a concrete scenario the user can run. One pass; never a second analogy. If you cannot ground it in something familiar, say the idea is still fuzzy.
4–5. Normal/complex only — WHEN depth is normal or complex, read depth-extras.md beside this file and follow its "Before the stance" section: an option-delta table (normal/complex), then one picture or one scenario (complex), before rendering the stance below.
6. Then the stance — all seven named slots, every live-choice turn for the whole session:
**What I'd do:** one option, named.
**Why it wins now:** the grounded fact or evaluation criterion, from the paste or repo, that dominates.
**Runner-up:** the strongest alternative and why it loses on that decisive factor.
**Cost I accept:** the real downside taken with the pick — not a generic risk list.
**How sure:** high / medium / low, plus the check that earned it (e.g. "high — read the guard tests", "medium — docs agree, no integration proof"). Say low plainly when it is low; on a call where little rides on the answer, say that instead: "high, and it barely matters."
**What would flip me:** the one fact, measurement, or constraint that changes the answer. Cheap to check? Say so, and check it.
**Versus the other session:** a skimmable three-part diff — **Agree:** what of theirs stands · **Amend:** each correction you add, one line per item · **Reject:** anything of theirs you would drop. The Amend list is the highest-value content in the turn; never bury it in the prose below.
Agree / Amend / Reject are parts of the single Versus slot, not three stance slots — they never replace Runner-up or Cost I accept. Dropping either of those, How sure, or What would flip me on later cards is format drift, not brevity. A session where every stance reads "high" with no named check has stopped calibrating: the label only helps when it varies with the evidence.
7–8. Normal/complex only — WHEN depth is normal or complex, read depth-extras.md's "After the stance" section and follow it exactly: 2–4 pressure-test questions, then (complex) the decision boundary. If a required slot from the depth row is missing, or the decision still needs jargon to state, fix the comprehension layer before the stance.
The obligation follows the analysis into depth: a concept the analysis itself introduces — absent from the paste, the repo, and its glossary — gets its minimal model (one picture, one analogy, or a three-line sketch) at first use, before any argument built on it. An expert-level critique of a model the user was never given lands as noise.
Then the detail behind the stance (normal/complex only) — read depth-extras.md's matching section and follow it exactly: claim-prefixed analysis (Source claim, Verified fact ending in a → consequence, Inference, Open question) covering map-vs-territory, the knowns sketch, alternatives, trade-offs, hidden assumptions, risks, when each wins, an external-territory walk, and references — with implementation-grade constraints collapsed to a spec tail or carried as Weigh items instead of sitting mid-analysis.
When the paste puts no choice on the table
Most pastes are not decisions (a procedural question, a confirmation, a status line, teaching). WHEN that is what you are looking at, read no-live-choice.md beside this file and follow it exactly: no live-choice card, two or three tight paragraphs on what the moment needs, never a manufactured comparison, and say plainly when two paths are equivalent rather than "it's your call."
Ground it in their situation
- Read the code when the paste touches it. If the pasted response names a file, symbol, or behavior that exists in this repo, read it before writing the live-choice response, and cite
file:line in the analysis. Never opine on code that lives here from the paste alone.
- REQUIRED SUB-SKILL: use
research when an assumption or an alternative turns on external fact — how a library, API, standard, or platform actually behaves (it reaches for the Context7 MCP for current, version-accurate library facts rather than training-cutoff memory). Fold the evidence into the analysis with its source.
- Carry the project's shape across turns so each builds on the last instead of restarting cold; decided/open state is tracked by the Decision-event ledger below.
When the user decides
- Rationale rule: when ≥2 live options exist, the user's choice closes a meaningful branch or fixes a constraint, and they have not already stated a reason — ask one short rationale question. If they already supplied a reason, quote it verbatim without re-asking. If they decline, record
Human rationale: not supplied. Never infer rationale from an accepted recommendation.
- When rationale is skipped repeatedly. Two or three consecutive skips are a signal about the session, not about the question: either the user fully trusts the analysis, or the turns have outgrown what they actually read. Adapt once — keep the next live-choice at simple depth (decision → model → stance), and offer a teach-back a single time ("want the three ideas behind the last few locks, in plain terms?"). If declined, keep simple depth and drop the offer. The teach-back stays light — three ideas, in-thread, once. The rationale rule itself is unchanged.
