Part 1 · Design The Work Before You Place The Tool

1. Unit Of Value Hai Step, Session Nahi

“Main AI use karta hoon” aur “hamara workflow AI use karta hai” mein farq hai:

Isi liye is course ka unit step hai, session nahi. Session wo cheez hai jise aap baad mein judge karte ho, result achha laga ya nahi. Step wo cheez hai jiska decision aap pehle se lete ho, un grounds pe jo output achha lagne se nahi badalte. Opening ki 2 teams ke paas same product tha, farq ye tha ke ek ne step-level decisions liye, doosri ne session-level judgement.

2. Requirements Nikaalo, Phir Attack Karo

Real kaam kabhi clean brief se shuru nahi hota, 40-page document, adhoore emails, aur meeting mein kisi ki kahi hui baat se hota hai. Kuch banane se pehle, ye requirements banni chahiye: specific, traceable, testable. Ye 5 cheezein settle honi chahiye:

• Kya produce ho raha hai
• Kiske liye
• Kitni baar chalta hai
• Kaunse data se
• Kaunse format mein

Summary mat maango, structure maango, aur ek source column maango, isse koi bhi jo room mein nahi tha list check kar sakta hai. Extraction sirf pehla pass hai, trustworthy wala nahi, jo requirements teams ko bid haraati hain wo shayad hi seedhe likhi hoti hain, wo ek subordinate clause mein baithi hoti hain, extraction ye dhoondti hai jo likha gaya tha, wo nahi jo matlab tha, isliye extract karne ke baad, apni hi list review karwao.

3. Plan Utni Hi Achi Hai Jitne Uske Neeche Ke Numbers

Planning kaam mein 2 cheezein milti hain jinhe AI bilkul alag tarike se handle karta hai. Synthesis (considerations gather karna, options structure karna) AI achi karta hai. Calculation alag hai, is wajah se nahi ke model add nahi kar sakta, wo kar sakta hai, problem ye hai ke page pe aap bata nahi sakte ke usne kiya ya nahi. Ek likha hua figure aur ek calculate kiya hua figure identical dikhte hain.

“Agar ek result matter karta hai, usay compute karo. Kabhi likho mat.”

Data upload karo aur numbers code execution se produce karwao, sentence mein likhwane ki bajaye. Guessed utilisation rate pe bani staffing plan sirf table mein ek guess hai. Wahi plan actual timesheet data ke computed analysis pe bani ho to har line se defend ho sakti hai, kyunke har line ek calculation tak trace hoti hai jise koi bhi dobara chala sakta hai.

4. AI Ka Insight Kahan Actually Plan Badalta Hai

Ek operations lead ne poori capacity plan ek hi request mein maangi: ticket data analyse karo, growth trend nikalo, agle quarter ka headcount recommend karo. Jawab well-structured tha aur 3 analysts hire karne ki recommend kar raha tha. Team 2 hafte pehle announce hui hiring freeze mein thi. Trend analysis theek tha, recommendation usi se follow hui, plan useless is liye tha kyunke ek step ko aisi input chahiye thi jo dataset mein kahin nahi thi: ek meeting ka decision.

Research bhi ek synthesis step hai, isi liye window test usay bhi lagta hai, lekin ek alag tarah fail hoti hai: model ka knowledge ek date pe rukta hai, us ke baad ki har cheez window ke bahar hai:

Part 2 · The Delegation Map

Ye course ka core hai. Har step 3 decisions leta hai:

5. 3 Criteria Har Step Decide Karte Hain

AI ke sath redesign karne se pehle, steps order mein likho aur har ek ko classify karo: AI-appropriate, human-retained, ya collaborative.

6. Do Workflows, Same Criteria, Alag Answers

Contract review, poori tarah mapped:

Penalty-exposure row pe gauro se dekho, arithmetic hona step ko low-stakes nahi banata, wo delegable is liye hai kyunke result reversible hai aur agli row mein hi ek human ke saamne aata hai. Approval gate hata do, ye step AI-appropriate nahi rehta, chahe arithmetic wahi rahe.

