Skip to main content

Paperclip ke Saath Workforce Banana: 90-Minute Crash Course

6 Scenarios, Zero se Managed AI Workforce Tak

Paperclip, OpenClaw ke employee ke muqable mein company hai. OpenClaw crash course ne aap ko aik AI Worker diya tha jo aap ke laptop par rehta hai; yeh course aap ko woh company deta hai jo aise Workers ki fleet ko hire, govern, aur audit karti hai. Aik Worker aik function hai; workforce Workers ka org chart hai jo budgets ke against, approvals ke under, audit trail chhor kar chalta hai. Paperclip usi Worker ke gird company hai.

Paperclip naya hai aur tez move kar raha hai: 2026 ke darmiyan tak tens of thousands of GitHub stars, MIT-licensed, self-hosted, aur aik command se poori tarah aap ke laptop par chalne wala tool; na account, na credit card, na cloud setup. OpenClaw ne Worker ke liye jo friction-free property hasil ki, Paperclip wahi property us Worker ke gird company ke liye hasil karta hai.

In ninety minutes ke end tak aap ki machine par aik real AI-native company chal rahi hogi: defined purpose wali company entity, adapter ke against hire kiya gaya Worker jo work receive karne ke liye ready hai, aik inbound issue jo keyless local Worker ko assigned ho kar completion tak gaya, aik approval jo aap yani board ke paas file hua aur permanent audited record ke saath decide hua, aik real LLM-backed Worker taake budget ke paas enforce karne ke liye data ho, aur aik SQL query jo milliseconds mein sab kuch reconstruct kar deti hai. Yeh Friday tak bhool jane wala demo nahin; yeh woh substrate hai jise aap kal dobara touch karein ge.

Yeh crash course kaise kaam karta hai. Aap aik tiny folder download karte hain, usay apne coding agent (Claude Code ya OpenCode) ko dete hain, aur chhe scenarios se guzarte hain. Agent folder parhta hai, Paperclip install karta hai, pehli company set up karta hai, pehla Worker hire karta hai, pehla issue assign karta hai, pehli approval file karta hai, keyless stub Worker ko real LLM Worker se badalta hai taake budget kuch meter kar sake, aur activity log query kar ke CFO-grade audit question ka answer deta hai. Aap steer karte hain; agent kaam karta hai; Paperclip woh management plane ban jata hai jis ke through aap ki workforce chalti hai.

Reading path - prerequisites - deep version (kholne ke liye click karein)

Reading path (chhe scenarios + aik monthly habit):

  1. Paperclip khara karein aur apni company define karein. onboarding, dashboard up, aik company configured. ~15 min.
  2. Apna pehla Worker hire karein. keyless local Worker (worker-stub.py, download mein shamil) http adapter par registered, heartbeats receive karne ke liye ready. ~15 min.
  3. Company ko us ka pehla real issue bhejein. issue create hota hai, create time par Worker ko assign hota hai, next heartbeat par pick up hota hai, aur resolve ho jata hai. ~15 min.
  4. Risky action ke liye approval file karein. $750 refund ko board sign-off chahiye; aap request file karte hain, decision dete hain, aur audited record parhte hain. ~15 min.
  5. Real LLM Worker add karein; budget ko waqai effective banayein. keyless stub billable work nahin karta, is liye budget move nahin karta. real Gemini-backed Worker add karein aur spend accrue hota dekhein. ~15 min.
  6. CFO ki tarah audit trail query karein. aik SQL query, poori company history, milliseconds mein. ~10 min.
  7. (Mahine mein aik dafa, aaj nahin) workforce audit chalayein. waqt aaye to ~10 min.

Har scenario runnable success par khatam hota hai. Agar ninety minutes aik sitting mein zyada lagein to inhein separate sittings mein karein; har cheez persist karti hai. Scenarios 3 se 6, Scenario 1 aur 2 ki company aur Worker par build karte hain.

Prerequisites (yeh page yeh cheezein assume karta hai):

  1. Claude Code ya OpenCode installed. Dono mein se koi bhi kaafi hai. Agar dono nahin hain to pehle Agentic Coding Crash Course karein.
  2. Node.js 20 ya later. terminal mein node --version chalayein. Agar v20.0.0 se neeche hai to nodejs.org/en/download se LTS install karein; poochein to aap ka coding agent aap ko walkthrough de dega.
  3. Python 3. python3 --version chalayein. Scenario 2 mein jis Worker ko hire karein ge woh aik small Python file hai (worker-stub.py, lagbhag 120 lines) jo download mein shamil hai aur sirf standard library use karti hai: pip install ki zaroorat nahin. macOS aur zyada tar Linux mein Python 3 pehle se hota hai; Windows par coding agent install mein madad dega.
  4. Sirf Scenario 5 ke liye free Gemini API key. Scenarios 1 se 4 aur Scenario 6 bilkul API key ke baghair chalte hain. Scenario 5 keyless stub ko real LLM Worker se badalta hai, aur usay key chahiye. Google ka free tier kaafi hai; yahi key OpenClaw crash course use karta hai. Shuru karne se pehle ya Scenario 5 par pohanch kar aistudio.google.com se le lein.
  5. Aap AI Workers par previous crash courses mein se kam az kam aik kar chuke hon, ideally OpenClaw ya Digital FTE. Paperclip ka point Workers ki workforce manage karna hai; hiring shuru karne se pehle aap ko maloom hona chahiye Worker kya hai.

Patient version chahiye? Is track mein is course ke dono taraf do companion crash courses hain. From Digital FTE to Production Worker aik single Worker ko Inngest ke durability envelope mein wrap karta hai: workforce se bilkul pehle ka step. From Fixed to Dynamic Workforce hiring ko callable capability banata hai: is course ke foran baad ka step. Yeh page plus AGENTS.md brief management plane ko end-to-end cover karte hain; woh dono courses is par layer hote hain, isay replace nahin karte.

