The Software Manager
By Zeel Jain
THE EDITORIAL ARC · PART 2 OF 5 The new white-collar skill, in one playbook. Five fields, three real examples, and the promotion test for managers in an AI-augmented team.
WHY THIS EXISTS
You got promoted. Nobody told you.
If software does work in your day; drafts emails, books meetings, writes code, processes hires, you’re already a software manager. You probably never agreed to it. Nobody handed you a manual.
The skill that matters now isn’t prompting. It isn’t picking models. It’s an older skill that found a new use: writing job descriptions, except the hire is software. Bad managers of humans produce bad teams. Bad managers of software produce broken stacks. Same skill. New medium.
“You don’t need to learn to code. You need to learn to brief.”
What you’ll have when you’re done
- The 5 fields every software job description needs.
- Three filled-in examples from real Know AI agents.
- The promotion test for software managers in an AI-augmented team.
THE FIVE FIELDS
Every software job description needs these.
Same five fields a good HR job description has. Same five fields a working AI agent has. That’s not a coincidence, it’s the same management problem in two surfaces.
01 Role
One sentence. What is this software for? Action-first, not category-first. ‘Reply to inbound DMs from prospects’ is a role. ‘Sales agent’ isn’t. The same difference as ‘analyse customer churn weekly’ vs ‘data analyst.’
02 Inputs
What comes in, where from, in what shape. Specifically. The wrong assumption about input shape kills more agents than bad models. Treat this like the dataset spec for a junior analyst. If the data isn’t named, the work won’t happen.
03 Standards
What does good output look like? With examples. ‘Match our brand voice’ is a wish. ‘Match the tone of these three example replies, with sentences under 18 words and no exclamation marks’ is a standard. Software does exactly what you specify. Specify well.
04 Edge cases
What you’d teach a new hire on day one. The three or four scenarios that come up weekly and need a different rule. Don’t try to cover every possible case, cover the recurring ones.
05 Escalation?
When does the software stop and ask a human? Software without escalation rules is software that confidently does the wrong thing. Make it cheap for the agent to raise a hand. Reward escalation, don’t punish it.
THE PROMOTION TEST
Hand your job description to someone who’s never done the work. Ask them to do it from your doc alone. If they can, the software can. If they can’t, the software can’t either. The job description is the job.
EXAMPLE 1 / OPERATIONS
The onboarding agent.
Real. Runs at Know AI. Processes around 20 hires a month. Names changed; logic intact.
Role
Generate offer letters, employment agreements, and Day-1 documents for new hires. Send the document-requirement email and track returned files.
Inputs
- New hire row in the ‘Pipeline’ sheet (name, role, start date, comp, employee/intern flag).
- Returned documents in the hire’s Drive folder (6 documents per hire).
Standards
Offer letters use the correct template (employee vs intern). Names and dates correct. Comp matches the pipeline row. Email tone: warm, professional, signed by the operator.
Edge cases
- Intern conversions: pull prior intern record, don’t treat as fresh hire.
- Comp above approval threshold: route for partner approval first.
- Documents missing >5 days: send polite reminder template.
- Start date moves: regenerate Day-1 docs only, don’t resend offer.
Escalation
Stop and ask the operator if: role title isn’t in standard list, comp is above threshold, Drive folder template fails to copy, any document looks tampered with. Never auto-send the offer letter, always queue for review.
EXAMPLE 2 / CUSTOMER FACING
The WhatsApp sales agent.
Different category, same five fields. Customer-facing, real money on the line. Built for an Amazon FBA coaching business.
Role
Reply to inbound WhatsApp messages from prospects. Qualify, address objections, share the payment link when ready.
Inputs
Single WhatsApp message from a prospect. Conversation history with that prospect. Reference: program details, pricing, top 12 objections + responses, payment link template.
Standards
Brand voice: warm, direct. Hindi pronoun ‘aap’ only, never ‘tum.’ Sentences under 20 words. No emojis. No exclamation marks. Objections handled from the playbook verbatim, not generated.
Edge cases
- Never mention EMI or instalments. Never imply discounts.
- Never break Hindi pronoun rule even if prospect uses ‘tum.’
- If conversation loops on the same objection 3 times, change strategy.
Escalation
Hand off to a human if: prospect asks for a custom price, claims a previous purchase, raises a refund or legal issue, or sends a message in a language other than Hindi or English.
EXAMPLE 3 / Internal Content
The brand voice agent.
An agent we use to keep our content consistent across writers and channels. Different domain again, same five fields.
Role
Review draft content (carousels, captions, reels) against the brand voice guidelines. Suggest edits where the draft drifts. Never rewrite without flagging.
Inputs
Draft content (text). Brand voice doc (rules + 20 examples of approved writing). List of phrases banned in our voice.
Standards
Output is a list of suggested edits with the rule each one references. Never an alternative rewrite without explanation. Suggestions are surgical, one edit per drift not whole-paragraph rewrites.
Edge cases
- If the draft is fine, return ‘no edits needed’, don’t invent suggestions.
- If the draft is unrecognisable from our voice, flag for human rewrite, don’t patch.
- If a banned phrase appears, flag with rule reference, don’t silently substitute.
Escalation
Flag for human if: the draft uses an unfamiliar topic the brand hasn’t covered before, the suggested edits exceed 5 per draft (signal of bigger drift), the writer pushes back on a suggestion repeatedly.
THE PROMOTION TEST
Are you a software manager yet?
The same way we’d evaluate a junior manager getting promoted to senior, here’s the test for a knowledge worker becoming a serious software manager. Honest answers only.
01. Can you write the role in one sentence?
If your description starts with a category ‘sales agent’ then you’re thinking in titles, not jobs. Senior managers think in jobs.
02. Have you defined the standards with examples?
Vague standards (‘match our brand’) are a junior move. Senior managers show, not tell. They include 3 examples of what good looks like.
03. Are your edge cases real, not hypothetical?
List edge cases that have actually happened in the last month, not imagined ones. If you’re inventing edge cases, you don’t know the work yet.
04. Is escalation a path, not an apology?
If your escalation rule reads ‘ask if unsure,’ you haven’t defined when. Senior managers name the conditions that trigger escalation.
05. Are you reviewing the brief, or just the output?
Junior managers review outputs. Senior managers review the brief and update it monthly. If your prompt hasn’t been touched in 8 weeks, you’re hoping, not managing.
THE HONEST VERDICT
If you got 5 out of 5, you’re managing software seriously. If you got 3–4, you’re on the curve. If you got 0–2, you have the same opportunity every great manager has at some point: a clear thing to get better at, with feedback in days instead of quarters.
If this was useful
This is part 2 of a 5-part editorial series on building AI workflows that actually compound. Real receipts from systems we run at Know AI. No jargon, no hype.
Follow @zeeel.ai on Instagram for the rest of the series.
ONE MORE THING
Write the job description for one piece of software you’re currently managing badly. DM it to me on Instagram with the keyword MANAGER. I’ll send back a few notes, not a rewrite, just the things I’d ask if I were running this with you.