Playbook row bhi same logic pe chalti hai, wo yahan AI-appropriate is liye hai kyunke koi agli row mein flags parhta hai. Wahi step ek workflow mein daalo jahan koi flags nahi parhta, wo collaborative ho jati hai. Step nahi badla, ye badla ke koi uske baad khada hai ya nahi.

7. Map Galat Hone Ke 3 Tareeke

Opening ki doosri team careless nahi thi, unhone hafton AI ko achi redlines draft karte dekha, phir easy clauses approve karne diye. Ye evidence ka reasonable response hai. Ye ek workflow ke liye risk kamane ka sabse common tareeqa bhi hai.

“Drafting quality draft ke baare mein evidence hai. Decision ke baare mein evidence nahi hai.”

8. Map Ka Owner Kaun Hai

Ek delegation map ek dafa likhi jati hai, phir chupke se galat ho jati hai, kyunke workflow uske neeche move ho gaya. Ek step add hua, ek reviewer ka role badla, quarterly volume double ho gaya. 3 controls isi tarah decay karte hain, koi signal diye bina:

Iska sirf ek countermeasure hai: ek scheduled read. Map pe 2 cheezein likhi honi chahiye: ek owner (ek named person, team nahi), aur ek review date (quarterly default hai). Review 4 sawalon mein 20 minute leta hai:

• Koi step add ya remove hua hai jo map pe nahi hai?
• Har collaborative step ke liye: pichle hafte kisne review kiya, kitna waqt diya? Agar koi jawab nahi de sakta, step automated ho chuka hai
• Kya volume ya stakes badle hain?
• Kya har configuration jis pe map depend karta hai abhi bhi current hai?

Part 3 · Build It, Aur Dekho

9. Loop: Ideate, Prototype, Feedback, Refine

AI ek design collaborator hai, vending machine nahi. Solution ek dafa maang ke nahi milta, loop chala ke milta hai, aur ise ek Project ke andar chalao takay context stable rahe.

Ek team ne 3 cycles mein bina code likhe ek dashboard artifact banaya. Cycle 1 mein sab kaam kar gaya. Cycle 2 mein date filter aur totals row maangi, filter theek tha, totals galat, subtle tarike se, kyunke totals likhe gaye the, compute nahi hue the. Cycle 3 (colour, print layout) pehli baar mein kaam kar gaya.

10. Jab Ye Prompt-and-Iterate Se Bahar Nikal Jaye

Wahi dashboard 6 mahine baad 3 departments har Monday khol rahe hain, aur ek board pack mein numbers ja rahe hain. Ye ab ek alag cheez hai, aur kisi ne decide nahi kiya ke ye ho.

Dependency escalation signal hai. Jis waqt doosre log kisi cheez ko infrastructure ki tarah rely karte hain, usay uptime, access control, aur ek fixed guarantee chahiye hoti hai, ye prompt-and-iterate ka kaam nahi rehta.

Part 4 · Jab Ye Bigadta Hai

11. 4 Causes, Timing Se Pehchano

Jab output disappoint kare, log 2 unproductive kaam karte hain: tool ko impossible declare karte hain, ya random words badalte rehte hain. Dono ek hi sawal skip karte hain: symptom sabse pehle kab dikha?

12. Sequence Sabse Sasta Pehle Chalta Hai

1. 1

   Prompt Dobara Parho

   5 components ke against check karo: role, context, task, constraints, output format

2. 2

   Conversation Length Check Karo

   Context overload ho sakta hai, restart ya summary chahiye

3. 3

   Feature Aur Model Check Karo

   Kya ye ek calculation hai jise code execution chahiye, ya complex task jo speed tier pe hai?

4. 4

   Configuration Check Karo

   Kya instructions, knowledge, aur Skills current hain?

5. 5

   Sirf Ab, Task Fit Pe Sawal Uthao

   Sabse expensive conclusion, sirf tab jab upar wale 4 rule out ho chuke hon

Part 5 · Fix Ko Permanent Banao

13. Reaction Ko Instruction Mein Badlo

Ek reactionbatati hai output kaisa laga (“too generic”). Ek instruction batati hai kya badle takay agla output different ho. Ek sawal se jump lagti hai: exactly kya present hona chahiye takay ye sahi ho, aur setup ka kaunsa hissa usay control karta hai?

