> practices
Practices
Atomic, dated best-practices — each a when X → do Y (because Z)
unit tied to a guide section, with how to verify it still holds. Machines
read the same corpus at
/practices.json and
/llms.txt.
- When
- Pricing an AI-assisted feature for a developer product, or writing the pricing page and launch messaging for one.
- Do
- Expect and anchor to the category's two-part shape — a per-seat license the buyer can budget, plus metered model consumption — and publish all three numbers on the page — the seat price, the meter's unit rate, and the monthly cap or included allowance. Don't let the meter be discovered on the first invoice.
- Why
- By mid-2026 devtools converged on this tariff across generation and review — GitHub Copilot Business/Enterprise at $19/$39 per user with AI credits that drain by token usage (usage billing since 2026-06-01), GitHub Code Quality's GA at $10 per active committer plus usage-based AI detection and Autofix (2026-07-20), CodeRabbit Pro at $24/dev/month plus on-demand credits. A developer audience prices in the meter's tail risk; showing the seat alone reads as hiding the real cost, and an uncapped meter is the same trust failure as a silent deprecation.
- Since
- GitHub Code Quality GA (2026-07-20) extended the seat-plus-meter shape from code generation to code review, making it the category default rather than a Copilot quirk.
- Verify
- Check the current pricing pages for GitHub Code Quality, Copilot, and CodeRabbit still show seat-plus-meter; if a major entrant has moved to flat per-seat pricing with no meter, the convergence claim is weakening.
The numbers are the edge: a bare model will say “be transparent about pricing,” but not that the AI-devtools tariff settled on seat + meter in mid-2026, which specific prices anchor the category, or that the cap is the third number a skeptical developer looks for.
- When
- Building a landing page, a pricing page, or launch messaging for a developer audience.
- Do
- Swap social proof a visitor must take on faith for numbers a stranger can reproduce — live weekly downloads, public star history, a deploy counter — and put a real price on the page with a spend cap where usage-based.
- Why
- Developers independently verify claims, in public. A logo wall proves a contract was signed once; a live metric proves the thing is used, and the skeptic can go check. Pricing opacity reads as something to hide; transparency is itself a trust signal.
- Since
- 2026 DevGTM operator consensus (PostHog, Tinybird, GitBook, Mozilla, APIwiz — Hackmamba roundup)
- Verify
- Every quantitative claim on your homepage links to something a visitor can check; your pricing page shows a number without a sales call.
The test that generates all the tactics: pick your most skeptical prospective user and assume they will fact-check every claim in public.
- When
- Sizing a market, setting awareness targets, or choosing between broad reach and segment depth in a developer GTM plan.
- Do
- Plan against a plateau, not a boom — pick a specific developer segment and go deep rather than spreading spend across "all developers." Re-check the population figures against the latest SlashData wave before quoting them.
- Why
- SlashData's 2026 research puts the global developer population at roughly 47M with growth decelerating to about 10% year over year, aging, and shifting toward South Asia and Greater China. Broad-awareness plays priced for a fast-growing pool stop paying back when the pool flattens; depth in a well-chosen segment compounds instead.
- Since
- 2026 SlashData wave — ~47M developers, growth decelerated to ~10% y/y
- Verify
- Check the current SlashData developer population report for the latest headline figure and growth rate; the direction has been durable, the numbers move each wave.
The plateau is the dated fact; “pick a segment” is the old advice it re-prices. A bare model will still quote population figures and growth rates from earlier waves — carry the current wave’s numbers and the deceleration, not just the headcount.
- When
- Designing or marketing an API where an autonomous agent could evaluate, integrate, or transact on a developer's behalf — anything that sells or provisions something programmatically.
- Do
- Ship quote-then-execute (a quote call returns the exact price and a short-lived token before anything commits), require a per-attempt idempotency key so retries can't double-purchase, record explicit consent for agreements, and serve docs as markdown plus OpenAPI with a quickstart that says "hand this spec to your LLM."
- Why
- Agents moved from reading docs to selecting and transacting in 2026, and the first mass-market platform built for them makes over-purchase impossible by construction — that safety is what makes a developer willing to delegate the transaction at all. There is no card-number field anywhere in GoDaddy's API; the guardrails are the positioning.
- Since
- GoDaddy Developer Platform launch, 2026-07-14 — the first mass-market incumbent shipping agent-first docs and an agent-safe transactional flow as one product surface
- Verify
- Open developer.godaddy.com and confirm the pattern still holds; then check whether an agent given your own OpenAPI spec could complete your golden path unaided, and whether a retried purchase call on your API can double-charge.
The dated shift is the actor, not the reader: a bare model will tell you to publish OpenAPI specs, but not that incumbents are now shipping quote-tokens, idempotency keys, and consent records specifically so an agent can buy — copy the trio before your category’s comparison-shopping layer forms.
- When
- Planning docs work, an information-architecture overhaul, or deciding whether to publish llms.txt.
- Do
- Make "is this page retrievable by an LLM" a first-class docs review question — one section answers one question, golden path up top, descriptive link text — and publish an llms.txt as a curated index. Treat llms.txt as cheap insurance, not the strategy.
- Why
- A growing share of your docs' readers are coding assistants and answer engines, and a developer's first impression is increasingly a machine's paraphrase of your content. Chunk-retrievable structure is what gets you cited — and it helps human readers for free. The llms.txt caveat is now measured, not speculative — Ahrefs' June 2026 study of 137K domains found 97% of llms.txt files receive zero requests, and AI retrieval bots are about 1.1% of what little traffic exists.
