Product managers spend extra time on text than maximum other folks discover. Specs, consumer stories, reputation standards, purpose notes, stakeholder updates, the abnormal govt brief. The work is a component structure, section translation, edge negotiation. The writing is in which alignment will get outfitted or broken.
ChatGPT can aid right here, not as a substitute for judgment, but as a drafting partner that speeds up the stupid portions and sharpens the messy ones. Used well, it might store hours according to week and reduce transform throughout engineering and layout. Used poorly, it floods your backlog with primary fluff that no one reads. The change is in the way you prompt, while you convey it into the process, and the way you review what it produces.
I have deployed ChatGPT throughout discovery, making plans, and birth on groups starting from scrappy 5-adult startups to service provider packages with compliance gates and multi-zone timelines. The styles in this newsletter come from that revel in. They attention on specifications and consumer testimonies considering that the ones artifacts raise most of the ambiguity hazard. Good prose here is leverage.
Where ChatGPT suits within the PM writing flow
Think of PM writing as 3 stages: exploration, crystallization, and communique. Exploration is wherein you pull threads and gather inputs. Crystallization is in which making a decision. Communication is where you get the group moving. ChatGPT works top-quality on the transitions between these levels.
During exploration, it facilitates you fall down a messy set of notes into issues, spot contradictions, and cartoon unique solution frames. In crystallization, it could generate substitute spec outlines, suggest attractiveness criteria variations, and stress scan assumptions with edge circumstances. For communique, it turns decisions and reason into clear person studies and stakeholder summaries tailored for the various audiences.
The critical idea is to anchor the brand with concrete context. Feed it true user fees, genuine improve tickets, hard details, design snapshots, and a one-paragraph technique abstract. The output good quality rises dramatically when the variety is constrained with the aid of the truth of your product and clients.
Drafting crisp specifications with no losing nuance
A spec could be a selection record. It must always clarify what's going to modification for the consumer, why that change topics, the constraints, and how we are going to are aware of it labored. The catch with ChatGPT is verbosity. It will luckily produce 5 pages where one may do. Your process is to pressure in the direction of tightness.
Start with a one-paragraph difficulty framing to your own words. Keep it to a few sentences: who is affected, what is damaged, and what takes place when it’s fastened. Paste relevant snippets of evidence: two consumer prices, a every single day question depend from analytics, a screenshot or redacted ticket thread. Then ask ChatGPT for 2 or 3 selection spec outlines, each and every beneath one page, with alternative industry-offs. You will aas a rule see a framing you had not viewed.
I prefer a spec layout which is heavy on what and why, and easy on implementation element. Engineering needs house to design. The spec needs to define obstacles, not dictate code paths. ChatGPT can guide with the aid of itemizing hazards, metrics, and non-objectives chances are you'll miss, especially should you ask it to generate pushback. Tell it to argue towards your plan from the attitude of a senior engineer or a skeptical compliance officer. Often that yields a line or two that ameliorations the scope in successful ways.
Resist the urge to repeat the primary coherent draft into your doc. Instead, sew together sentences from the most efficient adaptation, upload references in your genuine artifacts, and trim the rest vague. ChatGPT is an accelerant, now not an alternative to conviction.
User studies that align habit and intent
User memories are deceptively clear-cut. The vintage layout is quickly to jot down yet light to put in writing badly. A important story avoids resolution bias, describes the cause and the final result, and is testable. The attractiveness standards do the heavy lifting. ChatGPT can write good studies swift, but you need to be explicit about the persona, the adventure level, and the real-global triggers.
Give the mannequin a slender slice of context. Describe the character in two traces, paste a snippet of the present UI float, and be aware the widespread frustration. Ask for 3 story editions that replicate numerous constraints. You would say, make one model that preserves zero back-cease changes, one other that accepts a one-week API adjustment, and a third that assumes we are able to upload a brand new experience to the analytics schema. This manner surfaces commerce-offs in engineering terms.