14. Rule, Reference, Ya Procedure

Fix dhoondna aasan hissa hai. Mehnga failure use dhoond ke wapis kho dena hai. Test chhota hai: kya ye correction dobara chahiye hogi, mujhe ya kisi aur ko? Agar haan, configuration mein jani chahiye.

15. Optimize Karne Se Pehle Friction Dhoondo

Variance wo hai jo teams miss karti hain, kyunke koi individually experience nahi karta, sab ka apna output consistent hai, inconsistency sirf logon ke darmiyan hai, isliye ye reviewer pe dikhti hai, source pe nahi.

• Consolidate: jo steps saath chal saktay hain unhe saath chalao, 3 separate prompts jo same background maangte hain, ek prompt hain
• Promote: repeated pattern ko configuration mein le jao (Concept 14)

16. Wo Cheez Measure Karo Jo Actually Matter Karti Hai

Pehle baseline measure karo. Ek workflow ko badalne se pehle ek dafa chalao aur 3 numbers record karo: kitna waqt laga, kitne revision rounds chahiye huay, kitne manual steps kiye. Baseline ke baghair aap ke paas ek improvement hai jis pe yaqeen hai lekin bata nahi sakte.

• Naya workflow purane ke sath parallel run karo 2-3 cycles ke liye, evidence aur ek fallback milta hai
• Cost bata sako: usage cost + baaki insaan ka waqt (jo gates aap ne rakhe), saving = baseline × frequency × log, ek sentence jis pe manager act kar sake

Part 6 · Operate Karo, Explain Karo

17. Jab Bad Output Bahar Nikal Jaye

Ek gate ek busy Friday pe skip ho jati hai, configuration stale ho jati hai. Design karne wale ne pehle se ye 4 cheezein decide karni chahiye:

1. 1

   1. Ye Kaise Ruke

   Sabse tez tareeqa workflow rokne ka kya hai, kaun rok sakta hai, kya unhe permission chahiye?

2. 2

   2. Kya Already Bahar Ja Chuka Hai

   Ek output nahi, kitne outputs affect huay, kahan gaye?

3. 3

   3. Kisay Batana Hai, Kab Tak

   Customer, internal owner, risk/compliance function, ye pehle se decide karo

4. 4

   4. Map Mein Kya Badalta Hai

   Har incident design ke baare mein diagnostic data hai, map ko date ke sath update karo

18. Bolo Ye Kya Karta Hai, Phir Gate Ka Naam Lo

3 phrases chupke se overstate karti hain:

• "Fully automated" — almost kabhi sach nahi, pehla visible error isay public expose karta hai
• "AI handles X" — sentence se human gate hi gayab kar deta hai
• "It's basically as good as a person" — ek standard set karta hai jo tool aakhir kaar miss karega

Same workflow, 3 audiences ko 3 tareeke se:

“Stakeholders ek AI workflow pe zyada trust karte hain, kam nahi, jab human checkpoints explicit hoon.”

Poora Course, Compressed

“Output verdict nahi hai. Step ko us se judge karo jo wo cost karta hai aur jiska jawab dena hai, failure ko us se judge karo jab wo shuru hui.”

• Unit step hai, session nahi, personal habit alag cheez hai team workflow se
• Requirements ko structure mein nikalo, source ke sath, phir apni hi list pe ambiguity dhoondo
• Jo result matter karta hai use compute karo, kabhi likho mat, window test se decide karo kaunsa step aapka hai
• Har step 3 decisions leta hai: eligibility, ownership (reversibility/stakes/accountability), implementation, teesra doosre mein mix mat karo
• 3 mapping errors: halo delegation, unstaffed gate, mapping the tool instead of work, sab true statements se bantay hain, galat step pe
• Map ka ek named owner aur review date honi chahiye, scheduled read hi silent decay ka ilaj hai
• 4 causes timing se pehchano: under-specification, context overload, wrong feature/model, stale configuration, diagnostic sequence sasta pehle chalta hai
• Reaction ko instruction mein badlo (lever naam lo), fix ko rule/reference/procedure mein promote karo
• Optimize karne se pehle baseline measure karo, sahi metric chuno, naye workflow ko parallel run karo
• Bad output nikalne se pehle 4 cheezein decide karo: stop, scope, tell, map change, aur audience ke hisab se human gate ka naam hamesha lo

Ab Khud Try Karo: 5 Prompts

Sab ek sitting mein karne ki zaroorat nahi.

