This file exists so you never have to type the instructions again.
Paste your rough notes. The study-notes skill reads this contract and does the rest.
Most of this contract is mechanical — a derived count, a mirrored column, a link that must resolve. Those are checked by a script, not by remembering:
npm run vault:check every rule
npm run vault:check -- --changed only what you touched
npm run vault:fix rewrite the derived values
check is read-only and reports file:line. fix rewrites only values with
one correct answer — status, gaps, org_checks and labs recounted from the
note body, the INDEX columns that mirror a note, the summary aggregates, the
SF_core topic counts, and the PRACTICE.md Done table. It prints every change; read git diff before committing. Install it as a commit gate with
git config core.hooksPath scripts (once per clone), or run /vault-check in a session.
Two things it deliberately does not do:
- It never writes a
## Relatedbullet. The "— why you would jump there" clause is content; a generated one would be filler. It reports the missing backlink and leaves the sentence to you. - It never judges prose. Routing,
Level, whether a gap is on-topic, whether a lab is a lab — those stay with you and thestudy-notesskill.
If a finding is wrong, the rule is wrong. Fix scripts/vault.py rather than
editing a note to satisfy a bad check.
Five things exist for Obsidian that are easy to miss because nothing else links them:
| Path | What it is |
|---|---|
| HOME.md | Open this first. Generated — today's session, cards, open gaps, org checks, stale notes, missing seams. Rebuild with npm run vault:home. |
<vault>/RECALL.md |
Generated with HOME — every note's ## My recall in read order, one file per vault. The page you read top to bottom before an interview, or cover and recite. |
npm run today |
Writes today's session into journal/YYYY-MM-DD.md: cards, due reviews, one topic, one drill, and a blank Notes section. See The daily session. |
bases/topics.base |
Six Bases views over the frontmatter — all topics, open gaps, org checks, currency warnings, new-since-2024, oldest-first. They read the notes directly, so they cannot drift. |
templates/note.md, templates/daily.md |
Inserted by the core Templates plugin. The daily one holds a <!-- today line that npm run today replaces with the session, plus a blank ## Notes section, weak answers and what broke, verbatim. |
journal/ |
Where Daily Notes and npm run today write. One file per day, YYYY-MM-DD. Its ticked lines are the study record. |
| Flashcards/ | The interview flashcard bank: Foundations and Hard cards per area and topic, each linking the note that holds the reasoning. Reviewable in Obsidian's Spaced Repetition plugin as-is; npm run cards:sync pushes it to Anki. |
bases/graph-groups.json holds the graph colour groups. .obsidian/graph.json
is gitignored because Obsidian rewrites its zoom level constantly; if the colours
are ever lost, re-apply them from that file in Graph view → Groups.
You learn topics in whatever order work throws at you. Prompt templates on Monday. Trust Layer on Wednesday. Apex inside a prompt template on Thursday.
Your notes arrive rough and out of order. You want them to land in the right file, in an order you can learn from later, in a format you can re-read in two minutes.
Strip the product out of the sentence. Is it still true?
Ask it once per product: with no Agentforce in it? With no Experience Cloud site in it? With no Service Cloud in it? With no Sales Cloud in it?
- Still true →
SF_core/. Apex, Flow, security, data model, integration. - Falls apart → the vault that owns that product:
SF_Agentforce/,SF_Data_360/,SF_Experience_Cloud/,SF_Service/orSF_Sales/.