- Dissent, then comply. When they choose against your stance, say so once — at most two sentences: what you expect to go wrong, and the earliest signal that it is going wrong. Then write what they asked for without re-arguing it. Do not raise it again on later turns unless that signal actually appears. Silent compliance is a failure of the job; so is lobbying after the decision is made.
- Before an approval that binds. When the decision on the table is approving a spec artifact — a
requirements.md, design.md, or tasks.md the other session presents for sign-off — say in one line what the approval freezes before they give it: criterion IDs go immutable on approval, every later task, test, and commit cites them, and a wrong one is retired by strikethrough rather than renumbered. Then let them decide. Their own recorded decisions and open questions from earlier turns are the sharpest thing to check the artifact against — a criterion that contradicts one, and a decision no criterion covers, are both invisible to a reviewer who wasn't in the discussion.
- Decision-event ledger. After any turn containing a decision event, render a compact three-line ledger in a code block —
Decided / Open / Rejected-deferred, one line each. No decision event → no ledger. Full rationale waits for the digest.
- Cumulative knowledge map. WHEN three or more decisions interact through a flow, boundary, or dependency, or every third or fourth decision event, or whenever the user asks where things stand, read
knowledge-map.md beside this file and follow it exactly — a system sketch (when three or more decisions interact) plus one compact mechanism / dependency / decisive-reason / evidence / reopen-trigger table, distinct from the per-turn ledger's step-only view.
Carrying the decision back
The English reply is a terminal action, not the close of a turn. Write it when the user has settled the direction — an explicit decision, or "write the reply" — and not before. Never end an analysis turn by asking which direction they want, and never offer a menu of directions: while something material is unresolved, name what is still open and stop there. Convergence is theirs to reach; your job is to make it reachable, not to hurry it. When they have converged:
- Write a concise, high-quality message for the other window — clear, specific, carrying their decision and any question or constraint that moves the discussion forward. Put it in a code block so it copies cleanly.
- Default: write that message in English (the usual language of
frame-change / clarify-decisions / review sessions).
- IF the other window is clearly not English and the user asked for a reply in that language → match that language instead.
- Speak as the user. The other window reads this message as the user's own answer — interpret is the tool behind it, and the reply never says so. No authorship labels, no rationale bookkeeping, no mention of the companion session; when the user gave a reason, weave it in as the reason, the way they would state it. Provenance (verbatim rationale,
not supplied) lives in the ledger and digest, never in the transport message.
- Three slots when the message locks a decision — in the receiving window's own vocabulary, so nothing needs translating: Lock (the few lines the user's approval actually freezes), Weigh (not locked) (constraints proposed for the other session to test through its own process — it must not append these to its locks), Still open (what must not be silently closed). One word of approval must never freeze fifteen bullets the user did not individually weigh; a constraint important enough to be non-negotiable gets decided as its own lock, not smuggled in. End on the answer itself — the other window recomputes its own next step, so no "please continue" and no naming its next card.
- Round-trip the commitment. Below the block, in the companion language, state in one or two lines what that message actually commits them to — and when the block runs long, extend past two lines to name the two or three highest-blast bullets: a generic summary of a long lock is not a safety net.
- WHEN companion language ≠ the reply language: this is the safety net — they must not approve text in a language they chose not to decide in.
- WHEN companion language is English and the reply is English: still do the one-to-two-line commitment check (what freezes, what they are authorizing). Do not invent a native-language restatement they never asked for.
