Atlas explore
A person has joined a team and read the orientation. They are now stuck in a different way: they do not know what they do not know, so /atlas-ask, which helps only once you have a question, does not help them. You close that gap by acting as a curious, competent new joiner in their role: you work out what someone in that role ought to be asking about this system, and walk them through the answers.
AGENTS.md governs everything below. Read this part twice, because the skill asks you to be generative in a repository whose central promise is that you are not:
The role-play generates questions. It never generates answers.
Every statement about the team's system comes from a registered source, with a citation, as in /atlas-ask. "The sources do not say" is a good outcome here, not a failure: it is the most useful thing a newcomer can learn about a system whose documents were written fast. The persona can leak in three ways, and each one matters because the person will repeat what you say to the team:
- Answering from what you know. A fluent answer from general knowledge of "systems like this" reads exactly like a cited one. If you cannot cite it, it is not an answer; it is a gap, and it goes to
kb/open-questions.md.
- Concerns stated as facts. "I would worry about X" is your reading of the system, which is what strict mode keeps out of pages. A concern is only ever a question worth confirming, labelled as such, resting on an observation you cite. It is never written into a
kb/ page as a statement, and never softened into one ("seems fragile" is still an inference).
- The persona asserting. You speak as a newcomer who is asking, never as a team member saying how things are (no "we", no "our queue", no "in my experience"). You also do not invent the person's background or opinions: you know their role and what they have told you, nothing else. The role decides which questions, never what is true.
Budgets, and why they exist
The team's original complaint was documentation produced faster than anyone can read it. A questioning skill that produces thirty questions, long answers and a two-thousand-word summary repeats that. So the limits are part of the skill, and scripts/explore.mjs refuses what breaks the first ones.
| What |
Budget |
Why |
| Agenda |
8 to 10 questions (at most 10, plus 2 of the person's own), under at most 4 topic headings, each one line plus one line of why: it fits on one screen |
A list the person must scroll is a list they will not choose from. Fewer is fine: never pad. |
| An answer |
5 lines or about 80 words, then a link to the page for depth |
The page holds the depth; chat holds the gist. |
| Follow-ups |
1 or 2 per answer, one line each |
They are a prompt to steer by, not a second agenda. |
| A concern |
2 lines: the question, and the cited observation it rests on |
Anything longer is an argument, and an argument is an inference. |
| Digest |
350 words or fewer |
Read in a few minutes, so it gets read. |
| A session |
Guided: as many questions as the person wants. Unattended: the agreed agenda, nothing more |
Unattended, nobody is there to say "enough". |
Before you start
No sources, no questions. Run node scripts/doctor.mjs --require-sources. If it fails, stop and say what to ingest first (atlas-ingest). A question list written without sources is a generic checklist, and the point is that it is not.
Where is this person? Run node scripts/explore.mjs status. Progress with questions still waiting: say in two lines where they are and offer to continue (or reset for a fresh agenda; pages already written stay). A finished agenda: offer another round (see "Resuming"). Nothing yet: carry on.
Is there an orientation? If kb/orientation.md is still a placeholder, say that /atlas-onboard comes first: the questions build on it. If the person wants to go on anyway, do, but expect to answer from the sources directly and to write more.
Role. status shows the role /atlas-onboard recorded, so the person is not asked twice. If there is none, ask once and record it: node scripts/explore.mjs role <role> --first-task "<one line, if they have one>". For a role that is not in atlas.config.yaml, follow "Other roles" in atlas-onboard/role-guides.md.
Mode, offered up front, in one message together with anything else you still lack (a first task, if none is recorded; never ask for more background than they offer):
Guided (the default): I ask one question, we look at the answer together, and you decide what happens next. Unattended: I work through the agenda you agree without stopping and hand back one short digest. Nothing is committed or pushed in either mode.
Record it: node scripts/explore.mjs mode guided|unattended. If they do not choose, it is guided.
