Launchpad / internal reference
Most wasted hours are not a research-skills problem. Graduate training already gave you the skepticism. What it did not give you is a map of where commercial, regulatory and buyer facts live, a stopping rule, or permission to report that no answer exists. This page is those three things.
Which are you doing right now
01
Why this matters. A pasted deck can end up in training data, and we cannot get it back. It also risks the agreements that make founders willing to work with us in the first place.
This is the only thing on this page that is not a judgment call.
No decks, no data rooms, no protocols, no financial models, no unpublished results, no company name attached to a nonpublic fact.
On personal plans, including paid ones, training on your content is opt-out, which means it is on until you switch it off. Go switch it off today, and check it again at the start of each term. That shortens how long your chats are kept, and it does not replace the rule above. Company material only ever goes in an account covered by an agreement with the organization. Section 02 says which plans those are.
The way around it keeps nearly all the value, and it is a research skill in its own right: ask about the type of thing, never the company.
Who should [Company] sell their sepsis algorithm to first, and what is the buying process at UPMC?
Who typically holds budget authority for clinical decision support in a large academic health system, and what does the approval path look like from pilot to enterprise contract?
What regulatory pathway should [Company] use for their nerve-regeneration scaffold?
How have resorbable nerve conduits been classified and cleared historically, and what distinguishes the ones that went De Novo from the ones that found a predicate?
Where it stops. If you cannot ask the question in general terms without losing the point of it, it does not go to AI at all. Ask a person instead.
02
Why this matters. Skip these and you retype the same context every morning and hand-edit generic answers all semester. Twenty minutes here saves hours a week.
Four layers control what you get back, and they stack. Most people only ever use the fourth.
A standing block of context applied to every new conversation. In Claude it lives in Settings, under Profile. The other major tools have an equivalent under personalization or custom instructions. Three kinds of content belong here.
I support a program for pre-commercial life sciences startups that pairs fractional experts and interns with founding teams to close two or three strategic gaps before a funding inflection point. My work is commercialization research: market and buyer discovery, competitive landscape, regulatory pathway, evidence requirements, pricing and reimbursement, investor narrative. How to write: No em dashes. Prose over bullet lists in analytical work. Sentence case. State a position once and trust me to have read it. Always: - Flag which claims came from a source I gave you and which you inferred. - Mark anything you cannot verify as unknown rather than estimating it. - List decisions you made on my behalf at the end. - Do not produce a document until I ask for one.
Keep out of instructions: company names, nonpublic facts, and anything that will be stale in a month.
You can set a response mode from a menu in some tools, but the menus move and get renamed. Asking for it in the prompt works in every tool on every plan, so the wording to type is given for each option below.
| What you want | Claude | ChatGPT | Gemini | If there is no menu, type this |
|---|---|---|---|---|
| Teach me, do not answerStart here for a subject you will work in for months | Learning style | Studyor type @study | Guided Learning | The full wording below. Asking for Socratic teaching alone is too vague to work. |
| Answer plus the reasoningWhen you need to understand it and the clock is running | Explanatory style | No preset | No preset | Explain your reasoning as you go, including what you considered and rejected. |
| Short answersWhen you already know what you need | Concise style | No preset | No preset | Answer in three sentences or fewer, no preamble. |
| Formal registerFor something going outside the building | Formal style | No preset | No preset | Write this formally, for an external audience. |
| Match our voiceOutput that already looks like our documents | Custom style | Custom instructions | Gems | Paste two or three things we have written and ask it to match the style. |
Two things that will trip you up. In ChatGPT, Study is not available inside Projects or temporary chats, so if you have followed layer 3 and set up a project, you have to leave it to use study mode. In Claude, the Learning style is being folded into Skills, so the menu may have moved or gone. Both are reasons the last column exists: typing the instruction works everywhere, every time.
The full wording for teach-me mode. Asking for Socratic teaching does not work well on its own, because the useful part is in the specifics: one question at a time, tell me which step was wrong, quiz me at the end.
Paste this at the start of the conversation:
For this conversation, teach me instead of answering.
Do not give me the conclusion. Ask me one question at
a time, start from what I already know, and make me
reason to the answer. When I am wrong, tell me which
step was wrong rather than correcting the whole thing.
Before we finish, ask me to explain it back and tell
me what I got wrong.
Topic: [the concept]
What I already know: [one or two sentences]
This is the behavior to reach for whenever you are learning a domain rather than looking something up. Being made to produce the reasoning is what makes it stick, and the reason to accept the slower path is that you will be in this domain all semester. Two other instructions worth knowing: explain your reasoning as you go, including what you ruled out when you want the answer plus the logic on a deadline, and answer in three sentences or fewer when you know what you need.
Four toggles, and when to use each. Turn web search on any time you need something you can check, because it is the only mode that hands you links to open. Turn extended thinking on for analysis and off for lookups, where it costs time without helping. Save Research for a first pass at a whole landscape, never for a single fact. Check what memory is carrying before you rely on it, because it will quietly reuse an assumption from three weeks ago.