Collaboration Pattern​

Aik Worker aik function hai: aik queue, aik job, durable magar governed nahin. Workforce aik organization hai: Paperclip ka management plane company, adapters ke against hired Workers, create time par directly assigned issues, approvals as decision records, per-Worker budgets, aur activity log rakhta hai.

Is page par teen actors hain: aap, aap ka coding agent, aur Paperclip. Aap prompts paste karte hain aur woh calls karte hain jo sirf human kar sakta hai. Aap ka coding agent Paperclip CLI aur API chalata hai, logs dekhta hai, aur failures se recover karta hai. Paperclip management plane hai: yeh company rakhta hai, Workers hire karta hai, heartbeat par un tak work pohanchata hai, aur har action record karta hai.

Is page par teen actors hain: aap, aap ka coding agent, aur Paperclip. Aap prompts paste aur actions approve karte hain; coding agent Paperclip install aur configure karta hai; Paperclip heartbeats par work Workers tak le jata hai aur har action record karta hai.

Har scenario yahi five-step rhythm use karta hai:

  1. Aap apne coding agent mein aik sentence paste karte hain. Yeh brief hai, script nahin. Aap outcome describe karte hain; steps enumerate nahin karte.
  2. Aap ka agent AGENTS.md consult karta hai (yeh pehle hi context mein hai: folder mein CLAUDE.md session start par isay automatically import karta hai, is liye fetch step nahin) aur plan propose karta hai. Woh commands name karta hai jo chalana chahta hai, decision points flag karta hai, aur pehle destructive command se pehle poochta hai.
  3. Aap approve karte aur watch karte hain. agent install commands chalata hai, API hit karta hai, server log dekhta hai, aur aap ko dikhata hai kya nazar aa raha hai. known failure mile to brief se pattern recognize kar ke documented fix apply karta hai.
  4. Aap ka agent seam par ruk jata hai. Kuch moves sirf aap kar sakte hain: dashboard URL kholna, decide karna ke $750 refund grant hona chahiye ya nahin, API key export karna. agent seam name karta hai aur wait karta hai.
  5. Aap done tab hote hain jab aik observable cheez ho jaye. activity log mein nai row aa jaye, issue done par move ho jaye, SQL query expected history return kar de. Har scenario batata hai kis cheez ko watch karna hai.

Bas yahi hai. Agent woh kaam karta hai jis mein agent acha hai: install, configure, debug, query, recover. Aap woh karte hain jo sirf aap kar sakte hain: decide, approve, aur un cheezon par action lena jo aap ke accounts se tied hain. Yahi rhythm, goal describe karna, plan lena, approve karna, har step par verification ke saath execute karna, AI Prompting in 2026 crash course mein sikhaye gaye prompting pattern jaisa hai; neeche har scenario do short paste prompts use karta hai, instructions ki wall nahin, taake aap rhythm experience karein.

Har scenario wahi five-step rhythm use karta hai: aap aik sentence paste karte hain; agent AGENTS.md parh kar plan propose karta hai; aap approve karte hain; agent execute aur verify karta hai; scenario tab done hota hai jab aik observable cheez ho jaye.

Poore crash course ke liye aik recovery move

Agar kisi bhi point par kuch sideways jaye to aap ko CLI commands ya error codes janne ki zaroorat nahin. Yeh apne agent ko paste karein:

Kuch kaam nahin kar raha. paperclipai doctor chalayein, phir sab se recent Paperclip server log parhein, seedhi zabaan mein batayein kya nazar aa raha hai, aur aisa fix propose karein jise main approve kar sakun.

Aap ka agent diagnostic chalata hai, log parhta hai, jo dekhta hai usay name karta hai, aur fix propose karta hai. Aap approve karte hain. Yahi is page ke har scenario ka recovery loop hai.

Agar koi scenario zyada waqt le

Har scenario ka budgeted time H2 mein shown hai. Agar aap us budget ke 2x se aage nikal jayein, masalan 15-minute scenario par 30 minutes se zyada, agent ko wapas khenchein aur paste karein: "What's blocking us, in one sentence? Let's re-plan from there." Budget se aage spin karna aam tor par agent ke improvising ka signal hai; plan par re-anchor karna isay fix karta hai.

Jo folder aap download karein ge us mein exactly three files hain: AGENTS.md (Paperclip work karne wale kisi bhi coding agent ke liye compact operational reference), CLAUDE.md (aik line, @AGENTS.md, jo Claude Code ko brief automatically import karne ko kehti hai), aur worker-stub.py (woh keyless local Worker jisay aap Scenario 2 mein hire karein ge). Yahi poori environment hai.

Download paperclip-crash-course.zip

Kahin bhi unzip karein. unzipped folder mein terminal kholein. Apna coding agent launch karein:

cd paperclip-crash-course
claude

Ab aap ke agent ke paas brief loaded hai. Hum six scenarios aik aik kar ke chalayein ge; agla scenario shuru hone se pehle har scenario runnable success par khatam hoga. Yeh brief capable coding agent assume karta hai: Claude Code, ya OpenCode current frontier model ke saath. older ya smaller models drift karein ge, sab se zyada Scenario 2 ke Worker-create body ke precise JSON aur Scenario 6 ke SQL par; agar Scenario 1 mein agent ka first plan aap ki machine ke liye specific hone ke bajaye vague ya generic lage, to aage jane se pehle stronger model par switch karein.

Scenario 1 Se Pehle: Confirm Karein Agent Ne Brief Load Kar Liya Hai (~30 sec)​

Aik paste bata dega ke CLAUDE.md ne apna kaam kiya aur AGENTS.md agent context mein aa gaya:

Aap Paperclip ke liye kya kar sakte hain?