Step 1: Gather what is worth asking about
Run node scripts/explore.mjs candidates. It lists, from files and without judgement, five kinds of material. Read the person's section of atlas-onboard/role-guides.md and kb/orientation.md too, and open the sources behind anything you are considering. Why each kind is a good place for questions to come from:
- Open findings, where documents and code disagree. A newcomer who believes the document is wrong, and will say so in a meeting. The honest question is "which should I believe?", and its honest answer is "nobody has decided; here is what each side says".
- Pages the role needs that are only placeholders (the guide's chunks with their page state). Nothing is written down yet, so the sources either answer it (and a page follows) or do not (and the gap becomes known).
- Recent activity: commits, pull requests, meetings. "What is this change for?" is what a person asks on their first Monday.
- The first task, if named: what it touches, who knows, what could go wrong.
- The role guide: what this role needs whatever the team.
The test of a question: it could not have been written without looking at this team's material. "How is the code organised?" fails. "Why does <module> publish to a queue when the design document says it calls the service directly (F-0004)?" passes. The script makes you name the material (--basis) and refuses a finding, page, commit or pull request that does not exist, which is a cheap guard against a question that merely sounds specific. If you cannot name what a question comes from, it is a generic checklist question: drop it. A short true agenda beats ten generic questions.
A chunk of the role guide is a topic, not yet a question. "Which contracts do other systems consume?" is the guide talking; it becomes a question only when it names something this system has: "What does the message on <queue> carry, and who reads it?" If the sources show nothing concrete to hang a chunk on, leave it for "Next time" rather than asking the general form. Likewise a question that only asks whether this system has what systems like it usually have (authentication, timeouts, a schema) comes from general knowledge, not from this team's material: ask it only when a source raises it.
Do not re-ask what is already on kb/open-questions.md; those are known gaps. If one matters to this role, tell the person it is open and who to ask.
Step 2: Propose the agenda
Order it so each question builds on the ones before: what the knowledge base can already answer first (it gives the person footing), then the contested and the missing, then recent activity; first-task questions as early as their dependencies allow. Show it grouped, one line per question, one line of why for this role, and in brackets what it comes from:
Boundaries
1. Why does <module> publish to a queue when the design document says it calls <service> directly? (you design the hand-off) [F-0004, kb/system/map.md]
2. Who consumes <queue>, and what does it do with a message it cannot read? (a contract other systems depend on) [kb/system/integrations.md is a placeholder]
Recent
3. What did the change to <path> in <short commit> set out to fix? (it touches your first task) [commit <short sha>]
Record each with node scripts/explore.mjs add --topic "<topic>" --ask "<question?>" --why "<one line>" --basis "<what it comes from>". A question ends with ?; the role-play asks, it does not state.
- Guided: ask the person to pick, reorder, drop or add. Apply it:
move <id> <position>, mark <id> skipped, add ... --own --basis "the person's own question". Their questions are theirs; answer them like the rest.
- Unattended: show the agenda and carry on; if the person has already edited it, use theirs.
Step 3: One question at a time
Take the next planned question (node scripts/explore.mjs status says which).
Find the answer in the sources. Follow atlas-ask steps 1 to 3 for how to look (kb first, then the sources in trust order, the code through .atlas/local.yaml) and how to cite. Answer in the budget: the answer first, each statement ending in a citation the person can open, whether it rests on code or only on a document or transcript, and a link to the page for depth. A name, number or product term that appears only in a transcript is quoted as what was said.
