> can you generate the retention heatmap again
> who were the key people using us and for what
> spin up a bg agent for each and figure out what they are using the product for
> there are millions of logs in this how do i export them from observability?
> can we start making a unified map of api key ids
> give me signups graph day on day in PST last month
> transfer my existing status page to a new provider
> i am writing a post mortem doc for an incident
> audit all the SIEM rules we deployed they feel too noisy
> zip up all my claude code sessions and send it to a teammate
> with this design doc trigger a night sandbox
> find me a good mattress, use cdp do proper research
A field report · June 2026

Claude for knowledge work

How I run product, incidents, analytics, and hiring out of a terminal — Rahul
built by claude, from 10 weeks of my own session logs
The point

A coding agent — that I almost
never use to build an app

55%
analyze data
18%
debug & fix
12%
build
6%
plan & design
6%
quality
My own 30-day work-type split. The thing that makes Claude Code good at code — tools, files, verification loops — is exactly what makes it the best knowledge-work tool I've used. This talk is in three parts.
01

What I've done with it

Analytics, incidents, hiring, design docs, writing, ops — the same agent, every category of the job. A tour of real artifacts, not slideware.
High level

One agent, every category

Analytics
cohorts, funnels, retention heatmaps, growth projections — against the live warehouse
Incidents
a multi-day operational incident run end to end from the terminal — specifics redacted
Hiring
sourcing CSVs, take-home evals, onsite kits seeded with realistic data, ATS triage
Design docs
specs co-authored & published live, claims validated against the production DB
Writing
strategy docs by voice, creator briefs, a public incident disclosure
Ops
status-page migration, support triage, job-board sweeps, dashboards fixed via the browser
Six categories most teams split across six tools and six people. Here, one window. Let's look at a few.
Analytics

It argues with headlines

The ask
with the old dashboard versus the new dashboard… is the activation better?
The scary headline — one segment's activation looked like it fell sharply — was a mix-shift artifact. Overall 7-day activation actually rose.
Tens of thousands of users; most new signups deliberately route to a different surface. A relative gain, not a collapse.
activation: up, not down
overall 7-day activation rose once you correct for mix shift
a real chart the agent generated — the actual artifact
Analytics

Vague question in, defensible number out

bear / base / bull projection
15- and 30-day forecasts, with the growth rate itself increasing
a real chart the agent generated — the actual artifact
The ask
we are about to cross a big tool-call milestone… project out 15 days, 30 days and assume the growth rate itself increases
Same muscle, same week: a minute-level downtime ledger reconstructed from logs (with receipts attached) and a record-traffic day attributed user by user.
Research · fan-out

One sentence, ten analysts

The prompt — in full
in the last week who are the top users, can you spin up a bg agent for each and figure out what they are using the product for
10
parallel agents, one dossier per user
1 each
dossier per power user — usage pattern, integrations, and intent
0
of the heaviest users are interactive — all scripted loops
Out the other side: ICP observations, abuse flags, outreach candidates — and one misclassified account.
Research · agent teams

A team of subagents, named

The prompt
deeply go through the usage pipeline and find 50 good usecases… i want to spawn further team mates and sub agents to do this
# the team it spawned
finder      scoped ~hundreds of thousands of rows → candidates
understander scored hundreds of real sessions
synthesizer themes → 50 use cases, personas
skill-gen   dozens of ready-to-run skill files
100s
real sessions read & scored — every use case traceable to one
1
overnight, mostly — a research project that would've been someone's quarter
Incidents

An incident, run from one session

Specifics redacted
A multi-day operational incident — investigation, response, and follow-up — run end to end from a single terminal session, with every connected system reached through Composio's CLI. The only transferable lesson: reversible, dry-run-by-default scripts, and findings independently re-checked before anything ships.
Details — systems, numbers, and the response itself — are omitted from this public version.
1
session, multiple days — end to end in the terminal.
Writing

Writing that actually ships

Strategy by voice
A long voice ramble over a planning doc → Claude interviews me, restructures, assigns owners, pushes it back to the doc tool. Now the canonical plan.
Briefs under fire
Creator brief v0→v3 in one sitting with live taste feedback (“this tag line is dog shit”) — and a CONTEXT.md recording every cut so v4 starts warm.
Comms with stakes
A high-stakes external write-up: a careful long-form draft, a canon-facts header, and a [LEGAL/COMMS] flag on every claim needing counsel.
The pattern: taste stays human — structure, drafts, verification, and delivery don't have to be.
Ops

