Growisto AI Bootcamp · Session 2 · Take-Home Deck (v4)
Skill creation —
make the work run itself.
Held: 26 May 2026 · 2 hours · ~60 attendees. This is the refined post-session reference. Skim it before you start your homework. Keep it open while you build your first SKILL.md.
3 brands demoed · 1 shared skill
1 live build · Upwork scanner
1 homework · due Sat 30 May
Recap · what you saw on the call
Three demos. One core idea.
If you missed parts of the session or want to anchor what you saw, here's what happened in order — and what each part was meant to prove.
10:00 — first 15 min
Theory + the Full Loop. CLAUDE.md = project brain. SKILL.md = repeatable workflow. Build once, run forever.
25:00 — middle 30 min
3-brand Listing Refresher demo. Same SKILL.md file in three folders (HydroFit, TerraBrew, NomadPack), three different brand outputs — because each folder's CLAUDE.md was different.
55:00 — Q&A
Two big questions. Shivani: how do teams share a CLAUDE.md? Salama: is this 1 skill or 3 skills? Both answered on the call — and on the next slides.
75:00 — last 30 min
Live Upwork scanner build. Nishant built a skill from scratch in front of you, then duplicated the folder and re-ran it with Amazon-marketplace rules. Same skill, different context, different output.
The one idea: rules and project context belong in CLAUDE.md. Repeatable workflow steps belong in SKILL.md. Keep them separate, and one skill becomes infinite skills.
The framework
The Full Loop
Every skill you'll ever build follows this exact pattern. Four parts. Two files. Forever.
📝
Step 1
Write context once
CLAUDE.md captures everything that doesn't change between runs.
→
💬
Step 2
Do the work manually
Chat with Claude. Iterate until the output is what you want.
→
📜
Step 3
Capture as SKILL.md
Write down the steps so the next run doesn't need explaining.
→
⚡
Step 4
Trigger forever
One phrase runs the whole thing. From any folder. Any account.
"Skill is four hours of work that you would have done — and it now gets done with one phrase."
— Nishant, on the call
CLAUDE.md — the project's brain
Think of it as your client wiki. Permanent.
If you had to onboard a new team member to your account, what would you tell them in their first hour? That's exactly what goes in CLAUDE.md. You write it once. Every Claude session in that folder reads it first, before you type a word.
What goes in
Client name. What you do for them. Brand voice. Banned words. Target persona. Tone rules. Where the data lives. Listing/output format rules. Which MCPs to use.
What doesn't
One-off questions. This week's tasks. Anything time-bound. CLAUDE.md is for the things that are still true six months from now.
Granularity tip from the session: CLAUDE.md works at any level. You can have one CLAUDE.md per brand (HydroFit, NomadPack), or — for big Amazon accounts — one folder per category with its own CLAUDE.md (rules for "fitness", "kitchen", "outdoor" separately). Same skill, finer-grained context.
SKILL.md — the repeatable workflow
If CLAUDE.md is the wiki, SKILL.md is muscle memory.
A SKILL.md captures one workflow — the steps Claude follows to produce a specific output. One project can have many skills: one for listings, one for ads, one for reviews, one for keyword research.
1
PurposeOne sentence. What does this skill produce?
2
TriggerThe exact phrase that activates it. Specific enough not to fire by accident. No brand names — keep it portable.
3
StepsNumbered, concrete. No "analyse", no "review" — be specific. As Nishant put it: "Analyse can mean a hundred different things to a hundred different people."
4
OutputThe exact format you want — sections, headers, length, what's allowed.
5
ContextWhere data lives in the folder. What CLAUDE.md it relies on. Safety notes (read-only? edits live data?).
The worked example you saw
Listing Copy Refresher
One of the most repetitive tasks on the Amazon team. Every account does it monthly. Take an old listing + new priority keywords → produce a revised listing that fits the brand voice.
Inputs
The current listing (title, 5 bullets, description) + a CSV of 10–15 priority keywords tagged P1/P2/P3.
Outputs
Revised title + 5 bullets + description + keyword coverage table + brand voice check + summary of what changed.
Why this was chosen as the demo: the SAME skill works for any brand, but the CLAUDE.md per brand changes everything — voice, banned words, persona, format. So the contrast across brands is dramatic. You saw it across the next 3 slides on the call.