Agar reply specific Paperclip work name kare, pre-install probe, company-and-Worker shape, heartbeat contract, create-time issue assignment, approval payload, budget reality, activity-log queries, monthly audit habit, to aap loaded hain aur Scenario 1 ke liye ready hain. Agar yeh generic AI capability talk lage jis mein Paperclip-specific details na hon, import nahin hua: agent band karein, confirm karein ke aap unzipped paperclip-crash-course/ folder ke andar hain, aur relaunch karein.

AGENTS.md mein asal mein kya hai (woh file jo agent abhi parh raha hai)

Aap ko yeh file khud kabhi parhne ki zaroorat nahin; point yahi hai. Lekin is ki shape janne se behtar questions pooche ja sakte hain, masalan "approvals section walk me through" kaam karta hai kyun ke section maujood hai. brief order mein yeh cover karta hai:

PART 1 :: PRINCIPLES (apply everywhere)
  Versions this brief was verified against
  Source of truth, in order         <- live docs > this brief > the running install
  Critical: discover before you act  <- table of intent -> where to confirm
  Working pattern (every task)       <- read -> propose -> ask -> execute -> verify
  Past tense is for completed actions only
  Trust progression                  <- ask each, then blanket; re-acquire on anomaly
  Safety rails (non-negotiable)
  Secrets discipline
  Sourcing claims that exist only in this brief

PART 2 :: OPERATIONS (verified against a live install)
  Install and onboard            <- the pre-install probe + the onboard flow
  Configure                      <- second instance, keeping the server alive
  Companies                      <- create, goals, projects
  Agents (Workers) and adapters  <- the real adapter list + the verified create body
  The heartbeat contract         <- what the http adapter actually POSTs
  Issues and assignment          <- no routing engine; assign at create time
  Approvals                      <- the payload + why it's a decision record
  Budgets                        <- why they only move for LLM Workers
  Audit trail                    <- activity_log + the real connection string
  Diagnose and recover           <- the most common failures
  When you don't know what to do <- three-layer fallback

Baad mein AGENTS.md ka koi section relevant lage to action se pehle agent se us ka walkthrough maang sakte hain, masalan "walk me through the Approvals section of AGENTS.md before we file that request". brief is tarah likha gaya hai ke agent is se self-direct kar sake.

Lab path par note: poori pipeline keyless chalayein

Scenario 2 mein jis Worker ko hire karein ge woh worker-stub.py hai, keyless local Worker jo download mein shamil hai. Yeh Paperclip ke http adapter par chalta hai: har heartbeat par Paperclip poora issue stub ko POST karta hai, aur stub disposition wapas post karta hai. Na LLM, na API key, na second tool. Scenarios 1 se 4 aur Scenario 6 mein aap poora management plane, heartbeats, assignment, approvals, audit, end-to-end unhi cheezon par chalate hain jo pehle se hain. Phir Scenario 5 mein aap aik cheez add karte hain, free Gemini key, deliberate reason ke liye: keyless stub billable work nahin karta, is liye budget ke paas meter karne ko kuch nahin. real LLM Worker budget ko enforce karne ke liye signal deta hai. pipeline shape dono suraton mein identical hai; keyless stub bas drafted reply ke bajaye stub disposition produce karta hai.

Scenario 1: Paperclip Khara Karein Aur Apni Company Define Karein (~15 min)​

Goal: Paperclip aap ke laptop par running ho, aik company configured ho ("Acme Customer Support" ya jo naam aap pick karein), dashboard browser mein reachable ho. Do prompts: plan maangein, phir approve aur execute karein.

1a. Onboard Karein Aur Install Verify Karein​

First prompt: jo chahiye woh describe karein aur plan maangein.

Mein Paperclip apne laptop par chalana chahta hun aur Acme Customer Support naam ki pehli company walkthrough karna chahta hun. Kuch bhi touch karne se pehle apne brief wala pre-install probe chalayein aur plain language mein plan samjhayein: kya check kiya, kya change karein ge, aur kahan mujhe step in karna hoga.

agent AGENTS.md parhta hai, pre-install probe chalata hai, Node version, prior installs, port availability, prior data directory, aur running Paperclip processes check karta hai, phir plan propose karta hai. Agar stale state ho, jaise prior paperclipai install ya default port pakarne wala process, to touch karne se pehle decision ke liye pause karega. plan parhein. Kuch off lage to push back karein; "why are you doing that?" poochein, agent explain ya adjust karega.

Second prompt: approve karein aur run hone dein.

Plan thik hai. step by step aage barhein, har step par batayein kya nazar aa raha hai. Kisi destructive command se pehle pause karein. onboarding finish ho to dashboard URL batayein taake mein browser mein confirm kar sakun; kuch setups mein yeh auto-open ho jata hai.

agent brief wala onboarding command chalata hai, output watch karta hai, aur script ke operational values capture karta hai: API host, dashboard URL, data-directory path, embedded Postgres port, config-file path. Yeh values aik project-local file mein save hoti hain jo aap control karte hain; chat mein secrets echo nahin hote. Aik detail brief agent ko durust karwata hai: onboard banner API line ko /api suffix ke saath label karta hai, magar agent bare host capture karta hai taake later API calls mein doubled /api/api/ path na bane. default loopback bind mode mein bootstrap API key issue nahin hoti; trust loopback-scoped hai, aur poochein to agent yahi kahega. onboarding finish hote hi dashboard aam tor par default browser mein auto-open ho jata hai.

1a done when: agent report kare ke API server listening hai, aap browser mein dashboard URL open kar sakte hain, aur dashboard errors ke baghair render hota hai.

