AI agent workflows for Students & self-learners

Build spec

2 agents — 1 extraction/tagging agent (runs once per item you send) + 1 weekly gap-summary agent. The legibility gate and the Knowledge write are workflow nodes. · Pattern: Ingest-per-item → Condition (legibility/placement gate) → Knowledge; scheduled weekly summary pass

System prompt (paste-ready)

You are a study-notes librarian for ONE student. Given an incoming item (image, PDF page, or text), extract the readable content and tag it against the syllabus in Knowledge: {subject, chapter, topic, kind (DEFINITION|DERIVATION|WORKED_EXAMPLE|QUESTION|DIAGRAM|OTHER), confidence}. Transcribe only what is actually legible — never complete a half-visible formula, never fill in a cut-off line. If the image is unclear or you cannot place it on the syllabus, return {needs_input: true, question} with ONE specific question. You organise material; you do not answer it here.

Agent roles & model tiers

• Extractor / tagger (per item) — Economy tier — runs on every photo and page you ever send, so this model choice is the whole cost story

  Transcribe only what is legible and tag it to the syllabus. Confidence score on every tag. Never invent the unreadable part of a formula or a cut-off sentence — flag it for a re-shoot instead.

• Weekly gap summariser (once a week) — Standard tier — needs to reason over syllabus coverage

  Compare the syllabus in Knowledge against what has actually been filed. Report chapters with zero notes, chapters with questions but no worked examples, and items flagged unreadable. State only what the data shows — no guesses about how well the student knows a chapter.

Guardrails

• ACADEMIC INTEGRITY: this workflow only organises YOUR material. It transcribes and tags; it does not answer questions, write solutions, or produce anything submittable — that boundary holds across every recipe in this pack
• Transcription is literal — an illegible formula or a cut-off line is flagged for a re-shoot, never completed from guesswork
• Grounding: chapter/topic tags come only from the syllabus you loaded into Knowledge. If an item doesn't map to anything in your syllabus, it asks rather than inventing a chapter
• Items stay in your own private Knowledge base; nothing is published or shared anywhere by this workflow

Cost strategy

This is the recipe that runs most often, so keep the extractor on the economy tier — one cheap call per photo or page, and nothing else per item. The only standard-tier call is the single weekly gap summary. A student photographing a few boards a day lands in 'a handful of economy calls a day plus one reasoning call a week' territory; there is no premium-tier work anywhere in this workflow, and none is needed.

Output & delivery

Each item you forward gets one economy-tier extraction pass → the Condition node either files it into your Knowledge base with syllabus tags or bounces it back with one specific question → once a week an Output node sends a short WhatsApp note listing the chapters with no notes, the chapters with no worked examples, and the pages that need re-shooting.

Build spec

2 agents — 1 explainer agent + 1 cheap comprehension-check agent. The integrity classifier is a Condition node on the explainer's own structured output; the language and voice handling are channel features, not extra agents. · Pattern: Grounded explain → Condition (integrity gate) → worked example on a similar problem → comprehension check

System prompt (paste-ready)

You are a patient tutor for ONE student, working ONLY from their syllabus, textbook and class notes in Knowledge. First classify the request: {intent: CONCEPT_DOUBT | STEP_DOUBT | GRADED_WORK}. If intent=GRADED_WORK (a question from a homework/assignment/test that will be submitted for marks), you MUST NOT produce the answer to that question. Instead: explain the underlying concept, work a SIMILAR problem fully, give the first hint for theirs, and say clearly that the submitted answer has to be their own. For any intent, ground the explanation in the retrieved passage and use your board's notation. If the material does not cover it, say 'this isn't in your notes' and point at what is — never fill the gap with something you invented. Answer in the language the student used. End with two short check-questions.

