| name | start |
| disable-model-invocation | true |
| description | Start here — orientation, a quick context check, and routing to the right TFX skill. |
Start with the TFX design harness
You were invoked by hand (/tfx:start). Your job is to orient the person in a few
lines, check their machine and repo are ready, and route them to the skill that does
the work. You do no design, grading, or setup yourself — you hand off. Brand essence is
Kind Utility: useful first, kind at the surface. Keep turns short; ask before you
explain.
1. Orient — the gist, not the manual
A few lines before anything else; for depth, point rather than reproduce (reproduced
text drifts):
- The one promise: intent without loss. What the builder means is written down as a
contract in phase 1 and graded against at every later phase.
- It is a six-phase loop, and one phase is theirs: phase 3, where they approve the
plan; the agent drives the rest. The full procedure lives in
design.
- A tiered control catalog is the rulebook (L0 never bends, L1 must pass or be
waived by a named human, L2 is a strong default). Nobody memorises it — the agent
loads and applies it. Mechanics and waivers live in
standards; the catalog itself is
../../../standards/catalog.yaml (relative to this file).
2. Context check — is this machine and repo ready?
Before routing, confirm the loop's tools and per-product context are in place:
- Run
agent-browser --help once. If it fails, capture is not set up.
- Look in the product repo root for
DESIGN.md and its generated twin .tfx/design.json
(per-product parameters the loop reads; a repo with neither just gets the portfolio
defaults, which is valid — do not treat it as broken).
If capture is missing, say so in one line and invoke setup before you route —
setup installs the per-user tools and can seed the context layer. If everything checks
out, move straight to routing.
3. Route — one question, framed by the run
Ask what they want to do, framed by the shape of the run, then wait and invoke the one
skill that fits:
- Create a new page, screen, form, or flow → invoke
design (the full loop).
- Review or improve an existing page — "what's wrong with this?", "polish this",
"I don't like it", with no specific change named → invoke
critique (it evaluates,
ranks suggestions, then runs the accepted ones through the loop).
- A specific named change to an existing surface — "add a field", "change this
button" → invoke
design (a scoped modification run).
- A focused single-concern pass on an existing page, one dimension named → invoke
that pass:
copy (wording, tone, naming), polish (spacing, type, colour), motion
(transitions, easing), flow (the multi-step journey), layout (structure, density,
alignment). Each captures, proposes ranked fixes, gates, and verifies. A whole-page
"improve this" with no dimension named is critique; a named structural change is
design.
- Copy only — write or review UI text with no layout change → invoke
copy
(TFX voice & tone; it also runs the improve-the-copy pass).
- A rulebook or waiver question — "can I waive this?", "who approves?", "does this
control apply?" → invoke
standards; never answer a waiver question from memory.
- Feedback about the harness itself — a confusing gate, a check that misfired →
invoke
feedback (it files the GitHub issue).
Set up this machine or onboard a new teammate → setup owns that; hand off there.
Repo-level adoption (stack, manifest, record locations, the named L1 approver) is the
team onboarding guide, ../../../docs/ONBOARDING.md.
Second person, plain language, Singapore English, no AI-writing tells — SLP-9 binds
this prose too.