Three folders, one skill
Same workflow. Three radically different brand voices.
Three brand folders. Each contains an identical SKILL.md but a completely different CLAUDE.md. That's the only difference — and it's enough.
HydroFit
Hydration
Energetic · gym-going
"Train hard. Sip cold. Built to last."
Banned: diet · premium · slim
TerraBrew
Coffee
Considered · sensory
"Notes of bergamot, jasmine, dark chocolate."
Banned: cheap · best · premium
NomadPack
Backpacks
Functional · specs-first
"40L. 1.1kg. 600D polyester. Honest specs."
Banned: stylish · premium · perfect
Each folder has 4 files: CLAUDE.md (different per brand) · SKILL.md (byte-identical across all three) · listing.md (current copy) · keywords.csv (priority-tagged).
The skill itself · byte-identical across three folders
SKILL.md — what the listing refresher actually says
SKILL.md · lives in all three brand folders# SKILL: Listing Copy Refresher
## Purpose
Take an existing Amazon listing + a list of priority keywords,
produce a revised listing that weaves in new keywords without
breaking brand voice.
## Trigger "Refresh listing copy"
## Steps
1. Read listing.md (title + 5 bullets + description).
2. Read keywords.csv — P1 must-include, P2 should, P3 nice.
3. Read CLAUDE.md for brand voice, banned words, persona, rules.
4. Rewrite title — lead with priority-order keyword from CLAUDE.md.
5. Rewrite 5 bullets — CAPITALISED benefit + colon, weave P1 keywords.
6. Rewrite description — 4–6 short paragraphs, plain text.
7. Run keyword coverage check + brand voice (banned-word) check.
8. Output 3–5 line summary of what shifted from original.
## Output Revised Title → Bullets → Description → Coverage → Voice Check → Diff
## Context Inputs in this folder. Brand rules in CLAUDE.md. Read-only.
Notice: the file never says HydroFit, TerraBrew, or NomadPack — because it doesn't need to. Whichever folder it runs from, that folder's CLAUDE.md carries the brand specifics.
The unlock you saw live
One trigger phrase. Three brand-true outputs.
All three folders. Same SKILL.md. Same trigger phrase: "Refresh listing copy". Three completely different outputs because each folder's CLAUDE.md is different.
HydroFit · 158 chars title
BUILT FOR YOUR WORKOUT: 32oz stainless steel insulated water bottle keeps your drink cold for 24 hours, hot for 12. Train hard, sip cold.
TerraBrew · 188 chars title
ORIGIN, NAMED: Grown in the Yirgacheffe region of Ethiopia at altitudes above 1,800 metres — a single origin coffee from one washing station, not a blend.
NomadPack · specs-led
CARRY-ON COMPLIANT, 40L: 55x35x20cm, 1.1kg empty. Fits major airline cabin allowances. Verify your specific carrier before flying.
The depth moment: for NomadPack, the skill refused to claim "Ryanair compliant" because the dimensions don't fit Ryanair's underseat limit and CLAUDE.md said "be honest about trade-offs". Brand rules in CLAUDE.md don't just shape voice — they shape what the skill will and won't say.
The unplanned highlight · live build
Upwork Scanner — built from scratch in 25 minutes, on the call
Nishant built a real skill in front of you. Here's the arc — and what to take from it for your own homework.
1
Started with manual workOpened Upwork, took screenshots, pasted into Claude. No skill yet. Just doing the task.
2
Reviewed Claude's first cut, line by lineFor each job, asked: do I agree? Where's it wrong? Encoded each correction as a rule.
3
Built CLAUDE.md alongsideCapability pillars, strong verticals, job evaluation criteria — all the rules Claude was learning got captured in CLAUDE.md.
4
Pulled in Chrome MCP for automationOnce the rules were clear, plugged in Claude in Chrome to do the scanning automatically — Cloudflare blocks and all.
5
Duplicated the folder for Amazon workCopied the folder, updated CLAUDE.md for Amazon Marketplaces context, ran the same skill — got Amazon-PPC-relevant jobs instead of Shopify jobs. Same skill, different context.
