DISCUSSION COMPANION · 28 SEPTEMBER 2026
Questions worth asking.
Short answers for the room. The field guide explains the basics; the technical reference supplies file paths and implementation limits.
What is Lintel actually adding?
A reusable engineering method, project knowledge and inspectable verification around the agent you already use. The strongest case is preserving your team’s decisions, constraints and lessons across work, people and sessions. Generic instructions to “think carefully” are weak differentiation.
Does it guarantee quality?
No. It makes the intended method, expected depth and acceptance checks explicit. A check only adds evidence when it runs on the relevant result. Missing checks, stale review evidence and unavailable independent review must remain visible. A fluent explanation is not a passed acceptance test.
Why not just use AGENTS.md and native skills?
That is the right baseline. Give both alternatives the same project facts, tests and CI. Lintel has to earn its overhead through its methods, coordination, reusable content and recovery. For a small task in a well-maintained repository, the simpler baseline may be enough.
Does “universal” mean every feature works everywhere?
No. Current source names four supported families — Copilot, Claude, Codex and Cursor — across 13 named surfaces, plus a manual other route. Fifteen core wrappers expose a starter set; 96 canonical skills are the full library. Delivered discovery files, vendor documentation and observed host execution are separate. Most surface records remain not_run.
Read the exact client matrix before promising a customer a capability.
Are the safety rules enforced?
Instructions explain policy; host permissions, platform controls and CI enforce real boundaries. Compatible Claude Code registration can run Lintel’s hook bundle. The portable client adapter does not translate or install that bundle. The same file on another host is not evidence of equivalent enforcement.
What changed about reviews?
The shared Review Method asks for a bounded input packet, applicable obligations, coverage, findings and uncertainty. Evidence is bound to the inspected content and work context. Changed content or a later applicable rejection can invalidate an older pass. A digest does not authenticate who performed the review.
Is MARS just more agents?
MARS means Multi-Model Adversarial Review & Screening: distinct models independently inspect the same packet, with a challenge round when claims remain contested. The coordinator adjudicates against evidence, not majority vote. It requires consent and host-verifiable model selection; the result is advisory and cannot provide release clearance.
Swarm divides implementation into attributable lanes under one coordinator. MARS adds review perspectives. Neither turns self-review into independent review. No live MARS panel was executed for this presentation.
Where does the knowledge live?
Personal setup and packs normally live under the resolved Lintel home. The repository holds its kit and deliberately committed plans, decisions and lessons. Local runtime is ignored and does not travel with an ordinary clone. The .claude/ project storage convention does not require Claude.
See the storage map. A fresh session must read the durable records and reconcile them with the actual checkout.
What should a customer profile change?
The questions, permitted stack, required checks, review obligations and handoff. If the only visible difference is the greeting, the profile is not earning its place. Role preferences and mandatory policy are different inputs; a missing required pack must not silently become a neutral profile. A pack is not a substitute for configured platform permissions.
Can it help someone new to Copilot or vibe coding?
Potentially: one small request, a visible plan, a focused acceptance check and a recoverable next step can teach good habits while producing something useful. Keep the first working set small. Measure the time to an accepted result and how often a colleague has to rescue the work; more process is not automatically better.
Did the kitchen test prove that Lintel wins?
No completed controlled three-arm comparison is available in the inspected evidence. Protocol v2 and exact inputs are published here. An earlier timeout, uncontrolled demonstrations and reference/control calibration are different evidence, not three comparable participant results. Unknown billing stays unknown.
The Naked-versus-Lintel comparison also changes delegation and model mix; it does not isolate orchestration alone. Serial Lintel versus Lintel + Swarming is intended to hold the method and review policy constant while changing ready-package scheduling.
What do the GitHub screenshots prove?
The visible activity totals and the browser’s page-unresponsive dialog. They illustrate Johannes’s personal experience after enabling swarming. They do not establish a speedup, quality gain or a GitHub outage. The useful lesson is to keep parallel work bounded and delivery reviewable.
What can people take away today?
Lintel’s public repository, this presentation, its reference and the kitchen protocol. The CAIP pack’s distribution link is pending. Core and Benchmark are planned packages with links to follow. The current reference follows 0.12.0 unreleased source, not a claim that a tagged release or all client acceptance is complete.
What is a sensible team pilot?
Choose a repeated, bounded task in a non-production repository. Compare a strong native baseline with Lintel using the same facts and acceptance checks. Record accepted outcomes, interventions, residual defects, recovery, total time and actual cost. Repeat, include failures and negative results, and remove process that costs more than it helps.