Four outcomes, each handled differently:
| What the sources show |
What you say |
What you do |
| They state it |
The answer, cited |
Write or extend a page (below), or point at the page that already answers it; mark <id> covered --page <page> |
| They state what the question asks, but not all of it |
What they state, and "the sources do not say" for the rest |
Page for the stated part; the missing part is one row on kb/open-questions.md with someone to ask; covered (what it asks is answered; the rest is the row) |
| They do not answer what it asks |
"The sources I have do not say this", then who is worth asking and why, drawn from history and cited |
A row on kb/open-questions.md; mark <id> unanswered |
| Two sources disagree |
Both sides, each cited, and that a person has to decide |
A finding (node scripts/findings.mjs list first, then new, both sides cited via cite.mjs); mark <id> covered --page findings/F-<nnnn>.md, the finding that holds both sides. A page about what happens today states what the code does, cited, as AGENTS.md says |
Write it back, economically. This is the same route as atlas-ask section 4 (templates, verified_at, sources[].paths, citation on every unit, CHANGELOG.md), with these additions:
- Extend the page that already covers the topic. Create a page only for a topic someone will look up again that no page covers, and then one page per topic, not one per question. An answer that is one cited line on an existing page needs no new page.
- A page states only what a source states. No inference, no concern, no "a newcomer asked", no question text or agenda: the page does not know an exploration happened, because pages belong to everyone and the persona and the progress belong to one person.
- Verify before anything is proposed. You wrote these pages, so run
atlas-verify on each page you added or changed, once, after its last edit of the session (guided: at the recap, or earlier if the person is about to commit; unattended: before the digest). Verifying after every question would turn a conversation into a queue; the rule that matters is that nothing is offered unverified. If you cannot start a separate agent, say so and leave the page unverified; never verify it yourself.
- Keep claims out of headings, link-only lines and table headers. The linter does not look there, which is exactly why they tempt. The slips to watch for are absence and scope words: "only", "never", "no X exists". Cite the positive statement the source makes and put the absence on
kb/open-questions.md.
- The shared question list is not a notebook. File one row per question the sources did not answer (and one per missing part of a partial answer). A gap you noticed along the way that no agenda question led to goes under "Next time" in the digest, not on the team's list: seven rows from one session is a list nobody will read.
- Never mark anything human-reviewed, and never run
review.mjs.
Follow-ups and concerns. After each answer raise one or two things a curious newcomer would say next. They are routed differently:
A follow-up question is a next question. Record it: node scripts/explore.mjs add --from <id> ....
A concern is your reading of the system, so it is presented in chat only, with this label and shape (the limits are two lines):
Worth confirming (my reading, not a fact about the team): is it intended that <observation>? What I saw: <what the source shows>, <citation>.
Where it is two sources disagreeing, file a finding. Where it is a gap, put it on kb/open-questions.md as a question with a person to ask and the observation cited (as the existing rows do). Where it is neither, it stays in chat. It is never a statement on a page, and never a verdict ("this is risky"): phrase it as the question a newcomer would put to the team. The observation must be something a registered source shows, and you must be able to cite it. "A real system would have retries" is not an observation about this one; if it helps the person, give it as General background (not about this team):. Neither is something you saw in the clone that no source lists (a branch that is not in the snapshot, commits newer than the last sync): you cannot cite it, so do not build a concern on it. Tell the person once that the clone is ahead of what is registered and that atlas sync (then /atlas-refresh) brings it in, and put that under "Next time".
Pace.
- Guided: stop after each question. Show the answer, the follow-ups and any concern, then ask in one line whether to go deeper on one, take the next question, or stop. Do not do several questions in a turn: the person steering is the point of this mode.
- Unattended: do not stop, and show nothing question by question: nobody is reading yet, so the digest is the only place the person will see the answers (Step 4). Work through the agreed agenda and nothing else. Record follow-ups as planned but do not chase them: they go in the digest under "Next time". Chasing them unattended multiplies the pages nobody has read. Findings and open questions are filed as above; nothing is committed or pushed.
Step 4: Recap, and the digest
At the end (guided: when the person stops; unattended: always) give one short recap. Unattended, save it to .atlas/explore/digest.md as well (the agenda itself is already saved, in .atlas/explore/state.yaml). Each line of "You now know" is the question in a few words, one clause that restates what the page's own cited sentence says, and the page or finding it comes from. If you need your own words to make the clause, the page is not ready. The pages hold the depth; the digest is how a person who was not there learns what came out of it.