1b. Apni Pehli Company Define Karein​

Yeh agent ko paste karein:

Ab company define karein. Naam Acme Customer Support rakhein. purpose: "Respond to customer inquiries within 4 hours, with refund decisions made consistently and within policy." Aik goal add karein: "Reduce average response time to under 4 hours," aur aik project: "Inbound email triage." Pehle live API se exact field names check karein, phir create karein. done hone par mujhe dashboard mein visible cheezein dikhayein.

agent brief ka Companies section parhta hai, running API se field names confirm karta hai, company, goal, project create karta hai, phir dashboard confirm karwata hai. brief note karta hai ke company ka purpose description field mein jata hai, aur unknown field name loudly fail nahin karta balkay silently miss ho sakta hai.

Scenario 1 done when: dashboard mein "Acme Customer Support" naam ki aik company, aik goal, aur aik project visible hon, aur activity log mein matching rows hon, jaise company.created, goal.created, project.created.

Scenario 2 mein carry-forward

Paperclip ne jo company ID assign ki hai woh ab agent ki created project-local file mein hai. later scenarios mein agent usi file se read karega, is liye aap ko ID yaad rakhne ki zaroorat nahin. agent ko unzipped folder mein running rakhein taake working directory mein file available rahe.

Scenario 2: Apna Pehla Worker Hire Karein (~15 min)​

Concept. Paperclip mein Worker configured role hai: name, runtime, permissions, budget, heartbeat schedule. Paperclip ki API inhein "agents" kehti hai; yeh page poori jagah "Worker" kehta hai taake woh coding agent se confuse na ho jis mein aap prompts paste kar rahe hain. runtime aik adapter ke through plug hota hai. Paperclip kai adapters ship karta hai: LLM runtimes (claude_local, codex_local, gemini_local, aur more), openclaw_gateway, process adapter jo har heartbeat par command chalata hai, aur http adapter jo har heartbeat aap ke control wale URL ko POST karta hai. Aap ka agent aap ke install ke against aik API call se current set list kar sakta hai.

Bring-your-own-agent heartbeat contract. Paperclip assigned issue ke saath Worker URL par heartbeat POST karta hai; authority envelope aur budget server-side rehte hain. Worker acknowledge karta hai aur disposition alag post karta hai.

Is crash course mein aap apna pehla Worker http adapter par hire karein ge, pointed at worker-stub.py, woh keyless local Worker jo download mein shamil hai. Yeh standard library use karne wali small Python file hai. Jab bhi Paperclip isay work assign karta hai, poora issue stub ko POST karta hai; stub issue parhta hai aur disposition wapas post karta hai, jis se issue done par move ho jata hai. Na LLM, na API key. pipeline shape, heartbeat, assignment, approval, audit, real LLM Worker jaisi hi hai; work-product stub disposition hai, drafted reply nahin. Scenario 5 mein aap isay real LLM-backed Worker se badalte hain.

Authority par note. Worker hire karte waqt aap plain language mein capabilities field mein likhein ge ke woh kya kar sakta hai: CRM records parhna, replies draft karna, $50 se oopar refunds aur outbound external email ko board approval chahiye. Isay Worker ka authority envelope samjhein. Paperclip 2026 release mein yeh envelope structured per-field runtime enforcement nahin, balkay aap ke prose description ke tor par store hota hai; real enforcement seam woh approval gate hai jo Scenario 4 mein wire hoga. mental model phir bhi sahi hai: jab task is line ko cross kare to action ke bajaye approval request banani chahiye.

Authority envelope yeh sochne ka mental model hai ke Worker kya kar sakta hai; yeh Paperclip schema nahin. Real enforcement seam approval gate hai.

Aap role define kar ke API ke through register karte hain. Do prompts: stub start aur role draft, phir register aur verify.

2a. Stub Worker Start Karein Aur Role Draft Karein​

First prompt: orient, phir draft.

Mein Tier-1 Customer Support naam ka aik Worker hire karna chahta hun, http adapter par, is folder mein aaye hue worker-stub.py ki taraf pointed. Pehle mujhe walk through karein ke worker-stub.py kya karta hai aur aap isay kaise start karein ge. Phir register karne se pehle brief wala agent-create request body dikhayein: role name, adapter aur config, capabilities text, permissions, budget, aur heartbeat schedule. monthly budget jaan boojh kar chota rakhein; Scenario 5 real Worker se dikhaye ga ke budget kya meter karta hai.

agent brief ka Agents section parhta hai, worker-stub.py explain karta hai, tiny HTTP server jo Paperclip ka heartbeat POST sunta hai, payload se issue parhta hai, aur done disposition PATCH karta hai. Phir batata hai isay kaise start karega: aik python3 command, listening port aur Paperclip API URL arguments ke saath. Phir agent-create body draft kar ke structure dikhata hai: name, adapterType (http) aur adapterConfig (stub URL), capabilities, permissions, budgetMonthlyCents, aur runtimeConfig ke under heartbeat schedule. field wrong lage to poochein; agent live API se confirm karega.

Second prompt: start, register, verify.

Theek lag raha hai. stub Worker start karein, role ko Paperclip mein register karein, aur confirm karein ke dashboard mein nazar aa raha hai. Mujhe Paperclip ka assigned agent ID, next heartbeat time, aur stub ki apni log file dikhayein taake mein dekh sakun ke yeh listening hai.

agent worker-stub.py start karta hai, background mein running rakhta hai, har heartbeat log karta hai, role API ke through register karta hai, assigned agent ID project-local file mein capture karta hai, aur dashboard refresh karne ko kehta hai.

Scenario 2 done when: dashboard mein "Tier-1 Customer Support" naam ka aik Worker ho, worker-stub.py running ho, log file maujood ho, aur activity log mein Worker-creation row ho, jaise agent.created. Pehle se jaan lein: Paperclip Workers creation ke baad immutable hain. adapter ya budget edit nahin kar sakte; naya Worker hire karna parta hai. Isi liye Scenario 5 is Worker ko upgrade nahin karta balkay fresh LLM Worker hire karta hai.