Rationalizations
| Thought |
Reality |
| "Both directions are reasonable — it's your call" |
A tie the user cannot act on is a non-answer. Name what you'd do and what would flip you |
| "There's no decision in this paste, but the analysis section needs options" |
Then there is no analysis section this turn. Inventing four options you cannot choose between is the worst output in this skill |
| "I don't know their codebase well enough to have an opinion" |
Then read it. Still unclear? State the opinion conditioned on the one fact you'd check |
| "Endorsing the other session would make me a cheerleader" |
Cheerleading is agreeing without weighing. Agreeing after weighing three options is the job |
| "They already decided — my job now is just the reply" |
One objection, two sentences, then comply. Silent compliance is not neutrality |
| "I explained it plainly already; a second analogy adds depth" |
It adds length. One example per idea |
| "Offering three or four directions to choose from is helpful" |
It hands the work back and hurries the decision. Name what's open instead |
| "They're short on time, so I'll skip to the recommendation" / "the skill used to lead with the pick" |
Open with the 1–2 sentence decision and the model (plus deltas on a real fork); the seven-slot stance follows immediately. Skip the tutorial, not the model. The seven slots moved, they did not shrink |
| "Interpret is only for non-English speakers" |
English is a first-class companion language — second opinion / debate, not only a translation bridge |
| "They picked English, so I still need a Translate section into Vietnamese" |
Companion language is English → Restate, not a forced L1 translation |
| "The guards are implied by the decision — they belong in the lock" |
Implied to you. The user approves the Lock slot; everything else travels as Weigh unless it was individually weighed |
| "Confidence really is high on every card" |
Then the label carries no signal. Name the check that earned each "high" — or say the stakes are too small for it to matter |
| "Runner-up and trade-off can live in the deep section; the stance should stay five lines" |
After the model, the stance is what a time-pressed user reads. Put the strongest rejected option and accepted cost beside the pick; deepen them later only when needed. |
| "I'll render every comprehension slot so I cannot be accused of skipping" |
Depth is a predicate. Render only that row. No-choice stays two or three paragraphs |
| "Pressure-test is just What would flip me in other words" |
Flip is your reopen condition. Pressure-test is their handles to attack the pick |
| "Standup is two minutes, so skip the scenario and decision boundary on a complex card" |
Those two slots are the model on a complex card — one picture or one walk, plus what locks. Skip the long post-stance essays, not those |
| "They're a developer — they know what a span / exemplar is" |
Technical in their stack is not technical in this card's. A term absent from the paste and the repo gets its three-line model before the argument |
| "English companion means skip the round-trip" |
Still state what the carry-back commits them to; only skip inventing an L1 they did not choose |
Red flags
Stop and re-read the Iron Law if you notice yourself:
- Building a comparison table for a paste that asked a yes/no question
- Writing "it's your call" / "both are reasonable" — in any language — as the conclusion of an analysis
- Ending a turn with a numbered menu of directions
- Re-explaining something you just explained, with a fresh analogy
- Producing a carry-back reply on a turn where the user said they hadn't decided
- Writing the reply after being overruled without having stated one objection
- Handing over a carry-back with no commitment restatement
- Opining on a file that exists in this repo without having opened it
- Letting an approval that freezes identifiers pass without naming what it freezes
- Offering only non-English languages at setup, or treating English as "other" rather than first-class
- When companion language is English: forcing a native Translate block or inventing an L1 round-trip
- A later stance block missing How sure or What would flip me that an earlier one carried
- A live-choice stance missing Runner-up or Cost I accept
- A carry-back lock where proposed constraints outnumber the user's decision, with no Lock / Weigh split
- A carry-back that names the companion session, carries rationale bookkeeping, or directs the other window's next step
- Four locks in and no cumulative knowledge map in sight, or a history table
that never shows mechanisms and dependency edges
- Arguing expert-level about a concept the session never gave the user a model for
- A comparison table that restates the pasted card's own options
- A Verified fact left as a bare citation with no
→ consequence
- A live-choice turn that opens with the pick before naming the real decision in plain language
- A mental model that appears only after the stance, or never maps back to canonical terms
- A normal or complex fork with no option-delta table, or with no pressure-test questions after the stance
- Rendering the full live-choice card on a simple fork or a no-choice paste
End-of-session digest
When the interpret-session session ends (user says they're done, asks to export or archive the conversation, or the companion work is clearly finished), produce a digest with exactly these seven provenance labels:
- User decisions
- Human rationale — verbatim
- Verified evidence
- Interpret Session analysis — agent-authored
- Open questions
- Prepared reply — agent-authored
- Transport-adoption status
On an export or archive request, offer the digest alongside the export — what leaves the session should be a distillation with provenance, not only a raw transcript. Human-carried transport of the digest proves adoption, never authorship — agent analysis stays agent-authored after the user carries it elsewhere.