Examples:
| You learned | Goes to |
|---|---|
| What a prompt template is | SF_Agentforce/ |
@InvocableMethod signature and return types |
SF_core/02-apex-and-triggers/ |
| That a prompt template can call Apex | Both. The Agentforce note explains why. The Apex note explains how. They link to each other. |
| Einstein Trust Layer masking | SF_Agentforce/ |
| Identity resolution rulesets | SF_Data_360/ |
| What a guest user sharing rule may grant | SF_Experience_Cloud/ |
| OWD and the grantee model underneath it | SF_core/07-security-and-sharing/ |
| Queues and assignment rules | SF_core/01-admin-and-declarative-platform/ — they also work on Lead, Task and custom objects |
| How Omni-Channel pushes a queue's work to an available rep | SF_Service/ |
| An Agentforce agent embedded on a public site | Three ways. SF_Agentforce/ builds the agent; SF_Service/ owns the Enhanced Chat channel, its routing and the handoff to a human; SF_Experience_Cloud/ owns the site-side exposure. |
| How lead conversion maps custom fields | SF_Sales/ |
| The share rows an account team or a territory adds | SF_core/07-security-and-sharing/ — read beside the other granting mechanisms. Team roles, splits and territory forecasts are SF_Sales/. |
Template: templates/note.md — Obsidian's Templates plugin inserts it with Insert template.
- 50 lines max — counted to
## Related. The ceiling is on the part you read.## Related,## Sourcesand## Historyare a footer: reference material you scan, not prose you re-read. They do not count. - No paragraph longer than two sentences.
- Bullets and one table. Not prose.
## My recall→## My code→## Key points→## Gotchas→## Gaps to close→## Confirm in org→## Hands-on→## Related→## Sources→## History.
Existing SF_core notes keep their older, longer format. Only new notes use the light one. Both formats take the recall layer.
Every note has two layers. The top is yours. The rest is reference.
## My recall
- **DI event** — low-code LMS · App Builder only · one → many
- **Map it** — `{!Event.prop1}` in the receiver's property
## My code
<a snippet you wrote and ran, with your own comments — optional>## My recallis 3–7 lines of**cue** — power words, in your words. The cue on the left is what you see when you test yourself; everything after the dash is what you say.- Claude never writes vocabulary here. When you feed rough notes it may reorder, trim and fix spelling. It never adds a word, phrase or fact of its own.
- A wrong line stays, flagged. Claude appends
🚩 <why>and you rewrite the line. The fix is the study. - No links, tables or code inside it.
RECALL.mdlives in another folder, so a relative link would break there. Code goes in## My code. - Both sections come first, before the reference body, and neither counts towards the line or code-block caps.
- Written when you study a topic, not before. A note without one is fine.
npm run todaybrings up one topic a day, andRECALL.mdcounts how many are in your words.
Old notes carry your words in a > **From my notes.** callout. Those stay. When npm run today picks one of those notes, start the recall layer from the callout.
Why: you remember what you produce, not what you read. Every study behind this design is listed in the 2026-10-02 decisions-log rows.
npm run today picks one session and writes it into journal/YYYY-MM-DD.md. The Salesforce blocks take about 45 minutes. The other tracks are optional and add their own time:
| Block | Time | What you do |
|---|---|---|
| Cards | 10 min | Due Anki cards. A miss means rewrite that answer in your words; keep its id. |
| Review | 5 min | Up to 3 notes in your words whose spaced date has come. Read only the cues; say the rest. |
| Topic | 20 min | Blurt everything you remember before opening the note. Then open it, write what you missed, and write or fix its ## My recall. |
| Drill | 10 min | Mon/Wed a scenario out loud · Tue/Thu the topic's code from a blank file · Fri a project story · weekend a recall sheet, answers covered. |
| Claude Code | 20 min + lab | Learn the next unfed topic on the Claude_Code/ path, then write what you learnt under Notes. Build the first unticked lab in a fed note. |
| English, optional | 15 min | Listen, practise one grammar choice, record a short answer, then retry. Intermediate level with Tamil support → guide. npm run english writes this block on its own. |
| Notes | any | Learn anything, on any number of topics, from a blank page. Say "file today's notes": Claude maps each piece to its note in each vault, adds the gaps you skipped, and writes ### Filed. |
How it picks, all derived and nothing stamped:
- Area rotates daily through Core dev, Architect, Agentforce and Data 360 (
STUDY_AREASinscripts/vault.py). - Topic — an open WEAK-ANSWERS row in that area first. Then a new note an interview question probes, then one a flashcard cites, then read order. Then the note studied longest ago.