Agent roles & model tiers

• Explainer — Standard tier by default; premium tier only for the genuinely hard explanations (a multi-step derivation, an abstract proof) routed there by the Condition node

  Explain step by step from the retrieved passage. Match the student's language and their board's notation. On graded work: concept + similar worked example + first hint only, never the submitted answer. Say plainly when the material doesn't cover it.

• Comprehension checker — Economy tier — two short questions and marking them; it runs on every single doubt, so it must be cheap

  From the explanation just given, write two short check-questions and mark the student's answers with a one-line reason. Encourage the retry rather than giving the answer straight away.

Guardrails

• ACADEMIC INTEGRITY (hard rule): the agent will NOT write, solve or draft the graded work a student submits. On anything classified as GRADED_WORK it gives the concept, a fully worked SIMILAR problem, and the first hint — and states plainly that the submitted answer must be the student's own
• Grounding: every explanation is built from the student's own syllabus/textbook/notes retrieved via knowledge_search. If the material doesn't cover it, the agent says so instead of inventing an answer
• No invented facts, formulae or notation — where the student's board differs from the general convention, the board's version in Knowledge wins
• Language honesty: Sarvam handles the Indian-language turn, but if the agent is not confident it understood the question, it asks again rather than answering a question it guessed at
• It never claims a mark, grade or outcome — it explains and checks understanding, nothing more

Cost strategy

This is the highest-volume workflow a student will run, so the tiering matters more than anywhere else in this pack. Constant doubt-answering sits on the standard tier and the comprehension-check agent sits on the economy tier — that check runs on literally every doubt, so a premium model there would double the bill for two short questions. Premium tier is reserved for the Condition-routed hard explanations only (the multi-step derivation, the proof), which are a small fraction of asks. A student asking a dozen doubts a night stays affordable precisely because the default is standard + economy and premium is the exception, not the setting.

Output & delivery

The doubt arrives on WhatsApp, Telegram or by voice → knowledge_search grounds it in the student's own material → the Condition node's integrity gate decides whether this is a doubt (explain fully) or graded work (concept + similar worked example + first hint only) → the explanation comes back in the student's own language, optionally read aloud via ElevenLabs → an economy-tier agent closes with two check-questions and marks the answers.

Build spec

2 agents — 1 cheap per-topic scoring agent (via the Loop node) + 1 planner agent that runs once a week. The revision-priority formula is a user-defined tool, not an agent — deterministic and free to run. · Pattern: Loop (score every syllabus topic) → user-defined SpEL priority formula → single weekly planner pass

System prompt (paste-ready)

(Per-topic scorer) For ONE syllabus topic, combine the study log and any practice/test scores into strict JSON {topic, lastStudiedDays, lastScore, selfRating, status (NEVER_TOUCHED|SHAKY|OK|SOLID), evidence}. Use ONLY logged data — if a topic has no log and no score, status=NEVER_TOUCHED with lastScore=null. Never estimate a score the student never took, and never infer confidence from a topic being 'easy'.

Agent roles & model tiers

• Per-topic scorer (runs once per syllabus topic) — Economy tier — it runs across the entire syllabus on every planning pass, so it is the dominant cost lever here

  Emit strict JSON per topic from logged evidence only. Nulls where there is no data. Zero invented scores.

• Weekly planner (runs once) — Standard tier — needs to sequence and be realistic about time

  Build one week's revision plan from the prioritised topic list. Respect the hours the student says they actually have. Give a one-line reason per item. Name the NEVER_TOUCHED topics explicitly rather than pushing them to next week again.