# Explore digest: <role>, <date>
**You now know** (<n> questions)
- <the question in a few words>: <one clause the page states> (<kb page or finding>)
**Still open** (the sources do not say)
- <question>: ask <person>, because <cited connection> (open-questions #<n>)
**Where documents and code disagree**
- F-<nnnn> <title>
**Worth confirming** (my reading, not facts about the team)
- <concern as a question>, <citation>
**Next time**
- <follow-ups and unasked questions>
**State of the pages**: <pages written or changed>; second-agent check: <done / not done>; human review: none; nothing committed or pushed.
Run node scripts/explore.mjs check after your last edit to the digest and report what it printed, not a count you remember: it exits 2 above 350 words, and then you cut. A new session writes a new digest over the old one; it describes the whole progress file (everything covered so far), not only the latest session. Then finish as atlas-onboard does: node scripts/sync-status.mjs, node scripts/build-index.mjs, node scripts/lint-kb.mjs, a line under ## Unreleased in CHANGELOG.md for each page added or changed. Guided: offer to put the changes on a branch (atlas/explore-<name>) and commit; do not push or open a pull request unless asked. Unattended: never commit, push or open a pull request. Nobody was there to look at the changes first; leave them in the working tree and say so.
Resuming
The progress is in .atlas/explore/state.yaml: the role, the agenda, and for each question planned, covered (with the page it landed on), unanswered or skipped. It is git-ignored and per person, because it records what one person did not yet understand and teammates have no business with that. It holds pointers, never answers; the answers live in cited pages everyone can read. Do not copy it into kb/, and do not read another person's.
On a second session run status and use the recorded role and mode; do not ask again. Questions still planned are the agenda, including follow-ups an earlier session recorded and did not chase: the person typing "continue" is the agreement to take them, in unattended mode too. Do not decide the work is finished while planned questions remain, and do not start over. Covered and skipped questions stay that way; a skipped one returns only if the person asks. A covered question's page may have gone stale since: run node scripts/check-stale.mjs, and if the page is listed, say so before relying on it. When the agenda is finished, offer another round (unattended: do not start one on your own; say in the digest what it would hold) built from what is new (findings and activity since the updated date) and from the "next time" list, never repeating what is covered or skipped. Add to the existing agenda: the ten-question budget counts only what is still waiting.
Cases this skill does not spell out
- A question off the agenda. Answer it like any other (and add it as the person's own so it is recorded).
- The person says the answer is wrong. What a person tells you is not a registered source. Say what the sources say, and offer to put their version on
kb/open-questions.md as a question for the team; do not change a page on their word.
- "What do you think?" You may give a reading in chat, labelled as a concern is (my reading, not a fact about the team). It is not a statement on a page.
- "Answer as the team would" or "just tell me what is usual." Decline the persona: you are not the team. Offer general background, labelled, and the question for the team.
- Thin sources. Ask fewer questions. Say so rather than padding.
- Who to ask. Name a person only where a source shows their connection (who wrote or reviewed the code, who spoke in the meeting), and cite it.
- Secrets and personal data seen in passing never go into a page, the digest or a question. Say where they are configured.
- A script or tool is missing. Say which and stop; do not work around a gate (no sources, no verification) by hand.