- Since
- 2026 — assistants and answer engines mediate developer discovery; Ahrefs study (2026-06-15, 137K domains) quantified that llms.txt itself goes almost entirely unread
- Verify
- Ask a coding assistant a task-shaped question about your product and check whether it cites your docs correctly; re-check the Ahrefs study (or a successor) for whether any major provider has started requesting llms.txt.
The durable investment is the underlying content — accurate, well-structured, answering real questions — which compounds regardless of which indexing standard wins. Ship llms.txt as an afternoon’s cheap insurance, but never let the checkbox file substitute for the docs work: today the checkbox is all it is.
- When
- Planning a launch, an onboarding overhaul, or deciding where the next quarter's developer-marketing budget goes.
- Do
- Instrument the median minutes from signup to first successful API call (or equivalent first success) and make that number a launch gate — fix the path to value before buying more top-of-funnel.
- Why
- For developer products the biggest funnel leak is almost always between "signed up" and "got something to work", and the DevEx research ties feedback-loop speed to adoption. Traffic pointed at a broken first-run experience is spend converting skeptics into detractors.
- Since
- DevEx research (Noda, Storey, Forsgren, Greiler — ACM Queue, 2023) and the 2026-07-06 deep dive
- Verify
- You can state your median time-to-first-success from instrumentation, not guesswork, and the last launch had it as an explicit gate.
Signups measure curiosity; first success measures whether the product kept its promise — only the second compounds.
- When
- Setting DevRel goals, building the quarterly report, or defending the function's budget.
- Do
- Wire an "influenced" flag into the CRM for deals DevRel touched, pair it with leading indicators the team can move (activation rate, docs and repo engagement, community responsiveness), and report on a 6–12 month horizon. Drop sourced-lead quotas.
- Why
- DevRel's returns run through peer trust and organic discovery — exactly the paths last-touch attribution undercounts. A capture-first quota pushes advocates to optimize for capture, which corrodes the trust the function depends on.
- Since
- 2026 practitioner consensus; "presence in AI-generated answers" newly entering as a discovery metric
- Verify
- Your CRM has a working influenced flag, DevRel reporting uses it, and no advocate carries a monthly sourced-lead number.
Leadership pressure to tie DevRel to business outcomes is fair — the answer is better leading indicators, not a false-precision attribution dashboard.
- When
- Reporting adoption of a seat- or license-based AI product — to your own leadership, in a customer-facing admin dashboard, or ahead of an enterprise renewal.
- Do
- Break seats into progression cohorts (first use → advanced/agentic use → orchestration) and name the licensed-but-unengaged segment explicitly as its own line, instead of hiding it inside an "active seats" denominator. If you sell the product, ship this view to the buyer's admin before they build a hostile version themselves.
- Why
- Raw seat counts hide shelfware, and buyers have started looking for it — HCLTech's 500-enterprise survey (with Raconteur, 2026-07) found 90% of decision-makers say GenAI is transforming workflows while only 18% see significant revenue impact, so the renewal conversation now runs on proof, not activity. GitHub shipped exactly this shape for Copilot on 2026-07-22 — three engaged phases plus a named "Passive (licensed but not engaged)" cohort, with per-cohort throughput comparisons. The vendor that names its own shelfware controls the honest number in the room.
- Since
- GitHub's Copilot impact dashboard (2026-07-22) — the first major vendor to name a "Passive" seat cohort in its own admin product — landing the same week as HCLTech's 90%-transformation / 18%-revenue-impact survey.
- Verify
- Open the GitHub changelog entry and confirm the Passive cohort still ships; check whether a second seat-based AI devtool has since adopted phase-cohort reporting (if several have, this is table stakes, not an edge).
The dated shift is who counts the empty seat: a bare model will tell you to measure activation, but not that the biggest AI seat-seller now names its own shelfware in the buyer’s dashboard — because an 18%-revenue-impact market negotiates renewals on that number either way.
- When
- Building an outbound or expansion motion for a self-serve developer product, or evaluating community- and intent-data vendors.
- Do
- Score accounts on roughly 80% behavioral product signals (usage volume, new users added in a 7-day window, in-app engagement) against 20% firmographics, cap the total play list at 10–12 so sales actually runs it, and land raw usage events in a warehouse you control before renting anyone's signal layer.
- Why
- Bought intent data — website visits, keyword tracking — is commoditized; every competitor gets the same feed. First-party product usage is proprietary by definition. Otter.ai reported a 2x jump in outbound pipeline on this weighting, and the vendors packaging these signals keep getting acquired into big GTM suites, so owning the raw events hedges their roadmap risk.
- Since
- Common Room's Otter.ai case study (2026-07-17); Zoom's acquisition of Common Room (announced 2026-07-02) — the third GTM-signal roll-up in eight months, after Clari+Salesloft (2025-12) and Apollo+Pocus (2026-03)
- Verify
- Re-read the Common Room case study for the weighting and play cap; check that your raw product events are queryable in your own warehouse, not only inside a vendor's UI.
The numbers are the edge: a bare model will say “use product signals,” but not the 80/20 behavioral weighting, the 10–12 play cap, or that the signal-vendor layer is consolidating fast enough that renting it is a roadmap risk.