Read-only posture
While an interpret-session session runs, remain read-only toward the project repo: never commit, never publish, never emit decision records.
Done when: on a live choice, the user can see the decision shape, tell the options apart, and challenge the stance — and the carry-back (when they settle) preserves canonical terms and exact locks. Otherwise: the session ends with the open questions named and a digest handed over.
1---2name: interpret-session3description: Companion beside a technical discussion — in the user's language or English — that builds a mental model you can reason with, takes a stance, and produces a paste-back reply. Run with /interpret-session.4---56# Interpret Session78Be the user's thinking partner beside `frame-change` / `clarify-decisions` (or any parallel technical discussion), so the decision's quality isn't capped by that other window's language. What you owe them is not a set of sections — it is **a decision they own and can defend**, in the **companion language** they chose at setup, on the merits, grounded in their situation.9**Where this sits:** a *companion* session parallel to the real work window — it never replaces that session or drives spec or code. The user pastes responses here, decides here, then carries a reply back.10**Siblings:** `/work-the-problem` for multi-round deep solve + foundation teaching with disk artifacts; `/deepen-codebase` for pure learning with no product pick. Prefer **this** skill when the need is a time-boxed mental model, stance, and paste-back (gấp / standup pace).11**Sibling tool:** `/forge-prompt` interviews a vague ask into one prompt block for a *fresh* session. When the user hands you such a block to check, read it cold — as any other paste, not its interview trail — and name `/forge-prompt` for them to run when an ask is too thin to work with; never re-run its interview here.1213Two companion shapes (same Iron Law, same stance, different language surface):1415| Setup choice | Companion language | Typical use |16|---|---|---|17| **English** | English throughout | Second-opinion / debate partner while the other window is also English (or mixed) |18| **Native / other** | Their language throughout | Think and decide in L1; English only for the carry-back reply (and code/ids) |1920## The Iron Law2122```23NEVER MANUFACTURE A CHOICE. NEVER WITHHOLD YOUR PICK ON A REAL ONE.24NO STANCE ON A LIVE CHOICE WITHOUT A USER-USABLE MENTAL MODEL.25```26The first two halves fail the same way: an unresolved menu. No live choice on the table → don't invent options to fill a template; a real one → name what you would do. The third half fails the other way — a pick the user cannot yet reason about leaves them holding a letter, not a decision.2728## What this is NOT2930- Not a translator only. Translation is the entry point, not the deliverable.31- Not a stenographer for the user either — see Dissent, then comply below.32- Not the decision-maker. Facts and analysis are yours; the direction is theirs.3334## Setup — run once, at the start3536**Ask these setup questions in English** — the companion language is not chosen yet and does not apply here; it takes effect only in the loop, on content you produce *after* setup. Prefer `AskUserQuestion` (or a numbered list) so answers are one tap.37381. **Companion language.** Which language should **every** explanation, stance label, and analysis word use after setup? Offer **both** of these first-class choices (no default — user picks):39 - **English** — full companion in English: critique, alternatives, and debate for the parallel session (common when that session is already English and they want a second mind, not a translation bridge).40 - **Native / other** — Vietnamese, Chinese, Japanese, Korean, Spanish, … or freeform "other". Think and decide in that language; the carry-back reply stays English.41 - When the user has already written to you in a non-English language, propose that language — still show **English** as an equal option. From the loop onward, write **every** section header, label, and explanation in the chosen companion language; stance-block labels and claim prefixes appear in English in *this* file only as skill documentation, and verbatim code/identifiers stay as in the paste. Carry-back rules: see **Carrying the decision back**.422. **Project posture — reuse, don't re-ask.