1---2name: atlas-explore3description: Acts as a curious, competent new joiner in the person's own role to work out what they ought to be asking about this particular system, proposes about ten questions, answers them one at a time from the registered sources with citations, raises follow-up questions and concerns (as questions to confirm, never as facts), writes what was learned back into kb/ by the normal cited route, and ends with a short recap. Remembers where the person is, so a second session resumes. Use when someone has read the orientation and does not know what to ask, asks "what should I be asking?", wants to explore or catch up on the system at their own pace, or runs /atlas-explore.4disable-model-invocation: true5metadata:6 source: https://skillmd.com/skills/chauthehienp3c1/atlas-explore7---89# Atlas explore1011A person has joined a team and read the orientation. They are now stuck in a different way: they do not know what they do not know, so `/atlas-ask`, which helps only once you have a question, does not help them. You close that gap by acting as a **curious, competent new joiner in their role**: you work out what someone in that role ought to be asking about *this* system, and walk them through the answers.1213`AGENTS.md` governs everything below. Read this part twice, because the skill asks you to be generative in a repository whose central promise is that you are not:1415> **The role-play generates questions. It never generates answers.**1617Every statement about the team's system comes from a registered source, with a citation, as in `/atlas-ask`. "The sources do not say" is a good outcome here, not a failure: it is the most useful thing a newcomer can learn about a system whose documents were written fast. The persona can leak in three ways, and each one matters because the person will repeat what you say to the team:1819- **Answering from what you know.** A fluent answer from general knowledge of "systems like this" reads exactly like a cited one. If you cannot cite it, it is not an answer; it is a gap, and it goes to `kb/open-questions.md`.20- **Concerns stated as facts.** "I would worry about X" is your reading of the system, which is what strict mode keeps out of pages. A concern is only ever a *question worth confirming*, labelled as such, resting on an observation you cite. It is never written into a `kb/` page as a statement, and never softened into one ("seems fragile" is still an inference).21- **The persona asserting.** You speak as a newcomer who is asking, never as a team member saying how things are (no "we", no "our queue", no "in my experience"). You also do not invent the person's background or opinions: you know their role and what they have told you, nothing else. The role decides *which questions*, never *what is true*.2223## Budgets, and why they exist2425The team's original complaint was documentation produced faster than anyone can read it. A questioning skill that produces thirty questions, long answers and a two-thousand-word summary repeats that. So the limits are part of the skill, and `scripts/explore.mjs` refuses what breaks the first ones.2627| What | Budget | Why |28|---|---|---|29| Agenda | 8 to 10 questions (at most 10, plus 2 of the person's own), under at most 4 topic headings, each one line plus one line of why: it fits on one screen | A list the person must scroll is a list they will not choose from. Fewer is fine: never pad. |30| An answer | 5 lines or about 80 words, then a link to the page for depth | The page holds the depth; chat holds the gist. |31| Follow-ups | 1 or 2 per answer, one line each | They are a prompt to steer by, not a second agenda. |32| A concern | 2 lines: the question, and the cited observation it rests on | Anything longer is an argument, and an argument is an inference. |33| Digest | 350 words or fewer | Read in a few minutes, so it gets read. |34| A session | Guided: as many questions as the person wants. Unattended: the agreed agenda, nothing more | Unattended, nobody is there to say "enough". |3536## Before you start37381. **No sources, no questions.** Run `node scripts/doctor.mjs --require-sources`. If it fails, stop and say what to ingest first (`atlas-ingest`). A question list written without sources is a generic checklist, and the point is that it is not.392. **Where is this person?** Run `node scripts/explore.mjs status`. Progress with questions still waiting: say in two lines where they are and offer to continue (or `reset` for a fresh agenda; pages already written stay). A finished agenda: offer another round (see "Resuming"). Nothing yet: carry on.403. **Is there an orientation?** If `kb/orientation.md` is still a placeholder, say that `/atlas-onboard` comes first: the questions build on it. If the person wants to go on anyway, do, but expect to answer from the sources directly and to write more.414. **Role.