Scenario 3 mein carry-forward

Worker registered hai aur worker-stub.py running hai, magar Worker ki issue queue empty hai. next scenario isay real work deta hai. stub process running rakhein; band kiya to heartbeats dead URL par jayein ge aur issues resolve nahin honge. Agar din ke liye stop karein to Scenario 3 resume karte waqt agent stub restart kar sakta hai.

Scenario 3: Company Ko Pehla Real Issue Bhejein (~15 min)​

Concept. company maujood hai, Worker maujood hai, magar kuch assigned nahin. production mein customer email event create kar ke Paperclip issue banaye ga. Paperclip 2026 mein rules-based routing engine nahin: "agar issue X match kare to Worker Y ko bhej do" layer nahin. Assignment direct hai, aur kab karte hain yeh matter karta hai. assignee ke saath create time par bana issue heartbeat orchestration mein wired hota hai: yeh ready born hota hai, next heartbeat isay pick karta hai. baghair assignee create kar ke baad mein assign kiya issue isi tarah pick up nahin hota. move yeh hai: issue create karein aur same step mein Worker name dein.

Agla interesting part hai: Worker ka next heartbeat fire hota hai, Paperclip poora issue worker-stub.py ko POST karta hai, stub parh kar done disposition post karta hai, aur issue todo se in_progress se done tak move hota hai, har step activity-log row ke saath.

Do prompts: issue create/assign karein, phir Worker ko pick up karte dekhein.

3a. Aik Inbound Issue Create Karein, Create Time Par Assigned​

Yeh agent ko paste karein:

"Inbound email triage" project ke liye aik inbound issue create karein: customer C-4429 damaged product par $30 refund ke bare mein pooch raha hai. description ko realistic two-sentence customer email banayein. Isay Tier-1 Customer Support ko create time par assign karein, same command mein, follow-up edit ke tor par nahin. create hone par mujhe Paperclip ka assigned issue ID aur current status dikhayein.

agent brief ka Issues section parhta hai, create command mein Worker ko assignee name kar ke issue create karta hai, aur issue identifier dikhata hai, jaise ACM-1.

3a done when: dashboard issue queue mein Tier-1 Customer Support ko assigned aik issue, status todo, aur activity log mein issue.created row ho.

3b. Worker Ko Pick Up Karte Dekhein​

Is prompt se pehle heads-up: Worker mein wake on assignment set hai, is liye create time par assign hote hi heartbeat fire ho sakta hai. Jab tak aap yeh parhein, issue already done ho sakta hai. Yeh system working hai, missed step nahin. Yeh prompt "make it happen" nahin; yeh "record mein dikhao kya already hua" hai.

Yeh agent ko paste karein:

Is issue ke saath already kya hua, mujhe dikhayein. worker-stub.py ki apni log file aur issue ki activity-log rows kholein, aur sequence walk through karein: assignment par heartbeat fire hona, stub ka issue receive karna, stub ka disposition back post karna, aur issue ka terminal status tak pohanchna. Mujhe batayein kaun se fields prove karte hain ke work waqai hua, sirf issue ko done flip nahin kiya gaya. Agar kisi wajah se issue abhi done nahin to Tier-1 ke liye heartbeat manually fire karein aur phir yahi sequence walk through karein.

agent dono records parhta hai aur already-run sequence narrate karta hai: Worker assignment par wake hua, Paperclip ne issue stub ko POST kiya, stub ne done disposition PATCH kiya. issue done par hai, startedAt aur completedAt timestamps set hain. activity log run record karta hai. Agar heartbeat abhi fire na hua ho to brief on-demand heartbeat command bhi rakhta hai.

Scenario 3 done when: issue todo se in_progress se done tak move ho, activity log transition dikhaye, aur worker-stub.py log received heartbeat aur posted disposition dikhaye. Is issue par koi cost record nahin, jo durust hai; Scenario 5 batata hai cost data kahan se start hota hai.

Scenario 4 mein carry-forward

Worker ke paas aik issue complete ho chuka hai, is liye approval step add karte waqt activity log empty nahin. Scenario 4 jaan boojh kar aise issue se kaam karta hai jis ka action, bara refund, Tier-1 Worker ke akelay decide karne se oopar hai, taake board sign-off file aur record karna dikhaya ja sake.

Scenario 4: Risky Action Ke Liye Approval File Karein (~15 min)​

Concept. Kuch actions Tier-1 Worker ke pay grade se oopar hote hain: $750 refund, contract change, ya authority envelope cross karne wali koi bhi cheez. Paperclip ka answer approval hai: tracked board decision record. real LLM Worker, $750 refund par reasoning karte hue, authority exceed recognize kar ke action ke bajaye approval request file karega. keyless stub reason nahin karta, is liye is scenario mein aap board ka hand bante hain: woh approval request file karte hain jo Worker file karta, phir isay decide karte hain.

Do baatein saaf rakhein:

  • Approval decision record hai, state machine nahin. approval file karna Worker suspend nahin karta; approve karna automatically Worker resume, linked issue status change, ya approved action execute nahin karta. decision par act karna alag explicit step hai. Paperclip durable audited record deta hai: kis ne kya decide kiya, kis rationale ke saath, kab. Yahi record auditor chhe months baad parhta hai.
  • Teachable moment audit trail hai, automated unblock nahin. issue ke khud move hone ke bajaye activity log row watch karein.

Approval tracked board decision record hai, state machine nahin. Approval file aur decide karna human decision aur timestamp record karta hai; yeh Worker ko automatically resume ya issue ko change nahin karta.

Do prompts: request file karein, phir decide karein.