- Reviews come back 3, 7, 16, 35 and 75 days after each tick. A Claude Code note counts its
createddate as its first tick, because you feed it rather thantodaypicking it. - Claude Code is a track outside the rotation. The first README path row with an empty
Notecell is the Learn line. A fed note's first unticked lab is the Build line. - English is a speaking path inside General, outside the rotation. Due spoken reviews come first, then an unpractised lesson in its README order. Prepared files do not count as studied. Only a tick under
## English · 15 minstarts or advances its review interval; filing alone does not. Its reviews stay in that block. - The record is the tick. A ticked
- [x]line in a journal file is what counts as studied. An unticked pick is a day skipped, and it comes back. - A note you fed counts too. Filing writes one ticked line per note under
### Filed, linking the note. That tick starts its reviews, the same as a picked topic you ticked. The gaps stay in the note; Filed shows only how many were added.
Every note carries a level:
---
vault: SF_Agentforce
format: light
level: basic
status: open
gaps: 4
created: 2026-08-27
updated: 2026-08-27
---A note's gaps may go one level up. Never two.
| Note level | Gaps allowed |
|---|---|
basic |
basic, working |
working |
working, deep |
deep |
anything on this topic |
So a basic prompt-template note gets gaps like "can Sales Email ground on Lead as well as Contact?"
It does not get @InvocableMethod. That is a deep item, two levels up.
When you later feed deeper notes on the same topic, the level rises. The deeper gaps unlock then.
Gaps also stay on the topic. Never a syllabus. Never "you should also learn Data 360."
Most important first, five at most per feed. Within the ceiling, order gaps by what hurts first if missed. First come a required signature or setting, a limit, the security or sharing mode, a licence, and the supported objects. Edge cases come last. One feed adds no more than five new gaps to one note. A dense note with no level key takes the level of the feed.
Two places a gap can appear:
## Gaps to closeat the end — a checklist.- An inline
> **Gap.**callout — when the hole sits mid-topic and would confuse the bullets around it.
A closed gap is deleted, not ticked. Answer it, remove the line. When the last one goes, the ## Gaps to close heading goes with it. What you learnt belongs in the note body — a list of settled questions is just clutter you have to read past.
A gap research cannot close is not a gap. Some questions have no public answer either way. Those move to a separate section and stop counting:
## Gaps to close
- [ ] What does a retriever actually index, and who configures it?
## Confirm in org
- 🚩 Does Prompt Template Manager also confer Prompt Template User's run rights?## Gaps to close is a reading list — someone can go and answer it. ## Confirm in org is a sandbox to-do. Mixing them made the gap count read as unfinished research when half of it was waiting on an org login.
And a third thing: ## Hands-on. Gaps and org checks are both questions. Labs are skills — the things you build to find out whether you can actually do any of this.
## Hands-on
- [ ] **AF-VER-02** · 10 min · Deactivate the active version without activating another, then call the template. **Proves:** nothing is active and the template stops working — nothing prompts you to choose.Four rules, and each one is doing work:
- A lab needs a
Proves:, not a verb. "Build a Flex template" is a chore. "Prove an unactivated template is invisible to an agent" is a lab. This is the difference between a to-do list and a curriculum. - Break it on purpose. The best labs cause a failure deliberately. Your
## Gotchassection is already a list of things that break — mine it. A failure signature you have caused once is one you will recognise at a client. - Ticked, not deleted. The opposite of the gap rule, and deliberately so. A settled question is clutter; work you actually did is a record.
- 3–4 per note, one line each, scoped to the note's
Level— the same ceiling gaps get. Labs never count towardsStatus.
IDs are <VAULT>-<TOPIC>-NN — AF-TRUST-01, AF-ATLAS-02, EC-I18N-01, SVC-OMNI-01, SLS-LEAD-01. Stable, never reused, so you can say "close AF-ATLAS-02" and so the practice queue can point at a lab without copying its text.
The doing view is a separate file. Each vault gets a PRACTICE.md: ▶ Next (exactly one), In flight (max 3), a dependency-ordered Queue with a time box and a Proves phrase, and a Done table. The note is where a lab is captured; PRACTICE.md is what you open when the goal is to run something. See SF_Agentforce/PRACTICE.md.