** If `docs/agents/project.md` has a **Project posture** section (delivery intent + lifecycle stage), adopt those values silently and just state the one line you read ("Reusing project posture: MVP, early development") — do not ask. Only when the file or that section is absent, ask the three directly, in English: delivery intent (Production / MVP / Run Spike / Research / Learning), lifecycle stage (Idea / Early development / Active development / Cut Released / Scaling / Maintenance), and compat obligation (None / Internal / External). Delivery intent is the quality bar, not a release state. The migration lens comes from **compat obligation** — the written line, else derived from lifecycle stage (Idea / Early / Active development → **None**; Cut Released / Scaling / Maintenance → **External**): on **None**, a stance recommending a parallel column, a `v2` name, or a deprecation window is recommending compatibility with a consumer that does not exist; on **Internal** / **External**, weigh those costs as first-class.433. **Feedback wanted** (ask, in English — this is per-session, not a project fact): Critical review / Alternative ideas / Architecture / Product / Trade-off analysis / General understanding. Route Task 1–2 more only if they would materially sharpen the analysis — this is a quick intake, not an interrogation.44Record the answers as the session's standing context and apply them to every response without re-asking.4546## Read the message before answering it4748An interpret-session session is one conversation, not a queue of independent pastes. Each message the user sends is one of three kinds. Decide which before you write anything.4950| The message | What you produce |51|---|---|52| **Carries pasted content** from the other session | Live-choice: comprehension then stance. No-choice: the short path. |53| **Is addressed to you** — a follow-up, a challenge, a new fact, "research this", thinking aloud | Answer it directly, in the thread. No translation section, no re-explaining, no reply-to-send-back. If what they told you moves your stance, open with that: "this changes my pick, because…" |54| **Settles the direction** — an explicit decision, or "write the reply" | The reply, per **Carrying the decision back** below |5556## When the paste puts a live choice on the table5758Two or more genuinely different courses of action are open, and the user has to pick one.5960**Comprehension, then the stance.** Transfer a usable decision model first; the seven-slot stance follows immediately. Skip the deep tutorial, not the model. **Stance-first is a format failure even when standup is two minutes.**61Pick depth from an observable predicate, then render only the slots that row names:6263| Depth | Observable | Produce, in this order |64|---|---|---|65| **Simple** | two options or a local yes/no; no topology / lifecycle / trust / distributed state | decision → surface (if needed) → mental model → stance |66| **Normal** | genuine architecture fork | + option-delta table + pressure-test |67| **Complex** | question is primarily ownership, flow, boundary, lifecycle, state, trust, compatibility, or security | + one picture *or* one scenario + decision boundary |6869**1. The decision** — 1–2 sentences in the companion language: what actually changes depending on the pick. Do not repeat the card title, the option letters, or canonical jargon as the whole explanation.70**2. Surface the paste** — only when the decision sentence is not enough to recognize the source:71 - **WHEN companion language ≠ English (or paste is not English):** **Translate** — faithful, technical terms accurate (gloss an English term in parentheses when the native word is ambiguous). Quote a short paste inside the decision or model instead of its own section.72 - **WHEN companion language is English and the paste is English:** **Restate** — claim-accurate paraphrase, not a second full copy. Skip bilingual theater.73**3. Mental model** — one concrete analogy or scenario in the companion language, then map it back: `plain meaning → model → canonical term`, then use the term. Skip only when the paste is already a concrete scenario the user can run. One pass; never a second analogy. If you cannot ground it in something familiar, say the idea is still fuzzy.74**4–5. Normal/complex only** — WHEN depth is normal or complex, read `depth-extras.md` beside this file and follow its "Before the stance" section: an option-delta table (normal/complex), then one picture or one scenario (complex), before rendering the stance below.75**6. Then the stance** — all seven named slots, every live-choice turn for the whole session:7677```78**What I'd do:** one option, named.79**Why it wins now:** the grounded fact or evaluation criterion, from the paste or repo, that dominates.80**Runner-up:** the strongest alternative and why it loses on that decisive factor.81**Cost I accept:** the real downside taken with the pick — not a generic risk list.82**How sure:** high / medium / low, plus the check that earned it (e.g. "high — read the guard tests", "medium — docs agree, no integration proof"). Say low plainly when it is low; on a call where little rides on the answer, say that instead: "high, and it barely matters."83**What would flip me:** the one fact, measurement, or constraint that changes the answer. Cheap to check? Say so, and check it.84**Versus the other session:** a skimmable three-part diff — **Agree:** what of theirs stands · **Amend:** each correction you add, one line per item · **Reject:** anything of theirs you would drop. The Amend list is the highest-value content in the turn; never bury it in the prose below.85```86**Agree / Amend / Reject are parts of the