4a. Approval Request File Karein​

Yeh agent ko paste karein:

Aik inbound issue create karein: customer C-1138 teen mahine pehle kharide gaye product par $750 refund maang raha hai; kehta hai product defective aya aur photos maujood hain. Isay abhi Worker ko assign na karein: yeh action board sign-off maangta hai. Phir aise act karein jaise Worker apni authority se oopar action dekh kar karta: is issue se linked board approval request file karein. board decision ke liye approval type use karein. payload mein action ("issue a $750 refund"), rationale (customer history, damage photos, documented defects policy), aur do teen alternatives considered (partial refund, store credit, no refund) daalein. approval ko pending status mein dikhayein, aur confirm karein ke issue khud abhi untouched hai.

agent issue without assignee create karta hai, is liye stub Worker isay pick up nahin karta; issue default starting status par baitha rehta hai. Phir company approvals API se approval request file karta hai: board-decision type, free-form payload jis mein action, rationale, alternatives verbatim, aur issue link. activity log mein approval.created row aati hai.

4a done when: issue exists, unassigned, untouched default status par; dashboard approvals queue mein aik pending entry; agent decision aap ko hand off kar ke stop kare.

4b. Isay Decide Karein​

Yeh hissa aap ka hai. request parhein: amount, rationale, alternatives. decide karein.

decision ke baad yeh agent ko paste karein:

Mein ne is refund ko approve karne ka decide kiya hai (ya: reject karna hai; aik line mein batata hun kaun sa aur kyun). Mera decision record karein. Phir is approval ki activity-log rows walk through karein aur woh field dikhayein jo prove karti hai ke mein, human board, decider tha, agent nahin. Aur saaf remind karein ke kya automatically nahin hua: issue abhi wahin baitha hai, aur is decision par actually act karne ke liye next kya karna hoga.

agent decision record karta hai, approval approved ya rejected ho jati hai, decider identity aur decision timestamp stamp hote hain. activity log mein approval.approved ya approval.rejected row aati hai, actor_type set to user, aur board identify karne wala actor_id. Phir honest part name karta hai: linked issue khud status change nahin hua. issue close karna ya refund "issue" karna, simulated because payment processor wired nahin, separate command hai.

Scenario 4 done when: activity log approval chain rakhta ho, request filed, decision recorded by you; linked issue decision se visibly unchanged ho; aap woh field name kar sakein jo human decider ko agent action se distinguish karta hai (actor_type); aur apne words mein bata sakein approval decision record hai, automatic unblock nahin.

Scenario 5 mein carry-forward

Ab aap ke paas small but real activity log hai: two issues, one approval, aur in ke behind heartbeat runs. Aik table abhi empty hai: cost data nahin. Scenario 5 mein yeh change hota hai, aur yahi woh scenario hai jisay API key chahiye.

Scenario 5: Real LLM Worker Add Karein; Budget Ko Teeth Dein (~15 min)​

Concept. Har Worker ka monthly budget hota hai, hire karte waqt cents mein set. Magar budget ke paas enforce karne ko tabhi kuch hota hai jab Worker billable work kare: keyless stub kuch billable nahin karta, is liye spend zero rehta hai aur budget move nahin karta. Yeh lab ki honest shape hai, bug nahin. budget ko waqai meter karte dekhne ke liye real LLM runtime-backed Worker chahiye.

Is scenario mein aap swap karte hain. Aap gemini_local adapter par new Worker hire karein ge, fresh hire, edit nahin, kyun ke Scenario 2 mein bataya Workers immutable hain. Isay deliberately tiny monthly budget dein ge, phir itna work dein ge ke budget run down ho. Phir dekhein ge spend limit ki taraf climb karte waqt Paperclip actually kya karta hai. Yeh wahid scenario hai jisay API key chahiye: free Gemini key, wohi jo OpenClaw crash course use karta hai. Gemini is page ka path hai kyun ke free tier scenario ko keyless-adjacent rakhta hai; agar aap ke paas kisi aur low-cost runtime, masalan codex_local with cheap OpenAI model, ki key hai to budget mechanic identical hai aur agent adapter accordingly adjust kar sakta hai.

Budget monthly cents value hai, aur yeh tabhi effective hota hai jab Worker billable work kare. Keyless stub Workers zero cost generate karte hain; LLM-runtime Workers per-run usage record karte hain aur monthly spend budget ke against climb karta hai.

Aik prompt LLM Worker hire karne ke liye, aik prompt isay work dene aur budget watch karne ke liye.

5a. LLM Worker Hire Karein​

Yeh agent ko paste karein:

Ab mein real LLM Worker add karna chahta hun. start se pehle mein apni shell mein GEMINI_API_KEY export karun ga; jab aap ko chahiye ho mujhe batayein. gemini_local adapter par Tier-1 Customer Support (LLM) naam ka new Worker hire karein, cheapest current capable Gemini model use karte hue. Isay deliberately small monthly budget dein, aik ya do dollars, taake spend accrue hota dekh sakun. register karne se pehle create body walk through karein, aur remind karein ke hum stub one ko edit karne ke bajaye new Worker kyun hire kar rahe hain.

agent remind karta hai ke Workers immutable hain, is liye new hire, edit nahin. Woh point name karta hai jahan aap key export karein ge, phir agent-create body draft karta hai: Scenario 2 jaisi shape, magar adapterType gemini_local, small budgetMonthlyCents. key export ke baad Worker register karta hai.

5b. Isay Work Dein Aur Budget Dekhein​

Yeh agent ko paste karein:

Ab LLM Worker ko work dein. taqreeban ten short inbound support issues create karein, thori variety ke saath, aur har aik ko create time par new LLM Worker ko assign karein. heartbeats fire karein taake yeh unhein work through kare. Paperclip server log tail karein. Mein do cheezein dekhna chahta hun: per-run cost data exist hona shuru ho (keyless stub ke liye yeh nahin tha), aur Worker ka monthly spend mere set kiye budget ki taraf climb kare. plain language mein exactly batayein Paperclip spend limit approach aur cross karte waqt kya karta hai; yahi behavior mein observe karna chahta hun.

agent issues create/assign karta hai, heartbeats fire karta hai, log tail karta hai. Is dafa Worker billable work karta hai, is liye per-run usage data heartbeat records par land hota hai aur monthly spend zero se move karta hai. agent budget run down hote waqt Paperclip ka actual behavior plainly narrate karta hai: yeh live observation hai, scripted outcome nahin.