The lesson for your homework: don't try to write the perfect SKILL.md first. Do the work manually, capture the rules as you go, then turn it into a skill once the pattern is clear.
Big question 1 · from Salama on the call
"Is this 1 skill with 3 steps — or 3 separate skills?"
Context: Salama was thinking about writing web page copy. Three considerations — copywriting tone, SEO keywords, CRO outline. Should that be one skill that does all three? Or three skills?
"Those are rules, not skills. Rules go in CLAUDE.md. The skill is just 'create web page content' — and it reads the rules from CLAUDE.md."
— Nishant's answer
Goes in SKILL.md (the action)
- "Create web page content for [URL/brief]"
- "Generate FAQ section for [page]"
- "Refresh comparison table"
- "Build H1/meta/canonical for [page]"
Goes in CLAUDE.md (the rules)
- Brand tone of voice + examples
- SEO keyword priorities for this category
- CRO best practices for this funnel stage
- Banned phrases, target persona
The test: if it changes from project to project → CLAUDE.md. If it's the same action repeated across many projects → SKILL.md.
Big question 2 · from Shivani on the call
"If 6 people are on a project, do we share one CLAUDE.md?"
Yes — and there's a pattern that works for teams. Some context is shared (what the brand is, who they are), some is role-specific (the listing person doesn't need to read the ads rules).
Shared base CLAUDE.md
Brand identity, what we do for them, do's/don'ts that apply to everyone. Everyone on the team reads this.
Role-specific overlays
Listings person has listing rules. Ads person has campaign thresholds. Each lives in its own folder.
Sync mechanism
Obsidian (recommended), Google Drive, or GitHub. One source of truth. Everyone pulls the latest.
Recommended pattern (from the session): use a sync tool like Obsidian as the team brain. All shared docs — CLAUDE.md, weekly reports, mind maps — live there. When CLAUDE.md updates, everyone gets the update automatically. No more "which version is current?"
What you saw briefly on the call
Org-wide skills — the slash-command path
When a skill is good enough to share with everyone, it stops being a file in your folder and becomes a command anyone can type.
Today — your folder only
SKILL.md sits in your client folder. You type the trigger phrase. Only works when you're in that folder.
Org-wide — the / command
Skill is uploaded once to the org. Anyone in the company types /skill-name and runs it. Hardik's Amazon Listing Generator works this way — you saw it in the session.
# In Claude Code, type "/" to see all installed org-wide skills
/amazon-listing-generator
/shopify-psi-report
/lead-audit-deck
/cro-audit-report
# ...and 10+ more across departments
The path: build it in your folder → use it for 2 weeks → refine → submit to Pooja/Shreenidhi → it becomes a slash-command for everyone. That's Session 3 territory.
Migrating from Claude chat to Claude Code
"I've been using Claude chat / projects. How do I bring that work into Code?"
A lot of you have months of work in Claude chat with custom projects already. That doesn't go to waste — Claude can help you do a clean handoff to Claude Code, where the agentic tools live.
# In your existing Claude chat project, paste this prompt:
I want to hand this project off to Claude Code where I can run
agentic workflows and MCPs.
Please prepare a handoff package by:
1. Summarising what this project is for, in 5–10 lines, ready to become CLAUDE.md
2. Listing every recurring task I've done with you here
3. For the top 3 recurring tasks, writing draft SKILL.md outlines
4. Listing every external tool I'd need MCP connections for in Code
5. Highlighting any rules or preferences I've taught you over time
Output as separate files I can copy into a new folder.
Why this matters: Claude chat has projects and connectors, but it's not agentic — most MCPs don't run there, and live tool use is limited. Code is where skills, agents, and end-to-end automation actually run. Don't restart from zero — port over.
The most important habit · 5-minute rule
Before every new session — ask Claude what it knows.
Every Claude session starts with a fresh context window. CLAUDE.md gets read, but Claude doesn't know what you've changed since the last session. Spend 5 minutes asking before you do anything else.
# First message of every new session in a project folder:
Before I start making changes — what do you currently know
about this project?
Specifically:
1. What rules have I set up in CLAUDE.md?
2. What skills are available in this folder?
3. What data sources/MCPs are connected?
4. What's the last significant work we did together here?
Read the relevant files first, then summarise. I'll correct
anything that's wrong before we proceed.