It runs live systems

Migrate the status page. Dozens of incidents, a hundred-plus historical updates, groups & components — recreated backdated, zero dashboard clicks.
When the public REST API turned out to be a crippled subset, it reverse-engineered the RPC API the vendor's own CLI uses, and scripted that instead.
Hundreds of support tickets deduped & themed; surfaced a twice-filed URGENT blocker sitting unanswered
Live BI dashboard cards fixed by driving the browser itself (CDP)
Job postings normalized & verified in the ATS — and it knew to leave an intentional “joke” posting alone
02

How to actually play with it

The range above isn't talent — it's a handful of habits. Treat it like a playground: one front door, lots of random asks, real access, and trust that ratchets up.
Habit · 01

One front door for everything

Observability
there are millions of logs here — can you pull them from observability through Composio's cli?
Board-level analytics
last tuesday PST was a very high day for us. who were the key people using us… and for what
Hiring
find all the platform/infra people at <target company>… and make a csv with their linkedins
Comms
zip up all my claude code sessions from the last 30 days and send it to a teammate
Same window, same muscle memory. The reflex: don't go to the tool — ask the agent first.
Habit · 02

…and a lot of random shit

Primary care
“find me a good mattress, use cdp” → 4 Reddit back-pain threads scraped, real prices, a buying-guide infographic
Late-night logistics
a campground for Saturday → checks federal + state + 3rd-party sites, finds live inventory
Social
scroll my whole Twitter following list, compile a report, post it to Slack — via computer use
Inbox by hand
screenshot all 34 of those emails in my mail client and DM them to a teammate
Just for fun
a WarGames-style steganography easter egg on a side-project site — one message for agents, one for humans
Personal site
pull my non-work Slack “musings” and build a writing page on cryo.wtf
The point of the random stuff: it's how you learn the edges. The reps transfer straight back to the work.
Habit · 03

Give it what a new hire gets

# the agent's badge access — all via Composio's cli
prod database     read-only viewer URL
warehouse         BI + columnar store via cli
observability     logs / metrics / SIEM
comms             chat, email, support inbox
docs & tickets    docs, issue tracker, git host
the browser       chrome devtools (CDP)
the computer      screen, mouse, keyboard
Production database Q&A became conversational: paste a UUID, get back the org, the owner, and the allowlist state.
support workflow
Every connection above is brokered through Composio — one CLI, one auth surface across all of them. An agent with no access is a chatbot; scoped, mostly-read-only access via Composio is what turns intelligence into work.
Habit · 04

Demand artifacts, not answers

Every session should end in files: a chart, a CSV, a report, a script that re-runs itself.
Chat scrolls away. Files compound — they get re-run, shared, mined, and (today) turned into decks.
$ ls 02-06/                    # the day we moved status pages
STATUS-PAGE-MIGRATION.md       # what happened + how
migrate-incidents-v2.ts        # dozens of incidents + updates
build-structure.ts             # groups + components
incidents_backup.json          # escape hatch, kept
sessions/                      # full transcripts, archived
721 markdown docs
110 charts
77 CSVs
370 archived transcripts
Habit · 05 — how far you can push it

Trust is a ratchet

L1
Ask
answers, lookups — where most people stop
L2
Analyze
the warehouse talks back — cohorts, funnels, projections, anomaly hunts
L3
Produce
documents, datasets, briefs, decks — artifacts that ship to humans
L4
Operate
live systems: status pages, dashboards, job boards, support queues
L5
Delegate
fleets of agents, overnight runs, war rooms — you review, it executes
Each level earns the next. A global hook pings me on Slack when a session needs approval — so I can walk away.
Habit · 05 — the far end

While you sleep

Midnight, after finishing a design doc
okay now with this design doc can you trigger a night sandbox which will go build this across the two repos, and keep the stacked prs ready for me in the morning to review?
3 repos checked out on a cloud VM with a watchdog loop
28 concurrent agent sessions across 7 repos in one 8-hour window
Close the laptop. The review queue is the new inbox.
03

Steal my setup

The folder, the hooks, the memory, the skills, the publishing. ~60 lines of bash and a few conventions — copy them into any repo this afternoon.
The setup · the workdir