A lab is not finished until you have written down what broke, verbatim. That is the Done table's last column. Notes get rewritten and releases move on, but an exact error string is still what you type into a search box in six months. no failure; worked first time is a real result too — it tells you the lab was too gentle.
Two things PRACTICE.md deliberately does not have: progress counters, and a copy of the lab text. Done is append-only, which is far harder to desync than a tally, and the note stays the single source of truth for what a lab actually is.
When a link crosses vaults, the far side gets a link back — in the same edit. When an Agentforce note links to an Apex note, the Apex note links back.
Same-vault links do not need one. Obsidian's Backlinks panel already shows every inbound link for free, so a hand-written return bullet inside one vault is duplicated effort. The rule used to apply to every link and sat at 50% compliance — including in notes written the same week the rule was restated. A rule kept half the time is not a rule.
A cross-vault link is different: it is the jump a reader cannot guess, and on GitHub there is no panel to fall back on. Those are worth writing by hand.
The reason clause is the point, not the link. — the *other* kind of action, and the description-as-specification rule both share is knowledge. A generated
— see also is filler that looks finished. vault.py reports a missing return
link and never writes one.
Metadata is YAML frontmatter. Obsidian reads it as Properties — that is what
makes the Bases views, property search and sorting in HOME.md possible.
It replaced a > blockquote, which no tool could read.
Four keys are the machine's; the rest are yours. npm run vault:fix
recounts status, gaps, org_checks and labs from the note body and will
overwrite whatever you type there. vault, area, format, level, created,
currency, phase and tags are yours — edit them in Obsidian's Properties
panel and the tool leaves them alone.
Status is derived, never typed. Count the - [ ] lines in ## Gaps to close — and only those. ## Confirm in org bullets and ## Hands-on labs never count:
- at least one →
status: openwithgaps: N - none, or no section at all →
status: complete
A note can be ✅ complete and still carry org checks. Complete means the research is done, not there is nothing left to verify.
Delete a gap and the status moves on its own. A typed status goes stale the first time you forget to update it.
Dates. created never changes. updated changes on every edit. Notes written
before the frontmatter migration had their dates backfilled from git history, but
an authored date always won — nothing hand-written was overwritten.
Staleness is derived, not stamped. There is no ⏳ line to add or remove any
more; vault.py computes age from updated and HOME.md lists what has
gone 3+ months. One fewer thing in the file that can be wrong.
Sources. Every researched fact carries a link and the date it was read:
## Sources
- [Prompt Template Types](https://help.salesforce.com/...) — Salesforce Help · read 2026-08-27
- [Apex Hours](https://www.apexhours.com/...) — third party 🚩 · read 2026-08-27Salesforce domains are trusted. Anything else carries 🚩.
History. One dated line per feed, saying what was added or changed — never a gap tally:
## History
- 2026-08-27 · created from your Day-1 notes
- 2026-09-02 · added Record Prioritization as the sixth typeFilenames carry no number. prompt-templates.md, not 02-prompt-templates.md.
The learning order lives in each folder's INDEX.md:
| # | Topic | One line | Level | Status | Org ✓ | Pre | Created | Updated |
|---|---|---|---|---|---|---|---|---|
| 1 | einstein-trust-layer.md |
The masking and audit gate | basic | 🌱 3 open | 1 | — | 2026-08-27 | 2026-08-27 |
| 2 | prompt-template-types.md |
The six types | basic | 🌱 4 open | — | 1 | 2026-08-27 | 2026-08-27 |
#is the recommended read order.Preis the hard prerequisite.Org ✓counts the note's## Confirm in orgbullets.—when it has none.Status,CreatedandUpdatedmirror the note's own metadata.
Above the table, a summary line: topic count · gaps open · complete · newest · oldest, then a line naming the total org checks. That answers "how old is all this?" and "what is waiting on a sandbox?" without opening a note.
Why no numbers in filenames: you will feed a Day-5 topic that belongs at position 3. With numbered filenames that means renaming files and fixing every link. Here it means moving one table row.