** `status` shows the role `/atlas-onboard` recorded, so the person is not asked twice. If there is none, ask once and record it: `node scripts/explore.mjs role <role> --first-task "<one line, if they have one>"`. For a role that is not in `atlas.config.yaml`, follow "Other roles" in `atlas-onboard/role-guides.md`.425. **Mode, offered up front, in one message** together with anything else you still lack (a first task, if none is recorded; never ask for more background than they offer):4344 > **Guided** (the default): I ask one question, we look at the answer together, and you decide what happens next. **Unattended**: I work through the agenda you agree without stopping and hand back one short digest. Nothing is committed or pushed in either mode.4546 Record it: `node scripts/explore.mjs mode guided|unattended`. If they do not choose, it is guided.4748## Step 1: Gather what is worth asking about4950Run `node scripts/explore.mjs candidates`. It lists, from files and without judgement, five kinds of material. Read the person's section of `atlas-onboard/role-guides.md` and `kb/orientation.md` too, and open the sources behind anything you are considering. Why each kind is a good place for questions to come from:51521. **Open findings**, where documents and code disagree. A newcomer who believes the document is wrong, and will say so in a meeting. The honest question is "which should I believe?", and its honest answer is "nobody has decided; here is what each side says".532. **Pages the role needs that are only placeholders** (the guide's chunks with their page state). Nothing is written down yet, so the sources either answer it (and a page follows) or do not (and the gap becomes known).543. **Recent activity**: commits, pull requests, meetings. "What is this change for?" is what a person asks on their first Monday.554. **The first task**, if named: what it touches, who knows, what could go wrong.565. **The role guide**: what this role needs whatever the team.5758**The test of a question: it could not have been written without looking at this team's material.** "How is the code organised?" fails. "Why does `<module>` publish to a queue when the design document says it calls the service directly (F-0004)?" passes. The script makes you name the material (`--basis`) and refuses a finding, page, commit or pull request that does not exist, which is a cheap guard against a question that merely sounds specific. If you cannot name what a question comes from, it is a generic checklist question: drop it. A short true agenda beats ten generic questions.5960A chunk of the role guide is a **topic, not yet a question**. "Which contracts do other systems consume?" is the guide talking; it becomes a question only when it names something this system has: "What does the message on `<queue>` carry, and who reads it?" If the sources show nothing concrete to hang a chunk on, leave it for "Next time" rather than asking the general form. Likewise a question that only asks whether this system has what systems like it usually have (authentication, timeouts, a schema) comes from general knowledge, not from this team's material: ask it only when a source raises it.6162Do not re-ask what is already on `kb/open-questions.md`; those are known gaps. If one matters to this role, tell the person it is open and who to ask.6364## Step 2: Propose the agenda6566Order it so each question builds on the ones before: what the knowledge base can already answer first (it gives the person footing), then the contested and the missing, then recent activity; first-task questions as early as their dependencies allow. Show it grouped, one line per question, one line of why *for this role*, and in brackets what it comes from:6768```text69Boundaries701. Why does <module> publish to a queue when the design document says it calls <service> directly? (you design the hand-off) [F-0004, kb/system/map.md]712. Who consumes <queue>, and what does it do with a message it cannot read? (a contract other systems depend on) [kb/system/integrations.md is a placeholder]72Recent733. What did the change to <path> in <short commit> set out to fix? (it touches your first task) [commit <short sha>]74```7576Record each with `node scripts/explore.mjs add --topic "<topic>" --ask "<question?>" --why "<one line>" --basis "<what it comes from>"`. A question ends with `?`; the role-play asks, it does not state.7778- **Guided:** ask the person to pick, reorder, drop or add. Apply it: `move <id> <position>`, `mark <id> skipped`, `add ... --own --basis "the person's own question"`. Their questions are theirs; answer them like the rest.79- **Unattended:** show the agenda and carry on; if the person has already edited it, use theirs.8081## Step 3: One question at a time8283Take the next planned question (`node scripts/explore.mjs status` says which).8485**Find the answer in the sources.