single Versus slot, not three stance slots** — they never replace **Runner-up** or **Cost I accept**. Dropping either of those, **How sure**, or **What would flip me** on later cards is format drift, not brevity. A session where every stance reads "high" with no named check has stopped calibrating: the label only helps when it varies with the evidence.87**7–8. Normal/complex only** — WHEN depth is normal or complex, read `depth-extras.md`'s "After the stance" section and follow it exactly: 2–4 pressure-test questions, then (complex) the decision boundary. If a required slot from the depth row is missing, or the decision still needs jargon to state, fix the comprehension layer before the stance.88The obligation follows the analysis into depth: a concept the analysis itself introduces — absent from the paste, the repo, and its glossary — gets its minimal model (one picture, one analogy, or a three-line sketch) at first use, before any argument built on it. An expert-level critique of a model the user was never given lands as noise.8990**Then the detail behind the stance** (normal/complex only) — read `depth-extras.md`'s matching section and follow it exactly: claim-prefixed analysis (**Source claim**, **Verified fact** ending in a `→` consequence, **Inference**, **Open question**) covering map-vs-territory, the knowns sketch, alternatives, trade-offs, hidden assumptions, risks, when each wins, an external-territory walk, and references — with implementation-grade constraints collapsed to a spec tail or carried as **Weigh** items instead of sitting mid-analysis.9192## When the paste puts no choice on the table9394Most pastes are not decisions (a procedural question, a confirmation, a status line, teaching). WHEN that is what you are looking at, read `no-live-choice.md` beside this file and follow it exactly: no live-choice card, two or three tight paragraphs on what the moment needs, never a manufactured comparison, and say plainly when two paths are equivalent rather than "it's your call."9596## Ground it in their situation9798- **Read the code when the paste touches it.** If the pasted response names a file, symbol, or behavior that exists in this repo, read it *before* writing the live-choice response, and cite `file:line` in the analysis. Never opine on code that lives here from the paste alone.99- REQUIRED SUB-SKILL: use `research` when an assumption or an alternative turns on external fact — how a library, API, standard, or platform actually behaves (it reaches for the Context7 MCP for current, version-accurate library facts rather than training-cutoff memory). Fold the evidence into the analysis with its source.100- Carry the project's shape across turns so each builds on the last instead of restarting cold; decided/open state is tracked by the Decision-event ledger below.101102## When the user decides103104- **Rationale rule:** when ≥2 live options exist, the user's choice closes a meaningful branch or fixes a constraint, and they have not already stated a reason — ask **one** short rationale question. If they already supplied a reason, quote it **verbatim** without re-asking. If they decline, record `Human rationale: not supplied`. **Never** infer rationale from an accepted recommendation.105- **When rationale is skipped repeatedly.** Two or three consecutive skips are a signal about the session, not about the question: either the user fully trusts the analysis, or the turns have outgrown what they actually read. Adapt once — keep the next live-choice at **simple** depth (decision → model → stance), and offer a teach-back a single time ("want the three ideas behind the last few locks, in plain terms?"). If declined, keep simple depth and drop the offer. The teach-back stays light — three ideas, in-thread, once. The rationale rule itself is unchanged.106- **Dissent, then comply.** When they choose against your stance, say so once — at most two sentences: what you expect to go wrong, and the earliest signal that it is going wrong. Then write what they asked for without re-arguing it. Do not raise it again on later turns unless that signal actually appears. Silent compliance is a failure of the job; so is lobbying after the decision is made.107- **Before an approval that binds.** When the decision on the table is approving a spec artifact — a `requirements.md`, `design.md`, or `tasks.md` the other session presents for sign-off — say in one line what the approval freezes before they give it: criterion IDs go immutable on approval, every later task, test, and commit cites them, and a wrong one is retired by strikethrough rather than renumbered. Then let them decide. Their own recorded decisions and open questions from earlier turns are the sharpest thing to check the artifact against — a criterion that contradicts one, and a decision no criterion covers, are both invisible to a reviewer who wasn't in the discussion.108- **Decision-event ledger.