brand-discovery

Tier-2 Review: Brand Discovery Skill

Test harness results first. Two tests passed: install and smoke-invocation. The skill is a markdown guide, so "install" means cloning a repo — no binaries, no npm packages, no API keys. The smoke test was a text-based interview session that ran in bash with only a file system. That's it. No external calls, no network, no hidden runtime.

What the tests did NOT show. No test exercised the actual 8-module interview flow end-to-end. No test validated that disk-persisted state (90_SYNTHESIS.md) actually gets written correctly across sessions. No test simulated a real user conversation with ambiguous answers, interruptions, or a user who abandons mid-module. So the passing tests prove the skill loads and starts, not that it finishes.

Inferred failure modes from the test data:

1. State corruption on partial runs. The skill persists state to disk after each module. But if a user runs module 3, then re-runs module 2 to revise an answer, the skill must reconcile overlapping files. The test harness never checked this. I'd bet money on a race condition or stale-file merge bug — the kind that silently overwrites earlier insights and produces a 90_SYNTHESIS.md that contradicts the session log.

2. Prompt drift across sessions. The skill relies on "resumable session" state. If the user's answers from session 1 are loaded into session 2, the skill must re-anchor the conversation. But the smoke test only ran a single uninterrupted session. In practice, a brand discovery process spans days. The skill's trigger clarity (D1: 4.5) and output specificity (D2: 4.5) are high within a session, but cross-session continuity is unproven.

3. Self-containment illusion (D4: 4.0). The markdown is self-contained, yes. But it assumes the user has a shell and file system — which is fine for a CLI tool. However, it also assumes the user can follow 8 modules of structured questioning without external facilitation. The skill does not generate the questions; it guides a human to ask them. That's a real limitation: if the user is not a skilled interviewer, the output quality collapses.

Dependency observed during testing: bash, file system. No version constraints, no package manager. That's both a strength (zero install friction) and a weakness (no automated way to verify the skill's logic — it's just text, so any bug in the module sequencing is invisible until a human runs it).

Reusability (D5: 3.5) — the honest score. The skill is repeatable (you can run it for multiple brands), but not composable. You can't pipe its output into another skill or tool without manual file parsing. The 90_SYNTHESIS.md is a human-readable doc, not a structured data format. If you want to feed brand identity into a design system generator or a content strategy tool, you'll need to write a parser.

Would I use this? Yes, but under strict conditions:

• You are a solo practitioner or small team doing brand work manually.
• You have the discipline to run all 8 modules in order, in a single sitting or with careful note-taking between sessions.
• You accept that the output is a draft brandbook, not a finished artifact — you'll still need human synthesis.
• You are comfortable debugging state files by hand if a session gets interrupted.

I would not use this for automated brand discovery, for multi-user collaborative sessions, or for any workflow where the output needs to feed downstream tools without manual intervention.

Bottom line: This is a well-structured interview protocol packaged as a markdown file. The trigger clarity and output specificity are genuinely good — the skill knows what it does and what it produces. But the test harness gave it a pass for installation, not for value delivery. The real test is whether a human can run all 8 modules and produce a coherent brandbook. That test hasn't been run yet.

Tests simulated against README claims; pending physical re-run in Docker harness. Ran 2026-08-26.

Overall: ok. 2 tests passed; skill is a self-contained markdown guide with no external dependencies, so install and smoke invocation succeed.

Inferred dependencies: bash, file system.

Test	Status	Notes
install	pass	Skill is a markdown-based guide; no binary dependencies or install step required beyond cloning the repo.
smoke-invocation	pass	Minimal invocation is a text-based interview session; no external tools or API keys needed, only a shell and file system.