Scenario 5 done when: new LLM Worker real issues resolve kar chuka ho, per-run cost data wahan exist kare jahan pehle empty tha, Worker ka monthly spend aap ke set kiye budget ke against move ho, aur agent plain language mein bata chuka ho ke budget limit approach hote waqt Paperclip ne kya kiya.

API key par aik baat

key apni shell mein export karein (export GEMINI_API_KEY=...); isay project file mein paste na karein, aur agent ko bhi file mein likhne na dein. AGENTS.md agent ko yahi kehta hai. Agar key kabhi ghalat jagah paste ho jaye to rotate karein. free-tier key low-stakes hai, magar habit asal point hai.

Scenario 6 mein carry-forward

Ab aap ke paas do Workers hain, keyless stub aur LLM-backed one, dozen-plus issues, one approval, aur is scenario mein new real cost data. Itni history kaafi hai ke Scenario 6 ki audit query parhne ke qabil result de.

Scenario 6: CFO Ki Tarah Audit Trail Query Karein (~10 min)​

Concept. management plane ka point yahi hai ke workforce se bahar koi shakhs, CFO, legal, operations, compliance, database alone se seconds mein reconstruct kar sake kya hua, kisi se pooche baghair. Paperclip yeh history embedded Postgres database mein rakhta hai. Sab se important table activity_log hai: mutating action ke liye one row, company created, issue created, approval decided, Worker hired, heartbeat invoked, actor, action, aur touched entity ke saath. Doosri table cost_events dollar story rakhti hai: provider, model, tokens, billable run ki cost cents mein. Is mein rows sirf LLM Worker ke billable work ke baad aati hain, isi liye Scenario 5 tak empty thi.

activity_log table company ki institutional memory hai: har mutating action par one row. cost_events dollar story rakhta hai. Dono mil kar stakeholders ko milliseconds mein reconstruct karne dete hain ke kya hua.

Aik prompt: Postgres open karein, history query aur cost query run karein, results parhein.

Yeh agent ko paste karein:

Ab CFO ban kar dekhte hain. embedded Paperclip Postgres se connect karein (connection string assemble karne ka tareeqa brief mein hai). Pehle hamari company ke liye activity_log par "what happened, in order" query run karein: har action, actor, touched entity, oldest first. Phir cost_events par cost query run karein: aaj ki total cost dollars mein. Dono ki SQL aur results dikhayein. Aakhir mein Scenario 4 mein mere decide kiye approval ki aik activity_log row walk through karein, aur woh fields point karein jinhein auditor use karega confirm karne ke liye ke decider mein tha.

agent brief ka Audit-trail section parhta hai, install config file se Postgres connection string assemble karta hai, shortcut command nahin; brief tareeqa dikhata hai, phir do queries chalata hai:

  1. What happened, in order: activity_log par SELECT, company ke liye filtered, time ke order mein. Yeh aap ke run ki full spine return karta hai: company/project creation, both Workers hired, every issue created, heartbeats, approval filed/decided. Yeh query LLM Worker ke baghair bhi work karti hai; audit story ka keyless heart yahi hai.
  2. Total cost today: cost_events par SELECT SUM(...), integer cents column ko dollars mein convert kar ke, today filtered. Scenario 5 se pehle nothing; ab LLM Worker runs ki real small cost.

Phir agent Scenario 4 ke approval decision ki aik activity_log row dikhata hai: action (approval.approved ya approval.rejected), actor_type set to user, board identify karne wala actor_id, timestamp. auditor aap ka decision isi tarah reconstruct karta hai.

Scenario 6 done when: history query aap ke whole run ki ordered story return kare, cost query Scenario 5 Worker se real number return kare, aur aap activity_log ka woh column name kar sakein jo human decision ko agent action se alag karta hai: actor_type.

Scenario 7: Aap Ka Monthly Workforce Audit (~10 min/month)​

Concept. company waqt ke saath accumulate karti hai: Workers hired, budgets set, approvals decided, schedules running. Har addition small decision tha jisay aap ne approve kiya; chain compound hoti hai. defense har step par vigilance nahin, kyun ke aap woh nahin pakar sakte jo abhi exist hi nahin; defense fixed cadence par ten-minute review hai. Yeh scenario first ninety minutes ka hissa nahin; yeh move aap company ki baqi life mein mahine mein aik dafa karte hain.

Waqt aane par yeh agent ko paste karein:

Meri Paperclip monthly workforce audit chalayein. last audit ke baad jo kuch hire, configure, schedule, decide, ya pause hua hai, usay walk through karein. Jo cheez mein ne explicitly sign off nahin ki, jo Worker productive work nahin kar raha, jo budget mere set kiye threshold se drift ho gaya, aur jo setting zaroorat se loose hai, sab flag karein. Isay single short report mein summarize karein jisay mein approve ya trim kar sakun.

agent activity log, current Worker roster, budgets, audit tables parhta hai, aur one-page report produce karta hai.

Done when: aap report review karne mein ten minutes laga chuke hon aur kam az kam aik decision kiya ho: Worker authority text tighten, budget prune, unused Worker retire, approval threshold raise. next month ke liye calendar mark karein.

Yeh Kyun Kaam Karta Hai​

Do cheezein fresh rehti hain; aik cheez durable rehti hai.

Fresh #1: is page ke scenarios book site par live hain. agent inhein read karta hai jab aap batate hain kaun sa scenario hai.

Fresh #2: current Paperclip commands aur API shape paperclip.ing/llms.txt par live hai, full docs ka LLM-friendly index, plus patient walkthrough ke liye docs.paperclip.ing. agent kisi bhi non-trivial operation se pehle inhein fresh read karta hai. Paperclip tez ship karta hai; brief isi tarah accurate rehta hai even when individual flags drift.