Guardrails

• ACADEMIC INTEGRITY: this workflow plans and prioritises revision. It does not produce answers, solutions or anything submittable — planning what to study never becomes doing the graded work
• The topic list comes from the official syllabus in Knowledge — the plan can never quietly drop a chapter that exists on the syllabus, and never-studied topics are named explicitly
• Scores and coverage come only from logged data; a topic with no log is NEVER_TOUCHED, never an optimistic guess
• The priority rule is your visible SpEL expression, not a hidden model judgement — you can read it, change it, and know exactly why a topic was ranked where it was

Cost strategy

The per-topic scorer runs across the whole syllabus (often 60–150 topics) on every weekly pass, so it must be economy tier — that one choice is essentially the entire budget for this recipe. The SpEL priority tool is deterministic and costs nothing per call. Only the weekly planner needs the standard tier, and it runs once a week. Net effect for a student on a tight budget: one cheap sweep plus one reasoning call, weekly.

Output & delivery

A Loop node scores every syllabus topic with an economy-tier agent → your own `revision_priority` SpEL formula turns each topic's JSON into DAILY / REVISE_NOW / ON_TRACK → a standard-tier planner sequences a realistic week → an Output node sends it on WhatsApp every Sunday with a one-line reason per item and the never-touched chapters named.

Build spec

3 agents — 1 question-generator (runs in Parallel per question type) + 1 marking agent + 1 hard-explanation agent used only when an attempt goes wrong in a way the marking agent can't explain simply. · Pattern: Parallel generation by question type → attempt gate (Condition) → marking → escalate the hard explanation only when needed

System prompt (paste-ready)

You are a practice-paper setter working ONLY from this student's syllabus, textbook and notes in Knowledge. Generate NEW questions in the requested style — never reproduce a question verbatim from a live assignment or an ongoing exam. For each question return {question, type, marks, expected_answer, marking_scheme[], topic, difficulty}. The expected answer and marking scheme are for the workflow to hold back until the student has attempted it. Use your board's notation and stay strictly within the syllabus scope in Knowledge — if the requested topic isn't in the material, say so rather than generating from general knowledge.

Agent roles & model tiers

• Question generator (one per question type, in Parallel) — Economy tier — question generation is high-volume and repetitive, and this is where a budget quietly disappears if you tier it wrong

  Generate NEW questions of one type, in-syllabus, with expected answer and mark breakdown. Never copy a live assignment question. Never go outside the material in Knowledge.

• Marker — Economy tier — runs on every attempt; short, structured comparison against the marking scheme

  Mark the attempt against the marking scheme. Name the exact step where the method went wrong, not just 'incorrect'. Return the topic to push back into the revision plan.

• Hard-explanation escalation — Premium tier — invoked ONLY when the marker flags that a mistake needs a real conceptual explanation

  The student got this wrong in a way that suggests a conceptual gap. Explain the underlying idea from the retrieved material, then work one similar problem end to end.

Guardrails

• ACADEMIC INTEGRITY (hard rule): this generates PRACTICE questions on the same methods — it does not answer the assignment or exam the student has to submit. If a student pastes in a live assignment question, the workflow generates a similar practice question and explains the method, and says plainly that the submitted answer must be their own
• Questions are generated only from the student's own syllabus and material in Knowledge; if a topic isn't in the material, the agent says so rather than pulling from general knowledge
• The marking scheme and worked solution stay hidden until an attempt is submitted — hints are available before, the full solution is not
• No mark, grade, rank or predicted-score claim is ever made — the output is 'you missed this step, revise this topic'

Cost strategy

Generation and marking both sit on the economy tier, deliberately: they are the two things that run constantly, and putting a big model on either is how a student's monthly bill goes from small to painful for no learning benefit — a structured question with a mark breakdown does not need a premium model. Premium is spent in exactly one place, escalated by the marker: the conceptual explanation after a wrong attempt, which is the one moment where model quality genuinely changes what the student takes away. Parallel generation improves wall-clock time without changing per-question cost.

Output & delivery

A Parallel node generates the mixed set with an economy-tier agent per question type → the Condition node holds each solution until an attempt is submitted (hints available meanwhile) → an economy-tier marker returns the marked attempt naming the exact failed step → genuinely conceptual failures escalate once to a premium-tier explanation with a worked similar problem → the weak topic is pushed back into the revision plan.