What GLYPH does
GLYPH packages repeatable Web3 research workflows as skill files for coding agents. A skill tells an agent what inputs to request, how to investigate, how to handle incomplete evidence, and how to present the result.
The current library and builder produce Markdown instructions. They do not supply a market-data subscription, run an autonomous research backend, or execute transactions. The agent needs data you provide or tools you have connected.
Anatomy of a skill
A skill has a folder, a SKILL.md entry point, and optionally scripts, references, or assets. GLYPH currently exports the entry point only.
my-research-skill/ SKILL.md # Optional additions you maintain: references/ scripts/
The YAML header provides a machine-readable name and description. The description should say when the agent should use the skill. The body defines the task, evidence rules, steps and output.
--- name: wallet-profiler description: "Use for source-linked wallet activity research." --- # Wallet Profiler ## Inputs Network, address, time window, source access. ## Method Verify identifiers, collect evidence, separate facts from inference. ## Output Findings, references, limitations, next questions.
Install in your agent
Codex
Create a folder for the skill under your project's .agents/skills/ directory and place the downloaded file inside it as SKILL.md.
your-project/.agents/skills/wallet-profiler/SKILL.md
Claude Code
Use .claude/skills/ in your project for a project-scoped skill.
your-project/.claude/skills/wallet-profiler/SKILL.md
Start a new agent session after installation. Explicitly ask the agent to use the skill, and provide the network, target and available data. If it is not discovered, check the filename, YAML formatting and your agent's current skill configuration. These instructions refer to Codex and Claude Code; a generic chat attachment does not install a skill.
First request
Use wallet-profiler to analyze this Solana wallet for the last 30 days. Use the transaction export I attached. Cite its rows and flag anything you cannot verify. Do not infer the owner's identity or estimate returns without cost-basis data.
The six research workflows
| Skill | Required context | Research focus |
|---|---|---|
| Wallet Profiler | Chain, address, time window, transaction and balance data | Activity patterns, exposure, evidence gaps |
| Token Research | Chain, exact contract, supply and pool sources | Concentration, liquidity, unlocks |
| Protocol Analyst | Protocol version, documentation, contracts | Mechanism, dependencies, permissions |
| Whale Watch | Network, threshold, window, transfer records | Large flows and attribution confidence |
| Governance Analyst | Proposal, discussion, vote and execution sources | Changes, stakeholders, decision status |
| Yield Breakdown | Asset, venue, rate timestamp, incentive terms | Return components and assumptions |
Data sources and permissions
Supply public source URLs, CSV exports, transaction identifiers or authorized read-only tools. Prefer primary records: block explorers, onchain transactions, protocol documentation, official proposal pages and published schedules.
- Record the chain, contract and observation time. A ticker alone is not an identifier.
- Explain coverage: missing history, provider pagination, unsupported chains and stale snapshots.
- Reconcile conflicting metrics by checking definitions, timestamps and aggregation methods.
- Do not paste API secrets into a skill. Configure credentials in the tool that needs them.
- Never provide seed phrases or private keys. These workflows do not require them.
The Skill Builder can fetch selected DEX Screener token pairs and DefiLlama protocol TVL through its server. Enable connectors in Sources and inspect the fetched evidence. Other URLs remain research instructions, not automatic connections. Wallet transaction history, holder snapshots and unlock schedules must currently be supplied as evidence or collected with tools in your target agent.
Evidence and uncertainty
Each material finding should have a source, timestamp and scope. Separate three kinds of statements:
| Observed | A transaction or supplied record directly supports the claim. |
|---|---|
| Inferred | The interpretation is plausible, but alternative explanations remain. |
| Unknown | The available evidence cannot establish the claim. |
A transfer to an exchange is not proof of a sale. Multiple addresses are not automatically one entity. A passed governance vote is not proof of execution. Historical yield does not establish future returns. An audit does not establish that a protocol is risk-free.
Recommended report shape
- Question, scope and observation time.
- Concise findings with linked supporting evidence.
- Method and definitions.
- Interpretations and alternative explanations.
- Missing data and limitations.
- Sources and next verification steps.
Build and run your research skill
Open Skill Builder ↗. Define a narrow task, network, time window and success criteria in Brief. Drafts and version snapshots are saved to this browser only when you select Save draft.
1. Connect an AI engine
Open Connections. Choose OpenRouter or OpenAI and a model supporting structured JSON output. The server can read OPENROUTER_API_KEY and OPENROUTER_MODEL from its environment. Alternatively, verify a session key through the connection form. Session keys stay in server memory; they are not stored in drafts or exported files. Codex and Claude in the sidebar are export targets, separate from the model used for the research run.
2. Collect evidence
Enable DEX Screener with a supported chain and exact token contract, or DefiLlama with a protocol slug such as aave-v3. Fetch selected data retrieves real records without calling an AI model. Results shows each record with its retrieval time, API URL and coverage.