A project is a container with its own instructions and its own attached files, and every conversation started inside it inherits both. This is the layer that stops you re-explaining the workstream every morning.
Set the three layers above first. Section 07 covers wording, and it does less work than any of them.
Almost everything above works on a free account. Two things do not, and one of them is the reason the confidentiality rule exists.
| Capability | Free | Personal paid | Org plan |
|---|---|---|---|
| Instructions, account-wide | Yes | Yes | Yes |
| Socratic and explanatory behavior by prompt | Yes | Yes | Yes |
| Saved instructions you can reuse | Check plan | Yes | Yes, shareable |
| Extended thinking | Yes | Yes | Yes |
| Web search | Yes | Yes | Yes |
| Memory across conversations | Yes | Yes | Yes |
| File creation and code execution | Yes | Yes | Yes |
| Projects | Capped | Unlimited | Unlimited, shared |
| Research, the long multi-step mode | No | Yes | Yes |
| Usage before you hit a wall | Lowest | Roughly 5x free | More than paid personal |
| Training on your content | Opt-out | Opt-out | None by default |
| Admin controls, SSO, retention settings | No | No | Yes |
The highlighted row is the one that matters. On free and personal paid plans, training is opt-out, so it runs until someone turns it off. On organizational plans it is off by default and admins control retention. That distinction, not the feature list, is what determines whether company material can go into a tool. Personal paid at roughly $20 a month buys usage and Research. An organizational seat at a similar per-seat price buys the data posture, which is the thing the workaround in section 01 exists to compensate for.
Tiers and prices change. Confirm at claude.com/pricing before quoting any of this. Last checked August 2026.
03
Why this matters. This is where the hours go. Days get lost to questions with no decision behind them, no time limit, and no permission to come back empty.
Each step has a time on it. Stopping is one of the steps, with its own place in the sequence, rather than what happens when you give up.
The question bounds the work. Answer the question you were asked, not the more interesting one you find on the way.
Ask what the standard of care is, how this category gets bought, who signs, what a specialist would call this, which agency or database governs it, what the failure modes have been for similar products. Going to Google before you have the vocabulary is what produces the flailing, because you cannot tell a real source from an SEO farm on a topic you cannot yet speak about.
Run the same orientation question in two tools. It costs nothing, and disagreement between them is a reliable signal that the ground is soft. Agreement is not proof of anything, but divergence tells you where to be careful.
Free tiers cap out, so spend your good prompts here. Do not spend them asking a model to write the competitive landscape section.
Google is not one of the three routes. It is how you reach a primary source you have already identified, which is why it sits inside the teal box rather than beside it.
AI can tell you where to look and what to call it. It cannot tell you the answer if the answer is a number, a name, a citation, or a current fact.
This is the protocol, not an admission. The note takes three minutes, and the last line is the one that matters: it converts the ask from "please do this for me" into "please confirm or correct my read," which is faster to answer and better for you.
Question: one sentence Tried: sources opened, terms used Found: what, if anything Stuck on: the specific thing My guess: next step, even a bad guess helps
Post in [#channel] and tag [name]. A time box only works if someone answers quickly.
You have the answer and you verified it. Every number, date, name and status in your output has its own opened source.
The answer exists but is not public. It is proprietary, behind an institutional subscription we do not have, or held only by operators. Write down where it lives and who would know.
The answer does not exist yet. No precedent for a category nobody has taken to market. No comparables for a first-in-class approach. This is a finding, and often the most useful one, because it reframes the milestone.
You are tired and thirty tabs are open. That is not a stopping point. That is the 0:45 step, arrived at late.
Because the question bounds this work, you are done when the question is answered. Resist widening it.
04
Why this matters. Every error below sat in a deck that several people had already read. They survived because nobody checked a definition, rebuilt a number, or read two slides against each other.
Five real cases from Launchpad reviews, with names changed, covering market sizing, evidence, regulatory, buyers and our own internal work. Look at each claim before you open the answer.
Case 01 / clinical software
An index is a ratio. A length of stay index of 1.15 means roughly fifteen percent above an evidence-based benchmark. It is not 1.15 days. The company's own brief cited 0.675 days elsewhere in the same materials.
The headline value proposition was overstated by about seventy percent, and everything downstream of it was built on the error.
Why AI misses this: carry the founder's phrasing into your question and you get the founder's error back, restated with more confidence. Google returns a definition of the term. Only opening that definition catches it.
The ruleAny metric with a proper name gets its definition opened before its number gets used. Index, rate, ratio and score are the words that should slow you down.
Case 02 / therapeutics
Four, and every individual input was real. A global case count multiplied by a US-only price. It assumes every single patient is treated, with no share taken by anyone else. A competitor's actual selling price applied to a product with no approval and no clinical data. And a comparable borrowed from a different mechanism, target and disease to validate an unproven program.
The same review found a competitor approved the prior year that set an efficacy bar the deck never mentioned.