One folder per day, since March

9,510 files · 721 docs · 110 charts · 370 transcripts
✓ = a slide in this deck
Mar
2 days
30-03 ✓
31-03
Apr
25 days
01-04
02-04
03-04
04-04
05-04
06-04
07-04
08-04
09-04
10-04
13-04
14-04 ✓
15-04
16-04 ✓
17-04
20-04
21-04
22-04
23-04
24-04
25-04
26-04
27-04
29-04
30-04
May
24 days
01-05
02-05
04-05
05-05
07-05
10-05
11-05
12-05
13-05
14-05
15-05 ✓
18-05
19-05
20-05
21-05
22-05
23-05 ✓
24-05
25-05
26-05
27-05
28-05
29-05
30-05
Jun
9 days
01-06
02-06 ✓
03-06
04-06
05-06
08-06
09-06
10-06 ✓
11-06
A SessionStart hook makes the folder; a SessionEnd hook archives every transcript into it; opening the laptop commits yesterday. This whole audit trail is why this deck could be built.
The setup · hooks

Three hooks run the office

# SessionStart — make today's folder + commit yesterday
TODAY=$(date +%d-%m)
mkdir -p "$DAY_DIR/sessions"
echo "export TODAY_DIR=$DAY_DIR" >> "$CLAUDE_ENV_FILE"
git add -A -- ':!'"$TODAY"   # everything except today
git commit -m "auto-commit prior day's work" && git push

# SessionEnd — archive transcript + write .md summary
# PostCompact — log that compaction happened
74
auto-commits — the git history is a diary nobody had to write
The whole convention is ~60 lines of bash. Drop it into any repo and that repo starts keeping its own memory.
The setup · memory + skills

Teach it once

Never take a lazy shortcut and then ask “would you like me to do this properly?”
memory/feedback_no_lazy_shortcuts.md
A thought partner, not a task executor… always go one step further than the literal request.
memory/feedback_full_initiative.md
Corrections become memory (33 files). Repeated workflows become skills (6 hand-built):
/consumer-metrics
/morning-brief
/status-page-incident
/observability-debugging
/build-overnight
/verify
On June 9 the metrics skill found its own config stale, fixed itself, and still delivered the numbers.
The setup · reach

One CLI is the universal arm

# small asks are one-liners
$ composio execute METABASE_POST_API_DATASET \
    --data '{"native":{"query":"SELECT ..."},"database":<id>}'

# big asks are typed scripts with real concurrency
$ composio run -f fetch_retention.ts   # Promise.all, not loops

# anything with auth, proxied
$ composio proxy --toolkit slack -- POST /api/search.messages
BI · warehouse · chat · email · observability · docs · issues · git — one tool surface, every system the job touches.
Plus the browser itself: chrome-devtools is my single most-used MCP — 158 calls in 30 days. When there's no API, it clicks.
The setup · publishing

Publish where the team already is

$ cp design-doc.md live/design-docs/
# …that's it. instantly at
https://<my-internal-host>/design-docs/<slug>.md

# laptop tunnel → edge worker → hosted fallback
# markdown rendered with mermaid, code, dark mode
8
design docs live in 4 weeks — watch a product direction evolve doc by doc
Feel free to send this doc to your clanker. If it has feedback, it should email my agent.
mandated footer on every design doc
Docs ship with an AI-readable review channel: your agent reads mine, then emails my agent.
The bill

What all of this costs

~1B
tokens in this folder over ~10 weeks — roughly four figures in API-equivalent cost
~10B
tokens across all the coding agents I run this spring, combined
~95%
of it is cache reads — the marginal token is nearly free
~400M
tokens on the single busiest day (a postmortem + a status migration)
Calibration: a normal week is ~50–100M tokens; a heavy day is several hundred million. The entire deck's worth of work — analytics, hiring, design docs, ops — cost less than two contractor-days.
Token logs only survive from early April — but git goes back further: Claude co-authored commits ramp steeply month over month, the first one in 2025. The usage predates the logs.
Beyond this folder

The harness is the asset

The same muscle ships code: the majority of all commits across our main repos since January are Claude co-authored — well over half.
And the setup outlives any one model — when I stood up another agent, the first prompt was to copy this harness. Other agents do the same knowledge work through computer use. The folder, hooks, memory and skills are what compound.
A new agent, day one — verbatim
Can you understand how the Claude is set up in this repo and make sure that the translation sort of copies over to you?
60 day folders
~10B tokens
8 design docs live
1 deck — this one, built by Claude from these logs, today
feel free to send this deck to your clanker → it can reach mine
1 / 26
Redacted · public version