Glossary, Saquif Alam

Build trap: The condition Melissa Perri named in her 2018 book, where a team measures itself by features shipped instead of behaviour produced, and quietly stops asking whether the plaza it just paved is the one anyone actually wants to walk across. The version that shows up in enterprise marketing software is more polite than the startup version, because the roadmap looks busy and the release notes read well, and the trap only reveals itself when the renewal call goes quiet six months later.
Desire path: The dirt track that gets worn across a Dhanmondi park when the official walkway points the wrong way, used here as the operating metaphor for what users do inside a B2B product the moment the formal workflow is too long. The PM's job, in this reading, is to read the worn grass before paving anything, and the team that paves a plaza without first watching the path is shipping for the dossier, not the user.
Kill ratio: The share of in-flight features killed before they reach general availability, treated as a leading indicator of whether the experimentation culture is actually working. A team with a kill ratio near zero is not winning, it is filtering ideas before testing them, which is the opposite of what the experiment platform was bought to do.
Eval set: The curated bank of inputs and expected outputs against which a model or agent is graded after every change, the way a compiler suite gets a regression run. The argument in the post is that the support queue, read for six months and tagged honestly, is already a better eval set than anything a PM will manufacture in a Notion doc on a Tuesday afternoon.
Agentic harness: The scaffolding around a language model that turns one-shot completion into a workflow, including the orchestrator, the tool registry, the eval loop, the observability layer, and the retry policy that catches the model when it tries to invent a customer record. Without that scaffolding, the model behaves as a parlour trick that occasionally produces useful text, and the version with the scaffolding behaves as a junior coworker who still needs an attentive manager sitting two desks away.
Vibe coding: Andrej Karpathy's term for the workflow where a person describes intent in plain English to a coding agent, accepts the diff if the app still runs, and only reads the code when something breaks. The version Saquif practises is a stricter one, because the harness is doing the heavy lifting and a senior engineer is reviewing the PR, but the original looseness of the term is doing real work in describing how product owners now ship their own prototypes.
Headless content: Content stored as structured data with no opinion about where it eventually renders, the way a JSON record sits behind any number of channel-specific surfaces. The contrast in the post is with token-based design systems, because tokens carry visual decisions while headless content carries semantic ones, and the failure mode of confusing the two is what produces a brand site that ships beautifully and reads like a settings menu.
Prompt-eval-observability stack: The three-layer kit Saquif treats as the actual PM stack for agentic features, replacing the older mock-up-plus-PRD pair. Prompts now carry the spec, evals carry the acceptance criteria, and the observability layer is the production telemetry that tells you whether yesterday's spec is still true or has quietly become fiction since the last model swap.
Type schemas not tokens: The argument that the durable contract between a content platform and the agents writing into it is a type schema, not a brand-voice prompt or a token theme. Schemas survive the next model upgrade because they describe shape, while tokens and prompts have to be re-tuned every time a vendor ships a new checkpoint.
Build to learn: The pre-PMF mode where the artefact's only job is to produce information about whether the bet is real, so the team writes throwaway code on purpose and treats the demo as a sensor rather than a product. The mistake the post names is carrying this mode into the post-PMF business, where the same instinct produces an unmaintainable codebase and a quarterly NPS dip nobody will admit is causal.
Build to earn: The post-PMF mode where the artefact's job is to produce revenue and renewals, so the team grades itself on durability, latency, and the boring telemetry that protects the customer base it already has. The trick the post argues for is running both modes in parallel inside the same company without letting either side write the other side's review.
Cornered resource: Hamilton Helmer's name in 7 Powers for an asset the rest of the market cannot replicate at any price, used in the post as the cleanest way to explain to a CFO why the existing enterprise customer base is also the cheapest possible discovery substrate. The startup testing the same hypothesis has to acquire the customer first, while the incumbent already sitting inside the workflow can test the next adjacent feature for the price of a feature flag.