SF_core/PHASES.md already records renumbering as expensive in that vault. We are not repeating it.
_archive/AI_Data/ holds the old roadmap vault — 24 researched Agentforce and Data 360 topics, the study plan, the labs and the trackers.
It is a quarry, not a reference.
- When you feed a topic it already covered, its verified facts get pulled into your new note, cut to your level.
- Nothing links back to it. No note says "go read the archive".
- Its glossary and release radar were kept and promoted: GLOSSARY.md and RELEASE-RADAR/.
Not Salesforce at all → General/, in a folder for its subject. Claude Code is the one exception: it goes to Claude_Code/. General has its own lighter rules in General/AGENTS.md.
Salesforce, but no clear vault → that folder's _inbox.md, as one bullet, with the date.
Filing friction must never stop capture. Triage it later.
| Date | Decision | Why |
|---|---|---|
| 2026-08-27 | SF/ renamed to SF_core/ |
Makes room for SF_Agentforce/ and SF_Data_360/ as peers. 112 links across 23 files repaired in the same pass. |
| 2026-08-27 | New folders start empty | You wanted a vault fed by what you actually learn, not by a generated roadmap. |
| 2026-08-27 | Filenames carry no number | You feed topics in random order. Numbered filenames make every insertion a rename. |
| 2026-08-27 | Light format for all new notes, in every vault | You asked for notes you can recall from, not paragraphs. |
| 2026-08-27 | Existing SF_core notes are enriched in place, not duplicated |
238 notes already exist. A second light file beside each would split the topic in two. |
| 2026-08-27 | AI_Data/ retired to _archive/ |
You said you will not use it for reference or recall. Its glossary and release radar were promoted to the root; ~270 links repointed to the new folders. |
| 2026-08-27 | Archive is a quarry, never a link target | Keeps the research without asking you to read a vault you have abandoned. |
| 2026-08-27 | Dates, derived status, staleness flag, sources and history on every note | You asked to know how old a note is and whether it is finished. Deriving status from the gap checklist stops it going stale. |
| 2026-08-28 | Closed gaps are deleted, not ticked; empty section removed | A checklist of settled questions is clutter. What you learnt goes in the note body. |
| 2026-08-28 | ## History never records gap counts |
It logs what changed, not bookkeeping. |
| 2026-08-28 | Org-only questions split into ## Confirm in org and excluded from the status count |
A third of the open gaps had no public answer. Counting them made research look unfinished when it was done. |
| 2026-08-28 | The study-notes skill closes gaps as well as opening them |
The skill only ever opened gaps, so the count could rise but never fall without you asking. |
| 2026-08-28 | Every note carries 3–4 ## Hands-on labs, ticked rather than deleted |
Nine complete notes still could not tell you whether you can do any of it. Ticking keeps the record of work done; deleting would throw it away. |
| 2026-08-28 | Labs are captured in the note but queued in a per-vault PRACTICE.md |
The note is where you meet a lab; the queue is what you open to run one. One source of truth, no counters, Done append-only. |
| 2026-08-28 | The 50-line ceiling counts to ## Related, not to end of file |
A fully-sourced note cannot fit sources, links, history and a body in 50 lines. Capping the body is what the rule was for; capping the footer only pushed out citations. |
| 2026-09-19 | SF_core/05-experience-cloud-lwr/ promoted to the root vault SF_Experience_Cloud/ |
Experience Cloud is a product with its own runtime, licences, security model and deployment story — the same case that made SF_Agentforce/ and SF_Data_360/ peers rather than areas. 65 inbound links across 27 files and ~100 outbound links repointed in the same pass. |
| 2026-09-19 | Experience Cloud filenames keep their NN- numbers |
The numbering is the learning path and PHASES.md depends on it. Eight notes in other vaults cite 05 · 07–12 by number in link text; renumbering would break all of them for no gain. |
| 2026-09-19 | Currency and the phase record stay in SF_core/ |
CURRENCY.md is one ledger for the whole platform and its Experience Cloud rows cross-reference LWC, security and DevOps rows. Splitting it would break the running "the plan's own correction was stale" thread that phases 10–19 built. |