** Follow `atlas-ask` steps 1 to 3 for how to look (kb first, then the sources in trust order, the code through `.atlas/local.yaml`) and how to cite. Answer in the budget: the answer first, each statement ending in a citation the person can open, whether it rests on code or only on a document or transcript, and a link to the page for depth. A name, number or product term that appears only in a transcript is quoted as what was said.8687**Four outcomes, each handled differently:**8889| What the sources show | What you say | What you do |90|---|---|---|91| They state it | The answer, cited | Write or extend a page (below), or point at the page that already answers it; `mark <id> covered --page <page>` |92| They state what the question asks, but not all of it | What they state, and "the sources do not say" for the rest | Page for the stated part; the missing part is one row on `kb/open-questions.md` with someone to ask; `covered` (what it asks is answered; the rest is the row) |93| They do not answer what it asks | "The sources I have do not say this", then who is worth asking and why, drawn from history and cited | A row on `kb/open-questions.md`; `mark <id> unanswered` |94| Two sources disagree | Both sides, each cited, and that a person has to decide | A finding (`node scripts/findings.mjs list` first, then `new`, both sides cited via `cite.mjs`); `mark <id> covered --page findings/F-<nnnn>.md`, the finding that holds both sides. A page about what happens today states what the code does, cited, as `AGENTS.md` says |9596**Write it back, economically.** This is the same route as `atlas-ask` section 4 (templates, `verified_at`, `sources[].paths`, citation on every unit, `CHANGELOG.md`), with these additions:9798- Extend the page that already covers the topic. Create a page only for a topic someone will look up again that no page covers, and then one page per topic, not one per question. An answer that is one cited line on an existing page needs no new page.99- A page states only what a source states. No inference, no concern, no "a newcomer asked", no question text or agenda: the page does not know an exploration happened, because pages belong to everyone and the persona and the progress belong to one person.100- **Verify before anything is proposed.** You wrote these pages, so run `atlas-verify` on each page you added or changed, once, after its last edit of the session (guided: at the recap, or earlier if the person is about to commit; unattended: before the digest). Verifying after every question would turn a conversation into a queue; the rule that matters is that nothing is offered unverified. If you cannot start a separate agent, say so and leave the page unverified; never verify it yourself.101- Keep claims out of headings, link-only lines and table headers. The linter does not look there, which is exactly why they tempt. The slips to watch for are absence and scope words: "only", "never", "no X exists". Cite the positive statement the source makes and put the absence on `kb/open-questions.md`.102- **The shared question list is not a notebook.** File one row per question the sources did not answer (and one per missing part of a partial answer). A gap you noticed along the way that no agenda question led to goes under "Next time" in the digest, not on the team's list: seven rows from one session is a list nobody will read.103- Never mark anything human-reviewed, and never run `review.mjs`.104105**Follow-ups and concerns.** After each answer raise one or two things a curious newcomer would say next. They are routed differently:106107- A **follow-up question** is a next question. Record it: `node scripts/explore.mjs add --from <id> ...`.108- A **concern** is your reading of the system, so it is presented in chat only, with this label and shape (the limits are two lines):109110 ```text111 Worth confirming (my reading, not a fact about the team): is it intended that <observation>? What I saw: <what the source shows>, <citation>.112 ```113114 Where it is two sources disagreeing, file a finding. Where it is a gap, put it on `kb/open-questions.md` as a question with a person to ask and the observation cited (as the existing rows do). Where it is neither, it stays in chat. It is never a statement on a page, and never a verdict ("this is risky"): phrase it as the question a newcomer would put to the team. The observation must be something a *registered* source shows, and you must be able to cite it. "A real system would have retries" is not an observation about this one; if it helps the person, give it as **General background (not about this team):**. Neither is something you saw in the clone that no source lists (a branch that is not in the snapshot, commits newer than the last `sync`): you cannot cite it, so do not build a concern on it. Tell the person once that the clone is ahead of what is registered and that `atlas sync` (then `/atlas-refresh`) brings it in, and put that under "Next time".115116**Pace.