Skip to content
Gregor Biswanger edited this page Aug 3, 2026 · 2 revisions

FAQ

Do I need to install anything?

No. FeatherSpec is Markdown and folders. No CLI, no Node or Python package, no build step, no dependencies, no lock file. You copy files into a repository and open it in your assistant.

Does it work with tools other than Claude Code and GitHub Copilot?

Partly, and it degrades gracefully. AGENTS.md is a widely-adopted convention, so most agentic tools will read the constitution. The /sdd-* commands depend on each tool's own slash-command mechanism — if your tool has one, add a loader for it the way .github/prompts/ does for Copilot; the workflow bodies themselves are plain Markdown with no tool-specific syntax. That is deliberate: see Design Decisions.

Why are there no skills?

Both tools read every skill's name and description up front so the model can decide to load it — which means nine workflow descriptions in the system prompt of every request. These workflows have side effects and should run only when you type them, so they are commands instead. Full reasoning: Design Decisions.

Can I use it in a language other than English?

Yes. /sdd-setup asks for your documentation language first and stores it as DocLanguage in AGENTS.md. From then on every interview, report, spec, plan and Memory Bank file is written in that language. FeatherSpec's own wiring stays English.

Do I have to use both tools?

No. Using one costs you nothing — the other tool's entry points are a handful of one-line files. If you want them gone entirely, Committing to One Tool has the mechanical conversion.

How is this different from just writing a good prompt?

A prompt is gone when the session ends. A spec, a plan and a Memory Bank are files in your repository: reviewable in a pull request, readable by your teammates, resumable next week, and usable by a different assistant. The traceability table also gives you something a prompt never can — when a requirement changes, you can see which steps, which code paths and which tests that change reaches, before you edit anything.

Does this make me faster?

Probably not on day one, and the honest framing matters here. Spec-driven development is a variance-reduction technique, not a velocity technique: fewer confident wrong turns, a written record of why the code looks the way it does, and a chain from every acceptance criterion to the test that decides it. Trust your own measurements over your own impressions — developers are documented to be quite bad at estimating this, in both directions.

Do I have to do all this for a typo?

No, and doing so would be a mistake. AGENTS.md says work spec-first for anything that changes behaviour — deliberately not always — and documents a fast path for changes smaller than the spec that would describe them. The only rule is that the assistant says it is taking the fast path, so you can disagree.

A workflow that demands three Markdown files for a config value teaches people that the workflow is theatre, and then they stop using it for the changes that actually needed it.

Why is /sdd-clarify a separate command instead of a better interview?

Because the context that produced an ambiguity is structurally the worst placed to find it. /sdd-specify checking its own output is a reviewer grading its own homework — it fills the gap from the conversation it just had, which is exactly what makes the gap invisible. /sdd-clarify reads the spec cold, with no conversation history, and that is its only real advantage. It also deliberately lists problems instead of resolving them: a model asked to resolve an ambiguity resolves it silently with whatever is statistically plausible.

Does the assistant follow this automatically, or do I have to type commands?

Both. The constitution in AGENTS.md is loaded in every session, so the spec-first behaviour, the style preferences and the automatic architecture/memory sync apply without you asking. The /sdd-* commands are the explicit workflows — they never fire on their own, by design.

What if my spec turns out to be wrong halfway through?

Change the spec, then run /sdd-plan again. It detects that the spec changed (Mode C), reads the traceability table in reverse, and reports which steps and which files the change reaches before editing anything. New steps get appended; existing step IDs are never renumbered, because they are referenced from commits and reports.

Can I add my own commands?

Yes — two Markdown files, no registration. See Extending FeatherSpec.

Why is Claude Code's auto memory disabled?

It is a machine-local note store your teammates never see and Copilot cannot read — effectively a second, invisible Memory Bank. FeatherSpec keeps one visible source of truth. Set "autoMemoryEnabled": true in .claude/settings.json if you want it back.

Is .claude/settings.local.json supposed to be missing?

Yes. It is git-ignored on purpose — Claude Code writes it when you grant a permission with "don't ask again", so it is personal to your machine. Anything the team should share goes in .claude/settings.json. See Configuration.

How big should activeContext.md get?

Two screen pages, maximum. It is a handoff context, not a changelog. If it outgrows that, the surplus belongs in a more permanent file — Memory Bank has the table of where each kind of content goes.

Where do I report a problem?

https://github.com/GregorBiswanger/featherspec/issues

Clone this wiki locally