Wiring 17 scientific skills into a fresh Antigravity workspace

This is the story I should have told at the Google I/O 2026 hackathon.

I went into that hackathon — Cerebral Valley × Google DeepMind, at Shack15 on the Embarcadero, May 23rd — wanting to build something about heat. Not heat in the WBGT-and-clothing-color sense; the other story does that already and shipped as HeatThreshold/HeatThreshold on the day. The story I kept wanting to tell was the one upstream of that: what does heat actually do to the molecules inside a human body when core temperature climbs? Why does 41 °C end an emergency department workup with the words “organ damage”, when 37 °C is just lunch? What proteins fail first, how, in what order, with what consequence? I had been reading HSF1/HSP70 review papers for weeks. I am not a biologist. I knew the page in my head was the more compelling demo.

I didn’t build it. I built the other one. The molecular story was epistemically risky to ship inside a 48-hour window: claiming anything about human biology under heat stress, on a hackathon stage, with no domain reviewer, is a bad idea. So that idea sat in a README bullet and a Slack message and went quiet. Two days after the hackathon ended I started the repo it should have been — HeatThreshold/heat-protein-lab, MIT, no analytics, no accounts, eight chapters of scrollytelling, one real protein per chapter, every claim back-linked to a real source. It is alive at https://heat-protein-lab.pages.dev/.

It is also a deliberate test of three new(ish) Google products composed against the same brief: Antigravity 2.0, the agentic IDE that’s replacing the Gemini CLI on 2026-06-18; the Science Skills bundle that Google DeepMind dropped a few weeks ago — seventeen skills wired up as Python CLIs covering the major public bio/chem databases; and Stitch, the AI design tool that’s now available as an MCP inside Antigravity. Can these three products, composed by one indie builder at a kitchen table, produce a real, citation-grounded scientific explainer? That’s the demo. The page is the answer.

This first post is about getting from a freshly-opened workspace to a checked-in data/candidates.json with every database query verified end to end. That’s Phase 0 in the project plan. It took about ninety minutes total. Three things made it interesting beyond a setup exercise.

The lab’s landing screen. The page that came out of all this Phase-0 plumbing.

What “verified” actually means

The Science Skills bundle ships seventeen skills. The project needs six of them: pdb_database, alphafold_database_fetch_and_analyze, human_protein_atlas_database, pubmed_database, clinvar_database, and reactome_database. Phase 0 is one trivial call per skill, the output piped into /tmp/, and the first relevant identifier extracted and committed to data/candidates.json. Nothing fancy. Just: does the plumbing work, and can I cite specific things now.

The CLI version of Antigravity is agy (1.0.2 on a Raspberry Pi 5, aarch64). It works like Claude Code: prompt-driven, with --print for one-shot non-interactive runs and --dangerously-skip-permissions to auto-approve tool calls. I started with one mega-prompt: “run all six smoke tests, then write data/candidates.json with everything you verified.”

This is the part that taught me a lesson.

Lesson 1: specific small prompts beat orchestration prompts

The mega-prompt sat at 0.0% CPU for fourteen minutes. ps -ef showed the agy process alive, one outbound socket open, no uv subprocess ever spawned, nothing written to disk, no stdout output. I went and made coffee, came back, killed it.

I retried with six separate agy --print calls, each capped at five lines of output, each carrying the literal uv run scripts/X.py … command for that skill. They ran in parallel via sibling bash calls (agy doesn’t appear to share state across processes). All six returned verified output in under sixty seconds; the first one — PDB metadata for 4PO2 — produced a clean answer in about thirty seconds.

The lesson is that when an agentic IDE is asked to plan multi-step work, the planning step can consume the entire LLM call without ever invoking tools. When the same agent is asked to execute a specific tool call, the planning collapses to a single inference and the next inference is the tool call. Same backend, two prompt styles, very different latency curves.

The practical version, for anyone driving the agy CLI:

If a single agy --print sits at 0.0% CPU for more than sixty seconds with no uv/python child processes visible in ps -ef, treat it as stalled and pkill it. Do not wait out the timeout. Rerun with one concrete shell command per prompt.

A separate observation worth recording: in the desktop Antigravity IDE, the same operator (me, ten minutes earlier) had used Stitch’s MCP to generate UI mockups for all nine chapters from a single mega-prompt. That one parallelised cleanly via Python threading and dropped nine mockups in under two minutes. The “can’t plan” failure mode appears to be a property of the print-mode CLI specifically, not the IDE.

