Dogfooding as Product Discipline

Sources#

Summary#

The cross-source claim that product sense is built, not innate — and the way you build it is relentless first-hand use of the thing you ship. Fiona Fung calls it "ant food" (Anthropic's term for dogfooding) and names it the answer engineers kept asking her for: how do you build the product-sense muscle? Use the product until you "feel it in your bones." The same discipline recurs as Cat Wu's lunchtime vibe-checks (Claude Character as Product) and John Glasgow's "I'm in every Slack channel with every customer" (Founder-Led Sales Discipline). When agents make generation cheap, taste is the scarce input — and dogfooding is its training regimen.

Why product sense becomes the scarce skill#

Once coding is cheap (Verification as the New Bottleneck), the constraint shifts to knowing what's worth building and whether it's any good — i.e., product taste. Fung's two engineering profiles she now hires for are "creative builders with product sense" and "deep system expertise." The first is the dogfooding-fed one. This is the org-level confirmation of Engineer PM Convergence and Evals as Product Spec: taste/judgment is the durable human skill as the harness shrinks.

Fung's mechanism: feel it in your bones#

Her recipe for building product sense:

Dogfood — especially for managers, who pre-Claude often had no time in the product. "You really start feeling your product in your bones; you remember what problem you were trying to solve." Without it, "you make product decisions based on metrics, dashboards, or PowerPoints."
Ship and iterate.
Talk to customers — her passion project is onboarding small-business friends (restaurateurs) onto Cowork and watching them struggle: "really humbling to see how many things we could be doing better" in onboarding. (Tied to Anthropic's Claude for Small Business.)

The Mr. Peanut story is the taste-calibration anecdote: she coded a holiday "snowman" Claude in the CLI; her designer said it looked nothing like a snowman — "you turned Claude into Mr. Peanut." Human product/design taste caught what the builder couldn't see — the kind of judgment the model lacks.

Claude as cross-functional gap-filler (the flip side)#

Dogfooding reveals where you're weak; Claude then fills those gaps across all roles: designers on Claude Code make polish/UX fixes themselves via Claude instead of red-lining and handing to engineering (closing the iteration loop); Fung, a self-described verbose engineer, used Claude as a "content design partner" for short copy. "Claude is augmenting all the areas where you may not be as strong." Dogfooding tells you where the gaps are; the tool helps you close them — but the taste to notice the gap stays human.

The Codex twist: degrade your own process to improve the product (Ambrosino)#

Andrew Ambrosino (OpenAI Codex) supplies the most extreme form of the dogfooding contract. The Codex desktop app was "entirely informed by the dogfooding loop" — but with a deliberately uncomfortable inversion:

"There is a desire among all of us to do as much as possible in the app even when it's not the best tool, so that it can become the best tool. We often don't improve our process so that we can make the product better to do it — which is a deeply uncomfortable place to be in."

This flips ordinary dogfooding. The usual bargain is "use your product to find its gaps." Ambrosino's is "refuse the better external tool and absorb the pain, so the gap is felt hard enough to force the product to close it." The team's own workflow is intentionally left worse so the product is pressured to get better — a self-imposed constraint that keeps taste and roadmap fused to lived friction.

It began as a personal dogfooding loop: Ambrosino's goal for the original app was "to make it the thing I wrote the code with." The virtuous cycle — "I can't do this thing → I fix that so I can → now I can do more things" — first ran on him alone, then his own usage shifted from writing code to product-discovery/steering as the app matured, so he re-pointed his usage at that ("now I need to build a spreadsheet that models this, do internal deep research on all the efforts in this area"). Aligning your own usage with the problem the product must next solve is the discipline.

The five faces (cross-source)#

Source	The dogfooding move
Fiona Fung	"Ant food" — managers dogfood to keep product sense; onboard small businesses to feel onboarding pain
Cat Wu	Lunchtime vibe-checks of the model as eval discipline (Claude Character as Product)
John Glasgow	Founder in every customer Slack channel; joins sales calls to drive engineering (Founder-Led Sales Discipline)
Dan Carey	"Use Claude Design to design Claude Design" — build and dogfood your own tool daily; Claude does first-pass analysis of every user conversation
Andrew Ambrosino	Use the Codex app even when it's the wrong tool so it's forced to become the right one; keep your own process worse to make the product better

All five treat direct, repeated, first-hand contact with the live product/customer as the non-negotiable source of taste — and Ambrosino adds that you may have to hold your own process hostage to it.

Connections#

Fiona Fung — "dog food, dog food, dog food… ant food"
Cat Wu — lunchtime vibe-checks as the same discipline in eval form
John Glasgow — founder-in-every-Slack-channel as the customer-contact version
Claude Character as Product — vibe-checks; character quality is judged by dogfooding, not metrics
Founder-Led Sales Discipline — staying close to the customer is dogfooding the sales surface
Evals as Product Spec — evals encode taste; dogfooding is how you acquire the taste evals encode
Engineer PM Convergence — "creative builders with product sense" is the dogfooding-fed hiring profile
Managers as ICs — managers-as-ICs exists so that managers keep dogfooding the codebase
Jagged Intelligence (Ghosts, Not Animals) — the Mr. Peanut catch is human taste catching what the ghost can't
Dan Carey / Claude Design — "use Claude Design to design Claude Design": the developer-tool-improving-itself case, plus Claude doing first-pass analysis of every user conversation
Compounding Loop Optimization — "talk to users, make it the easiest thing in the world" is dogfooding turned into an optimized, Claude-assisted loop step
Andrew Ambrosino — the "degrade your own process to improve the product" and personal-dogfooding-loop variants
Implementation Abundance Inverts Product Work — dogfooding is how the curating taste that becomes the bottleneck is trained
Codex — the app "entirely informed by the dogfooding loop"

Open Questions#

Dogfooding works when the team is the user (Claude Code) or near it (Cat Wu, Boris). How do you build product sense for users very unlike you — does "talk to customers" fully substitute, as Glasgow/Fung's small-business work suggests?

Resolved Questions#

Can dogfooding scale, or does it implicitly cap how large an AI-native product org can stay taste-driven before it reverts to dashboards? Answered: The Orchestrator's Real Workload: Decision Burden, Framing Discipline, and Whether Taste Scales — false binary: dogfooding itself never scales (first-hand use is per-person and breaks when the team stops being the user), but the taste it produces scales through three named mechanisms — encoding into runnable artifacts (Evals as Product Spec), concentrating the rare-trusted-evaluator role plus vibe-check rituals rather than diffusing taste with headcount (Claude Character as Product), and AI-extended contact surface (Carey's Claude-first-pass on every user conversation). The cap variable is not org size but team-user distance plus encoding discipline: an org reverts to dashboards when it stops converting felt use into evals and rituals, at any size.

Sources#

Running an AI-native engineering org
How Anthropic's product team moves faster than anyone else | Cat Wu (Head of Product, Claude Code)
The ERP for the AI Revolution is here
Designing with Claude: From prompt to production
OpenAI Codex lead on the new shape of product work — Ambrosino: "we often don't improve our process so that we can make the product better"