Customer base substrate: Saquif's coinage for the renewable asset that the existing enterprise customer base becomes once you treat it as a discovery surface, not just a revenue line. A platform running its experiments inside that substrate is paying for the moat and using it, while a platform running the same experiments in a detached sandbox is paying the same cost and quietly leaving the moat's primary advantage on the table.
7 Powers: Hamilton Helmer's catalogue of the seven durable advantages a business can hold, including scale economies, cornered resources, switching costs, network economies, branding, counter-positioning, and process power. Saquif treats it as the most useful taxonomy still in print for testing whether a strategy slide is making a real claim or a hopeful one. Most of his post-mortems on losing categories trace back to a mis-named power on page two of the deck.
Agentic A/B test: An experiment whose treatment arm is not a static variant but a small agent loop that adapts its behaviour mid-flight inside guardrails set by the PM. The eval and observability layer has to be richer than a normal A/B because the variant is moving, and the kill criterion has to be written before the test starts or it will never be written at all.
Brief-to-publish workflow: The end-to-end loop inside an enterprise marketing team that runs from the campaign brief through draft, review, approval, channel adaptation, and finally publication across a dozen surfaces. The reason it matters as a term is that the platform already sitting inside this loop has a structural advantage over any startup trying to insert itself at one stage, which is the practical form the cornered-resource argument takes in MarTech.
MVP: The minimum viable product, treated in the post not as a fixed definition but as the residue of a political fight between engineering, design, sales, and the customer success lead about which corners are safe to cut this quarter. The version a startup ships and the version an enterprise PM ships share a name and almost nothing else, because the enterprise MVP has to pass a procurement review the startup version has never heard of.
RAG: Retrieval-augmented generation, the pattern where a language model is given relevant documents at inference time so its answer is grounded in a known corpus instead of its training data. The field guide reads as a long argument that most production RAG failures are retrieval failures wearing a generation costume, and the fix is almost never a bigger model.
Retrieval grounding: The part of a RAG system that decides which chunks of the corpus get attached to the prompt, and therefore which facts the model has any chance of getting right. The post argues that the failure mode nobody wants to name is the retriever fetching plausible-looking but irrelevant context, which the model then dutifully cites as if it had been asked the right question.
Structured output: A model response constrained to a declared JSON schema, so the downstream system can rely on field shape without parsing prose. The contract is doing more work than people credit, because once the schema is the spec, the prompt becomes documentation and the eval becomes a schema-conformance test alongside a quality test.
AI Overview: The summarised answer that an answer engine renders above the blue links, synthesised from a small set of grounded sources rather than from the open index. The practical implication for any content site is that the page has to be both readable to a human and parseable to a retriever, because the citation in the overview is now a meaningful share of the traffic the page will ever see.
Deprecation grief: The five-stage emotional process that ships every time an enterprise platform retires an API or a UI surface, starting with denial in the marketing team and ending with bargaining in the support queue. Naming it as grief is the post's only real intervention, because once a team admits the project is bereavement work as much as engineering work, the migration plan starts including the things that actually slow it down.
Spec half-life: The interval over which half the assertions in a written spec stop being true, measured against the running code rather than against the author's intent. For a pre-PMF feature in a fast-moving agentic stack, the half-life is closer to a fortnight than a quarter, which is why the post argues for letting tests and evals carry the contract instead of the document.
Diagram versus document: The forced choice between a diagram that gets read in a meeting and a document that gets read alone at night, treated here as two different artefacts of the same PM job rather than substitutes. The post argues that the diagram earns alignment in the room while the document earns durability in the wiki. The team that picks only one ends up either with a great whiteboard photo nobody can act on, or a Notion page nobody will open after the first week.
Dhaka distributed system: Saquif's framing of the city itself as a distributed system, with bKash agents as edge nodes, the traffic on Kemal Ataturk Avenue as the network layer, and the chai stalls in Karwan Bazar as the eventual-consistency repair shop. The point of the metaphor is to make the engineering reader notice that the city has solved problems the textbooks are still arguing about, mostly through informal protocols nobody wrote down.