Good acceptance criteria examine like a agreement among product and QA. They ought to be deterministic, not poetic. Include the poor paths and limits. If a placing is purely noticeable to admins, say so. If rate restricting kicks in after five attempts in fifteen mins, lock it down. ChatGPT is appropriate at enumerating area situations when you set the boundaries and encompass the industry ideas that be counted.
When you are integrating with regulated archives or problematic pricing common sense, instantaneous the type to produce a insurance map. For example, ask it to checklist every mix of plan fashion, role, and characteristic flag which may regulate the user journey. You can then prune something inappropriate and flip the remainder into recognition standards.
A quick anecdote approximately speed and rework
On a crew construction a workflow automation characteristic for revenue ops, our first inside spec ballooned to eight pages with diagrams and a dozen user memories. Two senior engineers learn it and have been left with 3 definitely varied interpretations of the brink behavior. We rewrote with ChatGPT as a drafting accomplice. The immediate protected an anonymized CRM schema section, a transcript from a purchaser call, and a screenshot of the recent rule builder UI.
We requested for a spec that targeted on three conversion events, included choose-in habit by default, and assumed no modifications to the webhook structure. The model produced a two-web page draft. We then requested it to argue in opposition t the idea from the point of view of a overall performance engineer tracking queue intensity. It flagged a backlog possibility if we allowed cascading re-evaluations. That made it into the non-dreams and converted one recognition criterion. Result: the next engineering review took 25 mins, not 90, and we cut a sprint of remodel.
Turning discovery notes into a specific thing the workforce can use
Discovery dumps are the place readability is going to die. You have Figma links, call transcripts, a chatgpt AI chatbot spreadsheet with survey responses, and a 1/2-baked chart from analytics. ChatGPT is helpful at compressing this into a specific thing one can share in planning with no diluting the signal.
Paste a small sample of each input, now not the whole corpus. Summaries decay while the variety faces too much repetitive textual content. Ask it to extract person intents and objections with verbatim prices, then cluster them. Follow with a request for a single-web page transient targeted at the prime two intents and one counter-intent you propose to disregard, with a line on why. The forget about line matters as it facilitates stakeholders see that pronouncing no was taken into consideration, no longer unintended.
Treat the output as scaffolding. Replace time-honored phraseology with the language your customers certainly use. Swap adjectives for numbers. ChatGPT is much more likely to jot down “many clients combat” than “14 of 23 surveyed failed to find the export button within 30 seconds.” The latter is what your workforce desires to examine.
Avoiding the customary failure modes
There are three predictable ways this is going sideways. First, everyday tale templates that flatten context. Second, specifications bloated with redundant sections that nobody maintains. Third, reputation criteria that read like marketing copy.
To counter the first, shop a snippets library of your product’s vocabulary. A few strains for your roles, feature flags, naming conventions, and your quirky mistakes states will maintain the kind grounded. Paste snippets into your activates. The difference presentations up within the pronouns and the verbs.
On bloat, set not easy length limits and ask for omissions. A powerfuble trick is to ask for “solely the sections considered necessary for kickoff,” then record those you favor. There is not any prize for inclusive of a “Dependencies” header that stays empty for the next two years.
For recognition standards, push the mannequin toward if-then language and finite values. Avoid “may want to” and “could.” Ask it to enumerate adverse paths and parameter limits explicitly. If vital, supply it a skeleton with a number of proper values and request of entirety that preserves the development.
Acceptance criteria thoughts that carry up under test
Here is a small checklist that has kept me more insects than any unmarried software. Use it to sanity look at various kind-generated standards formerly handing them off.
- Define visibility, triggers, and submit-circumstance country explicitly, with at the least one unfavourable case for each. Include limits for lengths, counts, timeouts, and costs, despite the fact that provisional, so engineers ask approximately them. Capture position and permission interactions in plain phrases and map them for your genuine roles, not hypothetical ones. Specify what gets logged or instrumented for analytics, with event names and estate keys. Call out what is going to not come about, especially unfavorable actions that remain as-is.