From the session: Nishant called this out as the rule that prevents the most skill failures. "Spend 5 minutes standardising the context. Don't go heavy until you're satisfied with what Claude says it knows."
The compounding view
Each skill is a deposit. The library is the bank.
One SKILL.md saves you 5 hours a month. Ten SKILL.mds across the Amazon team — listings, digests, reviews, keywords — saves the team hundreds of hours a month. That's the destination.
📁 client-folders/
├── 📁 amazon-hydrofit/
│ ├── 📄 CLAUDE.md HydroFit voice rules
│ └── 📄 SKILL.md listing refresher — yours
├── 📁 amazon-technosport/
│ ├── 📄 CLAUDE.md Technosport voice rules
│ └── 📄 SKILL.md same file, copied in
└── 📁 amazon-nivia/
├── 📄 CLAUDE.md Nivia voice rules
└── 📄 SKILL.md same file, copied in
Distribution rule: share the file, not the folder, not the prompt. A copied prompt is a snapshot. A SKILL.md file is alive — update it once, re-share, everyone gets the improvement.
When to build a skill — and when not to
Not every task deserves a SKILL.md. Most don't.
SKILL.md is for the work you do more than once. If it's a one-off, just chat with Claude. Use this checklist.
Build a skill if…
✅ You do this more than once a week
✅ The shape is the same every time (inputs, steps, output format)
✅ Different accounts/clients need the same kind of output
✅ You can describe the steps to a junior teammate
✅ The output is reviewable in under 5 minutes
Skip the skill if…
✗ It's a one-off analysis
✗ The steps change wildly each time
✗ You can't explain the rules clearly
✗ The output needs deep human judgement at every step (not yet)
✗ You haven't done it manually 3+ times
Pro tip from the session: the best SKILL.mds are written after you've done the work manually with Claude in the folder 3–4 times. Tell Claude to note every rule it applies. By the end of the week, ask Claude to draft the SKILL.md from what you've already done together.
Watch for these
The 5 most common skill-writing mistakes
1
Vague steps ("analyse the data")If a junior teammate wouldn't know what "analyse" means, Claude won't either. Be specific: "calculate ACoS per ASIN, flag if >55%".
2
Trigger phrase too generic"Check this" or "run report" will fight with other skills. Use specific verb + noun: "Refresh listing copy", "Run weekly digest", "Score this lead", "Run upwork scan".
3
Brand info hard-coded in SKILL.mdIf "HydroFit" appears anywhere in your SKILL.md, it's not portable. Brand info belongs in CLAUDE.md. SKILL.md should be brand-agnostic.
4
No output format specifiedEvery run produces a different shape. Account managers can't scan it. Always lock the output format with explicit section headers and length limits.
5
Writing the SKILL.md before doing the workYou'll miss edge cases. Do it manually first — work in the folder for a week, then capture what you actually did as the skill.
Your homework · due Saturday 30 May
Pick one workflow.
Build the skill.
Run it on 2 accounts.
By next Saturday, every person on the call has one working SKILL.md in their client folder, tested on at least two real accounts, submitted to Pooja and Shreenidhi.
1
Pick a workflow you've done 3+ times this month
Listing refresh. Weekly digest. Review reply. Content brief. Audit report. Anything repetitive.
2
Open your account folder. Write CLAUDE.md + SKILL.md
Use templates on slides 20–21. Edit to fit your account.
3
Run it on at least 2 accounts
Same SKILL.md, two different CLAUDE.mds. Both outputs should feel right for that account.
4
Email it to Pooja + Shreenidhi by Sat 30 May EOD
Subject + 5 lines + your SKILL.md attached. Format on the homework brief doc.
Quick reference · copy this
CLAUDE.md — starter template
CLAUDE.md · drop in any client folder# [Client Name] — [Platform / Channel]
## What this project is
[1–2 sentences: client, what you do, the goal]
## Brand Voice
- [3–5 bullets: tone, sentence length, do/don't]
- Examples that fit · examples that don't
## Target Persona
[Who buys this. What they care about / ignore.]