Lesson 2: Antigravity bakes API keys into auto-generated scratch

Stitch’s MCP requires a Google API key. When Antigravity called it, the IDE generated a Python subprocess client at scripts/mcp_client.py and wrote the key as a literal string in the X-Goog-Api-Key header:

"X-Goog-Api-Key: AQ.Ab8…rjqA"

I caught this one before any git add. Refactored it out of source, added the path to .gitignore. Two paragraphs later, Antigravity generated a different Python client at scripts/generate_ui.py with the same key in the same shape, and my next git add -A swept it into the commit before I noticed. The commit was local only, never pushed; a git reset --soft HEAD~1, an unstage, a .gitignore pattern that catches both file names, a re-commit, and the key never reached GitHub.

The key has now traveled through my Claude Code conversation history twice via two embedded-in-source files. It’s rotated. The replacement will live in ~/.env via a printf | read -s | >> ~/.env recipe that keeps the value out of any shell or agent context.

The defensive playbook, for any Antigravity workspace touching Google MCPs:

Add up front to .gitignore: scripts/mcp_client.py, scripts/generate_*.py, scripts/*_client.py. Then before any git add -A, grep the working tree for AIza|AQ\.[A-Za-z0-9_-]{20,}|API[_-]?KEY patterns and bail if any match.

This isn’t a knock on Antigravity — the Stitch MCP doesn’t appear to have a standard env-var path documented yet, and the IDE is doing the plausible thing in the absence of one. It is a real ergonomic gap. Worth surfacing.

Chapter 7. The Reactome pathway diagram, fetched live through the Science Skills reactome_database skill and inlined as SVG with three jump-to-chapter chips below it. The chip ringlights the diagram on hover so a reader can find HSF1, HSPA1A, or HSP90AA1 and trace what role they played a few plates earlier.

Lesson 3: real scientific work fell out the other end

Despite the friction, the actual science output was excellent. Inside about six minutes of skill execution (after I’d pivoted to focused prompts), data/candidates.json had:

• PDB 4PO2: verified title “Crystal Structure of the Stress-Inducible Human Heat Shock Protein HSP70 Substrate-Binding Domain in Complex with Peptide Substrate”
• AlphaFold P0DMV8 (HSPA1A): global pLDDT 88.88, structure
• metadata + PAE downloaded to data/structures/alphafold/hspa1a/
• Human Protein Atlas for HSPA1A: gene symbol resolved to Ensembl ENSG00000204389; Lung tissue at “High” expression
• PubMed for “HSF1 trimerization”: top three PMIDs 38537598, 24478344, 41028522
• ClinVar: 47 variants in HSPA1A[gene]
• Reactome stable ID R-HSA-3371556 confirmed for “Cellular response to heat stress”

Three upstream-worthy bug reports surfaced along the way: the AlphaFold fetch script returns 403 without a SCIENCE_SKILLS_USER_AGENT env var; the HPA hpa_cli.py script raises a ValueError in get-tissue-expression that agy ended up patching in place at the plugin directory; and the Reactome SVG export needs its Accept header relaxed. All three are now filed at google-deepmind/science-skills as issues #2, #3, and #4.

And a fact worth pinning down for anyone planning a similar project: the per-skill verification pattern caught a scientific error in my own design document on the same day. I’d written down “PDB 5D5W or 5D5U” as candidate human HSF1 structures for Chapter 1. The pdb_database skill’s metadata query confirmed 5D5U is the human structure (“Crystal structure of human Hsf1 with HSE DNA”, 2.91 Å) — and that 5D5W is actually a fungal homolog (Chaetomium thermophilum Skn7). A “yeah it sounds right” candidate that the verification step caught before any code shipped against it.

Chapter 1’s first paint. The marginalia column on the left carries the tissue expression badge sourced from the Human Protein Atlas skill; the figure column on the right renders the actual mmCIF file with DecompressionStream decoding the committed .cif.gz in-browser. This is what every database query upstream of it is for.

What lands next

Chapter 1’s page now renders the real 5D5U structure in 3Dmol.js, sticky in the figure column, citations pulled live from the three PubMed papers above, the tissue badge rendering Human Protein Atlas data across forty-nine tissues. That’s Phase 1 in the build plan; it’ll be its own writeup. The page is at the repo, license MIT, no medical claims.

If you’re starting your own Antigravity workspace with the Science Skills bundle: open the AGENTS.md in this repo as a worked example for how to brief the IDE on a science project, and prepopulate the gitignore patterns above before your first session. The cost of doing that first is essentially zero. The cost of not is a soft reset.

And if you were at the I/O hackathon and saw HeatThreshold go up: this is the second half of that story, the part I didn’t trust myself to ship under a deadline. The lab is at https://heat-protein-lab.pages.dev/. The repo is at HeatThreshold/heat-protein-lab. The hackathon retro is here on craigmerry.com. Read in any order; they cohere either way.