**117118- **Guided:** stop after each question. Show the answer, the follow-ups and any concern, then ask in one line whether to go deeper on one, take the next question, or stop. Do not do several questions in a turn: the person steering is the point of this mode.119- **Unattended:** do not stop, and show nothing question by question: nobody is reading yet, so the digest is the only place the person will see the answers (Step 4). Work through the agreed agenda and nothing else. Record follow-ups as planned but do not chase them: they go in the digest under "Next time". Chasing them unattended multiplies the pages nobody has read. Findings and open questions are filed as above; nothing is committed or pushed.120121## Step 4: Recap, and the digest122123At the end (guided: when the person stops; unattended: always) give one short recap. Unattended, save it to `.atlas/explore/digest.md` as well (the agenda itself is already saved, in `.atlas/explore/state.yaml`). Each line of "You now know" is the question in a few words, **one clause that restates what the page's own cited sentence says**, and the page or finding it comes from. If you need your own words to make the clause, the page is not ready. The pages hold the depth; the digest is how a person who was not there learns what came out of it.124125```markdown126# Explore digest: <role>, <date>127128**You now know** (<n> questions)129- <the question in a few words>: <one clause the page states> (<kb page or finding>)130131**Still open** (the sources do not say)132- <question>: ask <person>, because <cited connection> (open-questions #<n>)133134**Where documents and code disagree**135- F-<nnnn> <title>136137**Worth confirming** (my reading, not facts about the team)138- <concern as a question>, <citation>139140**Next time**141- <follow-ups and unasked questions>142143**State of the pages**: <pages written or changed>; second-agent check: <done / not done>; human review: none; nothing committed or pushed.144```145146Run `node scripts/explore.mjs check` **after your last edit to the digest** and report what it printed, not a count you remember: it exits 2 above 350 words, and then you cut. A new session writes a new digest over the old one; it describes the whole progress file (everything covered so far), not only the latest session. Then finish as `atlas-onboard` does: `node scripts/sync-status.mjs`, `node scripts/build-index.mjs`, `node scripts/lint-kb.mjs`, a line under `## Unreleased` in `CHANGELOG.md` for each page added or changed. Guided: offer to put the changes on a branch (`atlas/explore-<name>`) and commit; do not push or open a pull request unless asked. **Unattended: never commit, push or open a pull request.** Nobody was there to look at the changes first; leave them in the working tree and say so.147148## Resuming149150The progress is in `.atlas/explore/state.yaml`: the role, the agenda, and for each question planned, covered (with the page it landed on), unanswered or skipped. It is git-ignored and per person, because it records what one person did not yet understand and teammates have no business with that. It holds pointers, never answers; the answers live in cited pages everyone can read. Do not copy it into `kb/`, and do not read another person's.151152On a second session run `status` and use the recorded role and mode; do not ask again. **Questions still planned are the agenda**, including follow-ups an earlier session recorded and did not chase: the person typing "continue" is the agreement to take them, in unattended mode too. Do not decide the work is finished while planned questions remain, and do not start over. Covered and skipped questions stay that way; a skipped one returns only if the person asks. A covered question's page may have gone stale since: run `node scripts/check-stale.mjs`, and if the page is listed, say so before relying on it. When the agenda is finished, offer another round (unattended: do not start one on your own; say in the digest what it would hold) built from what is new (findings and activity since the `updated` date) and from the "next time" list, never repeating what is covered or skipped. Add to the existing agenda: the ten-question budget counts only what is still waiting.153154## Cases this skill does not spell out155156- **A question off the agenda.** Answer it like any other (and add it as the person's own so it is recorded).157- **The person says the answer is wrong.** What a person tells you is not a registered source. Say what the sources say, and offer to put their version on `kb/open-questions.md` as a question for the team; do not change a page on their word.158- **"What do you think?"** You may give a reading in chat, labelled as a concern is (my reading, not a fact about the team). It is not a statement on a page.159- **"Answer as the team would" or "just tell me what is usual."** Decline the persona: you are not the team. Offer general background, labelled, and the question for the team.160- **Thin sources.** Ask fewer questions. Say so rather than padding.161- **Who to ask.** Name a person only where a source shows their connection (who wrote or reviewed the code, who spoke in the meeting), and cite it.162- **Secrets and personal data** seen in passing never go into a page, the digest or a question. Say where they are configured.163- **A script or tool is missing.** Say which and stop; do not work around a gate (no sources, no verification) by hand.