## Banned Words
- [word] — [why]
## Rules & Thresholds
- [Metric] target / Flag if [condition]
- Approval rules, char limits, format
## Output Format
[Where output lands and in what shape]
## MCPs Connected
- [Tool] — [what you use it for]
## Common gotchas
- [Things Claude got wrong before]
The test for what goes here
If it changes between projects → CLAUDE.md. If it's the same action repeated across many projects → SKILL.md.
How to start
Paste in: "Update my CLAUDE.md. Here's the project: [...]. These are the rules: [...]" — let Claude draft v1, edit by hand.
Sync across team
Use Obsidian / Google Drive / GitHub as the single source. Everyone pulls the latest. No "which version is current?"
Quick reference · copy this
SKILL.md — starter template
SKILL.md · brand-agnostic, drop in any folder# SKILL: [Skill Name]
## Purpose
[One sentence: what does this skill produce?]
## Trigger
"[Exact phrase that activates this skill]"
## Steps
1. [Read which file from this folder]
2. [What to read from CLAUDE.md]
3. [What to calculate or check]
4. [How to apply rules / thresholds]
5. [How to format the output]
[Be specific. No "analyse" or "review" — Claude won't
know what those mean for you.]
## Output
[Exact section headers. Exact order. Length limits.
What's allowed/banned in the output.]
## Context
- Data files: [filenames] in this folder
- Brand rules: see CLAUDE.md
- Edge cases: [missing data? new items?]
- Safety: [read-only? writes to live? approve only?]
Trigger rule
Specific verb + noun. "Refresh listing copy" ✓ · "Run report" ✗. No brand names in the trigger — keep it portable.
Steps rule
If a junior teammate wouldn't know what a step means, neither will Claude. "Calculate ACoS per ASIN, flag if >55%" not "analyse spend".
The 3-run rule
Do the work manually 3 times before writing the SKILL.md. By then you know the real edge cases — you're documenting reality, not guessing.
What's coming · Session 3
Two halves. Hybrid format.
Session 3 will be different from Sessions 1 and 2. Less lecture. More working with your actual skill attempts.
First half (60–70 min)
Skill Clinic
3–5 volunteers walk through their homework SKILL.md. Group debugs together. You learn what works (and what breaks) by seeing other people's attempts. Bring your screen ready to share if you want feedback.
Second half (50 min)
Distribution Lite
How to share your skill so anyone in the company can install and run it with one command. The slash-command path. Team CLAUDE.md sync via Obsidian/Drive. First look at the growisto-skills library.
To get the most out of Session 3: finish your homework. Run your skill at least 5 times by Friday before. Bring the version you're not yet happy with — that's the one we can help most with.
All resources in one place
Resources & help
📄 This deck (PDF)
Bootcamp_S2_Presentation_v3.pdf — the take-home reference you're reading now.
📝 Homework brief (DOCX)
Bootcamp_S2_Homework_Brief.docx — the assignment, with examples by department, common pitfalls, and the submission template.
📁 Demo folders (ZIP)
demo_folders.zip — HydroFit, TerraBrew, NomadPack. Unzip into your workspace, open in Claude Code, run "Refresh listing copy".
🎥 NotebookLM walkthrough
Bootcamp_S2_NLM_SKILL_Walkthrough_v2.md — feed into NotebookLM, "Generate Video Overview" for a 3–4 min training video.
🎙 Session 2 recording
Fireflies link — go back to anything you missed or want to re-watch.
💬 Get help
General: Nishant · Amazon: Hardik · SEO: Ajay · Tech-CRO: Shailesh · Advanced: Aniket
The key principle for getting unstuck: questions only come when you start building. Don't wait for clarity to begin — begin to find clarity. Reach out the moment you hit a wall.
Growisto AI Bootcamp · Session 2 · Take-Home v4
Build the skill once.
Run it on every account.
The first SKILL.md takes 30 minutes. The tenth takes 10. By the tenth, you've stopped doing work — you've started building systems.
Your next 7 days
Day 1–2: Pick your workflow, open the folder
Day 3–4: Write CLAUDE.md + SKILL.md, run it
Day 5–6: Tune, then test on second account
Day 7 (Sat 30 May): Email Pooja + Shreenidhi
Questions? Reach out. Don't wait until Session 3 to get unstuck.