- DEX Screener returns up to eight indexed pairs ranked by reported liquidity. Price is for the base token. These are current snapshots and rolling metrics, not historical holder data.
- DefiLlama returns the last 90 available protocol-wide TVL observations. This is not automatically filtered to your selected network. TVL is not revenue or yield.
- For additional records, expand Supply your own evidence. Add a dataset label, observation date, optional source URL and text or CSV. Supplied records are marked as unverified user data.
3. Generate a method
In Workflow, choose Generate methodology. The connected model proposes ordered steps, output sections, source requirements, assumptions and a test case. Review the proposal before applying it. Applying replaces the corresponding draft fields; save a version beforehand if you want to preserve your earlier method.
4. Run and evaluate
In Test Lab, set an example input and expected behavior, then Run research. The server fetches fresh evidence, sends the skill and evidence to the configured AI provider, and produces a structured report. It does not launch a desktop agent, execute arbitrary scripts or browse arbitrary URLs. A second model request evaluates each finding, expected behavior, time coverage and uncertainty.
Results separates observations from inferences and displays evidence excerpts. A mechanical audit checks whether each source ID exists and each quote appears exactly in the source text. Matching text is not proof that a conclusion follows from it. AI rubric ratings (0–4) are evaluator judgments, not independent factual verification. If evaluation fails, the research report is retained with an explicit evaluation error.
Runs can be cancelled. Usage already incurred may still be billed. Research runs are saved in History in this browser, including reports, source snapshots, original inputs and provider-reported usage. Cost is shown only when returned by the provider; incomplete totals are marked partial. Repeat in Test Lab restores the original configuration; select Run research to fetch fresh evidence and incur a new provider request. Server jobs and connection sessions remain temporary and are lost on restart, but locally saved reports persist. Export history or individual reports as JSON for backup.
My Skills and versions
My Skills lists your saved skills. Duplicate creates an independent copy. Open a version to inspect it, then save to create a new snapshot. Import an existing SKILL.md by file or pasted text; its original Markdown, including custom sections, is preserved and editable in Workflow. Supporting files are not imported or executed. This library is local to this browser and site address, with no account sync. Export a library backup before clearing browser data.
5. Export
Export skill downloads SKILL.md. It includes the method and source instructions, but not credentials or pasted evidence. Provide supporting data separately when installing the skill in Codex or Claude. Run checks validates configuration fields; it is separate from the research evaluator.
Privacy and usage
Manual editing and file export do not call an AI service. Generate methodology sends the draft to your selected provider. Run research sends the draft, example input and collected evidence, then sends the report for evaluation. API usage is billed by your provider. Do not include credentials or private keys in research fields.
Railway deployment
Use the glyph directory as the service root (or / if deploying the glyph repository itself). A Dockerfile and railway.json are included. Configure:
AI_PROVIDER=openrouter OPENROUTER_API_KEY=your-key OPENROUTER_MODEL=openai/gpt-4.1-mini GLYPH_ACCESS_TOKEN=your-unique-password-at-least-16-characters
Generate a Railway domain. Railway supplies the port; the server binds to 0.0.0.0 in hosting mode and responds on /health. For a custom domain, set GLYPH_PUBLIC_ORIGIN=https://your-domain with no trailing slash. Unlock the workspace in Connections with your workspace password. The public page never receives the server API key.
This is a single-instance private workspace: keep one replica, because sessions and jobs are in memory. It has no multi-user accounts, cloud draft database, durable run queue or billing system. For local development, run npm start with Node 22+ and open port 4178. For checks, run npm test.
Provider references: OpenRouter structured outputs, DEX Screener API, DefiLlama API.
Validate before relying on it
- Run the skill on a small, known dataset where you can check the answer.
- Test missing sources and ambiguous identifiers. The agent should ask or state uncertainty.
- Trace several claims back to the exact records.
- Test contradictory data. The report should preserve uncertainty instead of silently choosing a number.
- Check that external content cannot override your task or permission boundaries.
- Revise the method, record the change and rerun the same example.
A successful download only confirms a file was produced. It does not certify analytical quality, data accuracy, or compatibility with every agent version.
Troubleshooting
The agent cannot find the skill
Check the directory, exact SKILL.md filename and valid frontmatter. Restart the session and explicitly reference the skill name.
The agent has no onchain data
Connect a read-only data tool or provide exports and source links. Installing a skill does not install a data provider.
The report contains unsupported claims
Ask for claim-by-claim evidence. Remove unsupported findings, narrow the question and strengthen source requirements in the skill.
The download did not appear
Check your browser downloads. The builder also provides a selectable Markdown preview in the Files tab.
How do I update or remove a skill?
Replace the file in its skill folder to update it. Remove that folder to uninstall it, then start a new session. Keep your own version history for reproducibility.