1. 1

   1. Extract Karo, Phir Attack Karo

   Ek real document se har requirement table mein nikalwao (label, source section, direct/implied), phir usi list ko dobara review karwao ambiguity ke liye.

2. 2

   2. Wo Step Dhoondo Jo Undo Nahi Ho Sakta

   Apna workflow do, sirf 3 sawal poochho: undo ho sakta hai? bad-case cost kya hai? ye decision hai ya input? Recommendation abhi mat maango.

3. 3

   3. Timing Se Diagnose Karo

   Agla disappointing output, fix karne se pehle: describe karo kya hua aur kab, AI se ek-ek sawal poochwao jab tak diagnosis na mile.

4. 4

   4. Ek Fix Promote Karo

   Ek correction dhoondo jo 2 se zyada dafa ki hai, poocho ye rule hai, reference hai, ya procedure, phir sahi format mein likhwao.

5. 5

   5. 3 Tareeke Se Bolo

   Apna workflow describe karo, phir technical, executive, aur risk audience ke liye 3 descriptions likhwao, har ek mein human checkpoint naam lo.

Ek Project: Map, Break, Fix, Explain

Ek workflow, 6 hisse, roughly ek ghanta.

1. 1

   1. Map Karo

   Real workflow ko table mein likho: step, owner, reason, har reason reversibility/stakes/accountability naam le.

2. 2

   2. Features Rakho

   Har AI step ke liye batao kaun carry karta hai (Skill, knowledge, code execution, standing instruction), har human step ke liye person aur moment naam lo.

3. 3

   3. Ek Step Real Mein Chalao

   Sabse mechanical AI step pe real input chalao.

4. 4

   4. Jaan Bujh Ke Todo

   Ek figure prose mein maango jo compute honi chahiye thi, jab plausible galat number aaye, fix karne se pehle Concept 11 se diagnose karo.

5. 5

   5. Ek Fix Promote Karo

   Step 3/4 mein jo correct kiya wo sahi home mein rakho.

6. 6

   6. Baseline Measure Karo

   Ek unchanged run: time, revision rounds, manual steps.

7. 7

   7. One-Page Map Bharo

   Owner, next review date, aur "agar galat ho jaye" wali 4 lines ke sath.

8. 8

   8. 3 Descriptions Likho

   Technical, executive, risk, har ek mein gate ka naam.

One-Page Map: Copy Karo

DELEGATION MAP: [workflow name]
Owner: [ek named person]        Last reviewed: [date]      Next review: [date]

ELIGIBILITY (dono pass hone chahiye)
Data gate: [kya ye data is tool mein ja sakta hai? kisne confirm kiya?]
Input gate: [kya har step ki har zaroorat supply ho sakti hai?]

STEPS
#  Step          Owner          Criterion              Carried by
1  ...           AI             reversible, low stakes  Skill
2  ...           AI             reversible, checked     code execution
3  ...           Collaborative  high stakes, human judges  Skill + gate
4  ...           Human          ye decision hai         named reviewer
5  ...           Human          irreversible            named signer

GATES (har collaborative/human step ke liye ek line)
Step 3: review [naam], [kab], check kar raha [kya specifically]
Step 4: approve [naam], se pehle [aage kya hota hai]

BASELINE (change se pehle measured)
Time per run: ___   Revision rounds: ___   Manual steps: ___
Metric jo ye workflow optimize karta hai: [time | consistency | accuracy | variance]

AGAR GALAT HO JAYE
Roka kisne: [kaun, aur bina permission act kar sakte hain?]
Scope check: [kaise pata chalega kitne outputs affect huay?]
Bataya kisay: [kaun, kitne waqt mein]
Map change: [yahan record, date ke sath]

Is Chapter Ke Naye Terms

Exam ke liye ye poori glossary yaad rakho, koi bhi term skip mat karo:

Self-Test

Khud Se Poocho

Pehle khud answer do, phir sawal pe click kar ke answer check karo.