Durable: download folder mein AGENTS.md hai, Paperclip kya hai, docs kaise navigate karne hain, safety rails, verified operational shapes, recovery patterns; CLAUDE.md, one-line import marker; aur worker-stub.py, keyless Worker itself. AGENTS.md companies, Workers/adapters, heartbeat contract, issues/assignment, approvals, budgets, audit, diagnosis cover karta hai. Yeh page se longer hai kyun ke yeh coding agent se Paperclip ke bare mein kiye jane wale tamam possible work ko cover karta hai, sirf oopar ke six scenarios nahin. folder ki cheezein stale nahin hotin, is liye isay once download kar ke reuse karein.

Pattern jaan boojh kar small hai: aik operational reference, aik import marker, aik runnable Worker. Three files. intelligence files mein nahin; intelligence aap ke coding agent ke inhein read kar ke aap ke next ask par apply karne mein hai. Aap ne six disconnected demos nahin kiye; aap ne aik company assemble ki hai jisay kal touch karein ge.

Ab Asal Mein Kya Chal Raha Hai​

Six demos nahin: aik running system. Scenario 6 ke baad persist hone wali inventory:

ArtifactYeh asal mein kya haiKal kyun matter karta hai
Paperclip serverAPI aur embedded Postgres rakhne wala long-running Node.js processterminal band ya reboot ke baad company survive karti hai
data directorycompany data, encrypted secrets, server logssubstrate; is directory ko dotfiles ki tarah back up karein
one companypurpose, goal, project, aur company-scoped isolationhar Worker, issue, approval ki boundary
two Workerskeyless worker-stub.py Worker aur gemini_local LLM Workeraik pipeline keyless prove karta hai; doosra economics dikhata hai
worker-stub.pyheartbeats sunne wala small local HTTP processaap ka first Worker, aur next keyless Worker ka template
issues and activity logassigned/worked issues plus every action ka row-by-row recordreproducible workforce behavior, queryable history
one approvalfiled aur decided board decision, permanent recordrisky actions ke liye audited sign-off
budget with dataLLM Worker ka monthly budget, real spend ke saath"no runaway spend" ko decorative ke bajaye enforceable banata hai
audit queriesactivity_log aur cost_events par SQLCFO/legal/compliance sawalat fast answer
AGENTS.md and recovery linedurable brief jo agent har session mein parhta haicompany ki lifetime ke liye reusable skill

Is ke saath working day kuch yun lagta hai: real customer message aata hai; integration, jisay aap later wire karein ge, Paperclip issue create karta hai jo Worker ko assigned hai; next heartbeat isay Worker tak le jata hai; Worker reply draft karta hai; high-stakes ones dashboard mein aap decide karte hain; mahine mein aik dafa audit chalta hai.

Agar yeh artifacts baad mein missing ho jayein, laptop wipe, accidental delete, version upgrade gone wrong, to brief plus fresh onboard plus data directory backup restore aap ko isi picture par wapas le aata hai.

real customer channel wire karne se pehle

Yeh crash course single-user hai: aap human board hain, aur testing ke liye issues bhi aap create kar rahe hain. real inbound source, support inbox, contact form, existing systems ka event, wire karne ka matlab hai strangers aap ki workforce mein write kar sakte hain. Is se pehle do cheezein true honi chahiye: inbound content process karne wale har Worker ke paas tight, clearly described authority envelope ho, "customer support" hone ki wajah se broad latitude nahin; aur un inputs ke liye deliberate plan ho jo kisi Worker ko map nahin karte, orphaned issues nahin. Dono par sochne tak single-user rahein.

Aage Kahan Jayen​

Scenario 6 ke baad aap ke paas working AI-native company hai: two Workers, one demonstrated approval, real data wala budget, aur seconds mein stakeholder questions answer karne wala audit trail. Shuru karne ke liye zyada tar logon ko yahi surface chahiye.

Is page mein touched topics ki patient walkthrough ya skipped topics ke liye getting-started directory mein deeper crash courses saath maujood hain. Quick map:

Aap kya chahte hainKahan jayen
fleet manage karne se pehle single Worker ko durable workflow engine mein wrap karnaFrom Digital FTE to Production Worker: Inngest durability envelope
hiring ko callable capability banana, Workers jo policy ke under doosre Workers hire kareinFrom Fixed to Dynamic Workforce: is course ka successor
Paperclip REST API ko kisi aur agent ke andar tools ke tor par use karnacurrent MCP server package ke liye docs.paperclip.ing check karein; version pin karein
Paperclip ko cloud ya shared host par deploy karna, multi-machine, multi-userlive docs at docs.paperclip.ing; rely karne se pehle verify karein kya ship ho chuka hai
OpenClaw ko humans aur workforce ke darmiyan edge layer ke tor par wire karnaOpenClaw crash course

Baqi zyada tar platform aap ka AGENTS.md already cover karta hai. Apne coding agent se poochein: "What does AGENTS.md say about the heartbeat contract?" ya "Walk me through the approvals section of AGENTS.md." brief reference hai; yeh page tour hai.

Meta-lesson: unzipped folder mein sab se valuable cheez AGENTS.md hai. Aik evening le kar isay end to end parhein, install steps ke liye nahin, document ki shape ke liye: discover-before-act table, working pattern, operations by task type, diagnostic checklists. Phir agle tool ke liye aisi hi file likhein jis ke samne coding agent rakhna hai. pattern portable hai: har tool jis ki learnable surface hai, us ke liye AGENTS.md likhne ke qabil hai. Paperclip clean target tha kyun ke install agent-driven setup se benefit leta hai aur operations paste-and-watch scenarios mein decompose ho jate hain; aap ko doosre targets bhi milen ge. Agla one author karein.

Flashcards Study Aid​