Run by those five facets and you'll capture 80 p.c of the anomaly that produces manufacturing surprises. ChatGPT can pre-populate them. Your job is to substitute defaults and upload the single or two constraints specified for your machine.
Writing for distinct stakeholders without whiplash
PMs write for engineers, designers, QA, earnings, executives, and in some cases consumers. Each target market has extraordinary thresholds for detail and a other variety of endurance. The mistake is to produce one artifact and ship it everywhere. The larger pattern is to take care of a center spec plus tailored summaries.
ChatGPT is powerful at tone shifting. Feed it the core spec and supply a brief character transient for the audience. Ask it to provide a one-paragraph replace for the government channel that names the business outcomes, probability, and date. For gross sales, request a two-paragraph notice with beforehand-and-after screenshots and 3 strains of impression on objection managing. For QA, ask for a matrix of states and transitions, in textual content, no longer a photograph, that fits how they believe.
Avoid upbeat adjectives in these outputs. They slash credibility with senior readers. Replace them with concrete references to the plan. The kind will preserve seeking to add positivity. Edit that out.
Using ChatGPT for side case discovery
One of the best significance uses is to generate a catalog of edge situations right now. Given your constraints, ask for approaches the machine might fail with politeness. For illustration, for a bulk import feature, recommended the mannequin with your subject limits, document length cap, and blunders coping with design. Request edge circumstances grouped by user mistake, community habit, and knowledge form. You gets circumstances like “UTF-8 BOM at leap of CSV,” “mixed line endings,” or “header capitalization mismatch.” Compare the listing on your genuine-global logs and prune.
For intricate permission procedures, ask the form to enumerate eventualities where inherited permissions and direct can provide war. If you furnish your role hierarchy and inheritance suggestions, it will probably generate tale-sized cases in mins. Again, overview ruthlessly. The objective is to capture blind spots, now not to simply accept the list wholesale.
Measuring affect without slipping into self-esteem metrics
PMs are accountable for results. If you adopt ChatGPT on your writing circulation, you deserve to degree whether or not it reduces cycle time and defects. The most simple degree is evaluate loop remember. Track what number of evaluation cycles a spec calls for to attain engineering sign-off. If that drops by even a 3rd, you might be saving days in line with mission.
Another indicator is the illness combination put up-liberate. Look at the ratio of habits misunderstandings as opposed to implementation insects. If popularity criteria get clearer, you may still see a drop in misunderstanding-relevant things. It will now not exchange in a single day, but over two or Technology 3 sprints the style exhibits up.
You might also measure inner time saved. If you stay calendar blocks for spec drafting, notice total hours according to artifact ahead of and after. Most groups I have labored with see a 20 to forty p.c reduction in drafting time. The precise win, nevertheless, is less to come back-and-forth, which is more difficult to quantify yet very visible for your Slack historical past and standups.
Privacy, compliance, and what no longer to paste
Be cautious with sensitive records. Do not paste consumer names, PII, or proprietary algorithms into a 3rd-party instrument. Redact or replace with placeholders. If your supplier has an authorised ambiance for ChatGPT use, follow it. If no longer, write activates that incorporate schemas and platforms in place of raw files, and masks identifiers always. A risk-free sample is to hold a neighborhood notice with the unredacted context and change in placeholders sooner than sending the instructed.
Also feel criminal commitments. If your patron contracts restriction tips sharing, translating improve tickets into anonymized prices can nonetheless be unstable if blended with other indications. When unsure, work along with your defense team to ascertain a sanitization checklist.
A simple prompting sample you're able to reuse
The so much official consequences come from dependent prompts together with your very own language, tough constraints, and a clean output shape. Here is a development I use routinely with fantastic effects. Adjust the tokens to suit your product.
- Context: one paragraph dilemma statement on your words, plus two to 4 quick records features or costs. Constraints: the key limits and non-aims that subject for this alteration. Audience: who it's for and what they desire to do with it. Output shape: sections you desire, with individual or sentence limits according to part. Pushback: ask for two risks or disagreements from a described angle.