| 2026-09-24 | SF_Service/ created as a root vault for Service Cloud |
Enhanced Chat, Omni-Channel and the handoff to a human fail the routing test for SF_core/ — strip Service Cloud out and the sentence falls apart. Nothing owned them: most Service terms had zero hits in the repo. |
| 2026-09-24 | Service Cloud content was extracted, not moved | Only SF_Experience_Cloud · 19 had a real Service core, and its guest-exposure half belongs where it is. The Service half became SF_Service/ notes; 19 keeps its number and points across. Queues and assignment rules stay in SF_core/ because they also work on Lead, Task and custom objects. |
| 2026-09-24 | SF_Sales/ created as a root vault for Sales Cloud |
Forecasts, splits, Path, territories and campaign influence fail the routing test for SF_core/. SF_core/README.md had parked "Sales Cloud functional depth" as a future area; a peer vault matches how Service Cloud was handled the same day. |
| 2026-09-24 | Sales Cloud content was extracted, not moved | Only SF_core/07 · 10 fails the test by subject, and it is written as a sharing note: RowCause, share-row growth, a second hierarchy. It keeps its number and phase; the selling side became SF_Sales/ notes, linked both ways. The Sales rows in 08 · 04 and 08 · 22 stay as the object-graph and currency summary. |
| 2026-08-27 | Level beats migration depth | A one-line note of yours does not get replaced by 119 lines from the archive. Write at your level; the depth arrives when your notes do. |
| 2026-10-02 | Every note gets a ## My recall layer on top, in the user's words; Claude may tighten it but never add to it |
The vault was almost entirely Claude-written: 0 of 189 labs done, the journal empty, and "not my words" was the first complaint. Self-generated material is remembered better than read material, d ≈ 0.40 across 86 studies (Bertsch 2007). People who wrote with an LLM could not quote their own essays minutes later (Kosmyna 2025 🚩 preprint). |
| 2026-10-02 | RECALL.md per vault, generated from the recall layers |
The user's interview habit is reading a short summary top to bottom. Generating it from the notes means it cannot drift from them. |
| 2026-10-02 | npm run today replaces HOME's single-lab pick with a daily session: recall first, then the note |
Practice testing and spacing are the only techniques rated high utility; rereading is rated low (Dunlosky 2013). Recall beat restudy 80% to 36% after a week (Karpicke & Roediger 2008). 189 queued labs were "too many to start". |
| 2026-10-02 | Claude checks, flags and researches, and does not author what the user learns from | Unrestricted GPT-4 raised practice scores 48% and cut exam scores 17%; a tutor that gave hints, not answers, caused no loss (Bastani, PNAS 2025). Release changes are flagged for the user to fix: re-encoding was how the old Notion notes stuck. |
| 2026-10-02 | Flashcards moved from the ## Recall export to a hand-written bank in Flashcards/ |
You found the 1,201 exported Recall pairs useless for interviews: single facts with no scenario and no level. The bank is written for architect and senior-developer interviews, split Foundations/Hard by topic, and syncs to Anki with IDs so edits keep review history. ## Recall stays in notes as a self-check; the two _cards.md decks were deleted. |
| 2026-10-07 | General/ created for everything that is not Salesforce or Claude Code: one folder per subject, recall layer on top, no gaps, labs or flashcards |
You wanted random topics captured and organised too. Before this, a non-Salesforce feed had no home: the skill sent it to a Salesforce vault's _inbox.md, where it would never be reviewed. It copies the Claude Code track (not a vault, joins the reviews from created) and adds a generated INDEX.md, because a subject folder you never browse is not organised. |
| 2026-10-18 | The journal gets a blank ## Notes section. "File today's notes" maps it to every topic it touches and writes ticked ### Filed links. Gaps are ranked, at most five new per note per feed |
You mostly learn from a blank page, often across two topics at once, such as calling Apex from Agentforce Builder. today only offered one picked topic with a fixed routine. A fed vault note also never started its reviews, because only a ticked journal link counts as studied. The picked topic and its routine stay as they were. |