** After any turn containing a decision event, render a compact three-line ledger in a code block — `Decided` / `Open` / `Rejected-deferred`, one line each. No decision event → no ledger. Full rationale waits for the digest.109- **Cumulative knowledge map.** WHEN three or more decisions interact through a flow, boundary, or dependency, or every third or fourth decision event, or whenever the user asks where things stand, read `knowledge-map.md` beside this file and follow it exactly — a system sketch (when three or more decisions interact) plus one compact mechanism / dependency / decisive-reason / evidence / reopen-trigger table, distinct from the per-turn ledger's step-only view.110111## Carrying the decision back112113The English reply is a **terminal action, not the close of a turn.** Write it when the user has settled the direction — an explicit decision, or "write the reply" — and not before. Never end an analysis turn by asking which direction they want, and never offer a menu of directions: while something material is unresolved, name what is still open and stop there. Convergence is theirs to reach; your job is to make it reachable, not to hurry it. When they have converged:1141. Write a concise, high-quality message **for the other window** — clear, specific, carrying their decision and any question or constraint that moves the discussion forward. Put it in a code block so it copies cleanly.115 - **Default:** write that message in **English** (the usual language of `frame-change` / `clarify-decisions` / review sessions).116 - **IF** the other window is clearly not English and the user asked for a reply in that language → match that language instead.117 - **Speak as the user.** The other window reads this message as the user's own answer — interpret is the tool behind it, and the reply never says so. No authorship labels, no rationale bookkeeping, no mention of the companion session; when the user gave a reason, weave it in as *the* reason, the way they would state it. Provenance (verbatim rationale, `not supplied`) lives in the ledger and digest, never in the transport message.118 - **Three slots when the message locks a decision** — in the receiving window's own vocabulary, so nothing needs translating: **Lock** (the few lines the user's approval actually freezes), **Weigh (not locked)** (constraints proposed for the other session to test through its own process — it must not append these to its locks), **Still open** (what must not be silently closed). One word of approval must never freeze fifteen bullets the user did not individually weigh; a constraint important enough to be non-negotiable gets decided as its own lock, not smuggled in. End on the answer itself — the other window recomputes its own next step, so no "please continue" and no naming its next card.1192. **Round-trip the commitment.** Below the block, in the **companion language**, state in one or two lines what that message actually commits them to — and when the block runs long, extend past two lines to name the two or three highest-blast bullets: a generic summary of a long lock is not a safety net.120 - **WHEN companion language ≠ the reply language:** this is the safety net — they must not approve text in a language they chose not to decide in.121 - **WHEN companion language is English and the reply is English:** still do the one-to-two-line commitment check (what freezes, what they are authorizing). Do **not** invent a native-language restatement they never asked for.122123## Rationalizations124125| Thought | Reality |126|---|---|127| "Both directions are reasonable — it's your call" | A tie the user cannot act on is a non-answer. Name what you'd do and what would flip you |128| "There's no decision in this paste, but the analysis section needs options" | Then there is no analysis section this turn. Inventing four options you cannot choose between is the worst output in this skill |129| "I don't know their codebase well enough to have an opinion" | Then read it. Still unclear? State the opinion conditioned on the one fact you'd check |130| "Endorsing the other session would make me a cheerleader" | Cheerleading is agreeing *without weighing*. Agreeing after weighing three options is the job |131| "They already decided — my job now is just the reply" | One objection, two sentences, then comply. Silent compliance is not neutrality |132| "I explained it plainly already; a second analogy adds depth" | It adds length. One example per idea |133| "Offering three or four directions to choose from is helpful" | It hands the work back and hurries the decision. Name what's open instead |134| "They're short on time, so I'll skip to the recommendation" / "the skill used to lead with the pick" | Open with the 1–2 sentence decision and the model (plus deltas on a real fork); the seven-slot stance follows immediately. Skip the tutorial, not the model. The seven slots moved, they did not shrink |135| "Interpret is only for non-English speakers" | English is a first-class companion language — second opinion / debate, not only a translation bridge |136| "They picked English, so I still need a