A skeleton feels like this in simple textual content:
Context: Sales managers shouldn't bulk reassign leads throughout territories with no exporting and reimporting CSVs. This explanations a weekly backlog and 12 percent lead getting old past SLA. Support tickets express frustration with errors messaging on import screw ups.
Constraints: no alterations to territory principles engine this region, CSV upload pipeline need to stay, no background jobs longer than 5 mins, audit log required for every single reassignment.
Audience: backend and frontend engineers for kickoff, QA for check planning.
Output: a spec beneath 700 words with sections for Scope, Non-objectives, User Impact, Acceptance Criteria (8 to twelve strains), Metrics, and Risks. Use our phrases: Territory, Lead Owner, Audit Event. Do now not embody structure diagrams.
Pushback: include two hazards from the viewpoint of the SRE on queue load and a mitigation in line with menace.
Feed it two or three truly mistakes messages and a screenshot description of the current import dialog. The first draft you get is many times 80 percent of the way there. You finish the ultimate 20 %.
When to ward off ChatGPT and write by hand
If you are exploring a touchy strategic shift, writing a memo that argues for a sizeable guess, or coping with a delicate trade-off with workers results, write the 1st draft yourself. Models have a tendency to soft the edges of hard selections. Your voice subjects in those circumstances. Bring ChatGPT in later to tighten sentences or to create stakeholder-certain summaries, now not to provide the core argument.
Another state of affairs is while the area has authorized or safe practices implications. Medical dosing calculators, monetary guidance engines, or anything else that touches public safe practices merits area expert authorship, then structured QA, then criminal evaluate. The mannequin can assist with formatting and checklist iteration, however it should still no longer be asked to produce authoritative reputation standards with out heavy supervision.
Maintaining a living spec library
The worst destiny for a spec library is to grow to be a graveyard of stale documents. If you are likely to accelerate construction, you furthermore mght need a protection plan. I retailer a temporary “spec delta” segment on the upper of every artifact wherein transformations are logged with dates and a one-line rationale. Once a quarter, we archive whatever thing untouched for 90 days or more into a study-simply folder and elevate ahead solely what's lively. ChatGPT is fabulous right here for producing diffs in case you paste the antique and new types and ask for a modification log in three lines. That makes stakeholder updates more straightforward and facilitates engineers see what replaced devoid of rereading the full document.
Bringing the workforce along
If you adopt ChatGPT for specifications and testimonies, make it obvious in your team. Share activates and drafts. Invite engineers to suggest on the spot upgrades and to make a contribution their personal templates. When engineers see their vocabulary mirrored in the output, they consider it more. Designers recurrently realize sooner iterations on UX copy and empty-nation language. QA teams advantage while acceptance criteria stick to regular patterns. Adoption grows while every one characteristic sees a direct improvement.
The different cultural move is to deal with the style as a junior collaborator, not an oracle. Encourage disagreement with its drafts. Ask teammates to spotlight anything that sounds conventional or uncertain. Reward edits that add specificity. Over time, your prompts will become greater compact, your patterns more good, and your workforce will spend much less time wrangling words and greater time development.
Final emotions from the field
Specs and person tales are where procedure meets truth. The craft is in turning messy inputs into clean directions devoid of suffocating creativity. ChatGPT is a mighty software for that craft when you use it with guardrails and flavor. Anchor it with your product’s certainty, supply it constraints that mirror your international, and ask it to argue with you earlier than you dedicate.
The reward compound. Drafts come in combination swifter. Reviews tighten. Edge circumstances get stuck earlier. And most likely most importantly, the team spends less electricity decoding rationale and greater vigor solving authentic complications for users. That is the level of the process.
If you decide upon one habit firstly, make it this: write your worry announcement and constraints for your very own phrases, paste two authentic artifacts, and ask for a brief spec with laborious limits and pushback. Edit ruthlessly. Share early. You will think the difference the 1st week and see it in the work via the cease of the month.