The ruleA market size is a construction, so re-derive it rather than checking it. Ask what each factor is, where it came from, and what it assumes about geography, how many patients you would reach, and what you would be paid. A number nobody can decompose is a claim, and it goes in quotation marks with an attribution.
Case 03 / medical device
About twenty minutes, then stop. No agency, payer or health system has yet made a decision about a product that has not been used this way, so the answer is in no database. It lives in the working knowledge of people who have taken products through the same transition.
The right output states the question precisely, establishes that it is not publicly resolvable, and names the kind of specialist who could resolve it. The wrong output takes six hours and produces a folder of tabs.
The ruleWhen the question is how an existing rule will be applied to a novel situation, the answer lives in someone's experience. This holds the same way for regulatory classification, insurance coverage, purchasing and contracting. Recognizing it fast is the skill. Searching harder does not help.
Case 04 / digital health
An executive at the target buyer holds an equity position in the seller. That is a conflict of interest in purchasing, and it means the first customer in the plan is the deal least likely to survive a compliance review. It also means the validation the company cites from that system is not independent.
Nothing here required a database. It required reading two slides against each other and knowing that health systems have conflict-of-interest policies.
This is the category of finding AI is least likely to surface, because each slide is unremarkable alone. The problem only appears when you hold both at once.
The ruleRead the team and the go-to-market slides against each other, and check named advisors against named targets. Conflicts like this show up in the org chart long before anyone says them out loud.
Case 05 / internal memo
The platform has had a dependency column with finish-to-start modes and blocked-task tracking for years. The support page cited in the memo describes the feature. The author had been using the tool daily and would have known if he had opened the column.
The memo also cited a competitor's review site and an unrelated press release as neutral evidence.
The lesson is narrower than "do not use AI to think." Nobody here would let a founder cite an unverified AI-generated number in a pitch.
The ruleHold your own output to the same bar you would hold a founder to. The word verified means you opened it yourself. Claiming diligence you did not do is worse than not doing it.
05
Why this matters. Most wheel-spinning happens because you do not know a source exists. None of this is in the literature you were trained on, so you can search all afternoon while the answer sits in a free federal database.
Arrive with a question, filter, take the link. Deep links go stale, so if one fails, search the source name plus what you need.
Nothing matches. Try a broader term, or ask in the channel. If a source for this does not exist, that is a finding worth logging.
Syndicated market reports with a headline growth rate and no stated methodology. AI-generated content farms. A competitor's comparison page. A press release cited as neutral analysis. Segment definitions a practitioner would not recognize are the fastest tell that a report is repackaged rather than researched.
06
Why this matters. A number you pass along ends up in a deck and gets said out loud to an investor. It traces back to you, and all four failures below look correct until you open the page.
Check the citation is too vague an instruction, because people look for the first failure and stop.
The link resolves, the source is real, and it says something near the claim rather than the claim. The most common failure with search-grounded tools and the hardest to catch, because everything looks right until you read the paragraph.
The source says it, but the source is citing someone else. You have cited the messenger. Your number is now three steps away from anyone who measured anything.
The citation supports the sentence but not the figure inside it. A paper on complication rates gets cited for a prevalence number that appears nowhere in it.
The link resolves to a market report landing page with a headline growth rate, no methodology and a four-figure paywall. Technically a citation. Worth nothing.
Every number, date, price, name and status in a deliverable gets its own opened tab. If you cannot open it, it does not go in. If you cannot open it and it matters, that absence is the finding.
1 in 277
PubMed-indexed papers in early 2026 contained at least one fabricated reference, up from 1 in 2,828 in 2023. The fabrications were typically not obviously defective, which is what made them hard to catch.
Columbia-led analysis of ~2.5M papers, The Lancet, May 2026
That is peer-reviewed literature that passed three to five expert reviewers. You are not more careful than that by default.
07
Why this matters. The difference between a wasted afternoon and a usable answer is usually decided in the first message you send.
Worked out from how this program's work gets done. All of them assume the three layers in section 02 are already set up.
Several of them involve giving a tool a document to work from. Every one of those means a public document, a published paper, a standard, or something you wrote. It never means a portfolio company's deck, model, protocol, dataset or data room, and stripping the company name off does not make it safe, because the surrounding detail identifies it anyway. When a pattern would need a company document to work, you do that reading yourself and bring the tool a category-level question instead.
08
Why this matters. An unmarked guess becomes a fact the moment somebody copies it. Saying plainly what you could not confirm is what makes your work safe for a founder to act on.
Write the limitations as carefully as the findings. It also turns six unproductive hours into something the next person can use instead of six wasted ones.
Where a value could not be verified, the field carries one of these. It never carries a guess.
Not statedConfirmNone publicly statedOne markdown file per project, in the project folder. Append, never overwrite. Six lines per entry. This is the artifact that makes a dead end valuable, and it means the same question in four months is a file lookup rather than a repeat.
## [date] | [question in one sentence]
Serves: which decision
Searched: sources opened, terms used
Found: claim + link + date accessed
Not found: what does not appear to exist
publicly, and where you looked
Open: what remains, and who could answer