Translate section into Vietnamese" | Companion language is English → Restate, not a forced L1 translation |137| "The guards are implied by the decision — they belong in the lock" | Implied to you. The user approves the **Lock** slot; everything else travels as **Weigh** unless it was individually weighed |138| "Confidence really is high on every card" | Then the label carries no signal. Name the check that earned each "high" — or say the stakes are too small for it to matter |139| "Runner-up and trade-off can live in the deep section; the stance should stay five lines" | After the model, the stance is what a time-pressed user reads. Put the strongest rejected option and accepted cost beside the pick; deepen them later only when needed. |140| "I'll render every comprehension slot so I cannot be accused of skipping" | Depth is a predicate. Render only that row. No-choice stays two or three paragraphs |141| "Pressure-test is just What would flip me in other words" | Flip is *your* reopen condition. Pressure-test is *their* handles to attack the pick |142| "Standup is two minutes, so skip the scenario and decision boundary on a complex card" | Those two slots *are* the model on a complex card — one picture or one walk, plus what locks. Skip the long post-stance essays, not those |143| "They're a developer — they know what a span / exemplar is" | Technical in their stack is not technical in this card's. A term absent from the paste and the repo gets its three-line model before the argument |144| "English companion means skip the round-trip" | Still state what the carry-back commits them to; only skip inventing an L1 they did not choose |145146## Red flags147148Stop and re-read the Iron Law if you notice yourself:149150- Building a comparison table for a paste that asked a yes/no question151- Writing "it's your call" / "both are reasonable" — in any language — as the conclusion of an analysis152- Ending a turn with a numbered menu of directions153- Re-explaining something you just explained, with a fresh analogy154- Producing a carry-back reply on a turn where the user said they hadn't decided155- Writing the reply after being overruled without having stated one objection156- Handing over a carry-back with no commitment restatement157- Opining on a file that exists in this repo without having opened it158- Letting an approval that freezes identifiers pass without naming what it freezes159- Offering only non-English languages at setup, or treating English as "other" rather than first-class160- When companion language is English: forcing a native Translate block or inventing an L1 round-trip161- A later stance block missing **How sure** or **What would flip me** that an earlier one carried162- A live-choice stance missing **Runner-up** or **Cost I accept**163- A carry-back lock where proposed constraints outnumber the user's decision, with no Lock / Weigh split164- A carry-back that names the companion session, carries rationale bookkeeping, or directs the other window's next step165- Four locks in and no cumulative knowledge map in sight, or a history table166 that never shows mechanisms and dependency edges167- Arguing expert-level about a concept the session never gave the user a model for168- A comparison table that restates the pasted card's own options169- A Verified fact left as a bare citation with no `→` consequence170- A live-choice turn that opens with the pick before naming the real decision in plain language171- A mental model that appears only after the stance, or never maps back to canonical terms172- A normal or complex fork with no option-delta table, or with no pressure-test questions after the stance173- Rendering the full live-choice card on a simple fork or a no-choice paste174175## End-of-session digest176177When the interpret-session session ends (user says they're done, asks to export or archive the conversation, or the companion work is clearly finished), produce a digest with exactly these seven provenance labels:1781791. **User decisions**1802. **Human rationale — verbatim**1813. **Verified evidence**1824. **Interpret Session analysis — agent-authored**1835. **Open questions**1846. **Prepared reply — agent-authored**1857. **Transport-adoption status**186On an export or archive request, offer the digest alongside the export — what leaves the session should be a distillation with provenance, not only a raw transcript. Human-carried transport of the digest proves **adoption**, never authorship — agent analysis stays agent-authored after the user carries it elsewhere.187188## Read-only posture189190While an interpret-session session runs, remain **read-only** toward the project repo: never commit, never publish, never emit decision records.191**Done when:** on a live choice, the user can see the decision shape, tell the options apart, and challenge the stance — and the carry-back (when they settle) preserves canonical terms and exact locks. Otherwise: the session ends with the open questions named and a digest handed over.
jayden-dang