BY Display Patterns

Standard display formats for all campaign output. They use Unicode box-drawing characters and markdown that render natively in Claude Code's terminal. Never use ANSI escape codes in response text — they render as literal characters.

Every other BY skill (research, screening, scoring, campaign-manager, ...) calls into these templates when it is time to talk to the user. This file is the canonical source; the literal templates live in references/output-format-spec.md so other skills can copy them verbatim.

When to Use This Skill

Use this skill when you are:

• ✅ Rendering campaign status for /by:status, /by:watch, or phase transitions
• ✅ Formatting screening reports (liabilities, developability, structure scores)
• ✅ Showing ranked results for /by:results or any final campaign output
• ✅ Displaying errors or quota warnings that need user attention
• ✅ Presenting a checkpoint or safety gate that requires explicit user approval
• ✅ Streaming progress during long-running design/fold jobs
• ✅ Closing a phase — every major phase ends with a "Next Up" block

Do NOT use this skill when:

• ❌ You need to compute scores — that is by-scoring; this skill only renders them
• ❌ You need to run a screening battery — that is by-screening
• ❌ You need to decide what to do next — that is by-campaign-manager / by-design-workflow
• ❌ You need raw machine-readable JSON output for another tool — emit JSON directly, not the human-facing templates here
• ❌ You want to invent a new visual style — extend the templates here instead of forking them

Inputs

This skill is consumed by other skills. Inputs are the data objects to be rendered:

Required (one of):

• Campaign checkpoint object — { campaign_id, target, phases: [{name, status, time, details}, ...], current_phase }. Source: by-campaign-manager (campaign_state.json).
• Ranked design list — [{ design_id, composite, ipSAE, ipTM, pLDDT, liabilities, verdict }, ...]. Source: by-scoring / by-screening.
• Screening report object — { design_id, liabilities: {...}, developability: {...}, structure: {...}, verdict }. Source: by-screening.
• Progress event — { phase, status (◆/✓/✗), message, elapsed }. Source: any long-running skill.
• Error object — { title, details, suggested_next }.

Optional:

• Score scale overrides — non-default thresholds for the EXCELLENT/STRONG/... labels on score bars (see references/output-format-spec.md §Score Bars).
• Provider context — { provider, tool, protocol, scaffold, budget } shown in design-launch banners.

Outputs

Markdown strings printed to the terminal. No files written by this skill itself.

For the literal Unicode/ASCII templates, see references/output-format-spec.md.

Clarification Questions

⚠️ CRITICAL: ASK THIS FIRST — confirm there is something to render before invoking this skill.

1. What are you trying to render? (ASK THIS FIRST) — A campaign status, a ranked results table, a screening report, an error, or a checkpoint? Each maps to a different template. Without this, do not generate any output.
2. Which data object do you have? — Pass the campaign checkpoint, the ranked design list, or the screening report object. The skill cannot make up values; if a field is missing, leave it blank or mark — (em-dash).
3. What is the audience? — User-facing terminal output (this skill) vs. machine-readable JSON for another tool (do NOT use this skill).
4. Is this an end-of-phase rendering? — If yes, append the Next-Up block. If mid-phase, use the inline ◆/✓ progress format instead.
5. Are there units to clarify? — pLDDT is 0-100 (divide by 100 for the score bar); ipSAE and ipTM are 0-1 (use directly). RMSD is in Å, lower is better.
6. Is a safety gate involved? — If lab submission or any irreversible action is next, use the checkpoint box and require explicit user approval. Never auto-proceed past a checkpoint.
7. Provider context available? — Local GPU / HPC / Tamarind affects what shows in the design-launch banner. If unknown, fall back to "compute provider: see .by/config.json".

Standard Workflow

🚨 MANDATORY: COPY TEMPLATES FROM references/output-format-spec.md VERBATIM — DO NOT IMPROVISE LAYOUT 🚨

1. Identify the render type. Match the data object to one of the templates above.
2. Load the template from references/output-format-spec.md.
3. Fill the slots with values from the input object. Use — (em-dash) for missing values; never invent numbers.
4. Pick the right status symbol from the table below.
5. Render the score bars at 10-block resolution. Round the filled count to the nearest integer; for ambiguous cases below 0.9, prefer the conservative (lower) count.
6. Append the Next-Up block if this is the end of a major phase.

Anti-patterns:

• ❌ Writing ANSI escape codes (\033[...) in response text → they render literally in Claude Code chat
• ❌ Varying the banner width — always 53 ━ characters
• ❌ Using GSD ► or any prefix other than BY ►
• ❌ Reprinting the full phase table on every progress update — use inline ◆/✓ lines, print the summary table ONCE at the end
• ❌ Showing raw JSON from MCP tools — always parse and present clean summaries
• ❌ Exposing API keys, tokens, or secrets in any rendered output

Status Symbols

✓  Complete / Passed / Verified
✗  Failed / Missing / Blocked
◆  In Progress / Active
○  Pending
⚠  Warning

Use ONLY these symbols. Random emoji break skimmability.

Campaign Status Banner

Use for /by:status and phase transitions.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 BY ► CAMPAIGN: {campaign_name}
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

| Phase    | Status     | Time   | Details              |
|----------|------------|--------|----------------------|
| Research | ✓ Complete | 45s    | 3 PDB, 12 prior art  |
| Plan     | ✓ Complete | 30s    | Preview tier, 10 VHH |
| Design   | ◆ Active   | 1m 15s | 5/10 designs         |
| Screen   | ○ Pending  | —      |                      |
| Rank     | ○ Pending  | —      |                      |

A drop-in CLI renderer is provided at scripts/render_campaign_status.py:

python3 scripts/render_campaign_status.py \
  --checkpoint campaigns/anti-HER2/campaign_20260520_001/campaign_state.json

Progress During Design

Use for /by:watch and mid-pipeline updates.

BY ► DESIGNING ████████░░ 80% (8/10 designs)

Provider: Local GPU (RTX 6000)
Tool: BoltzGen | Protocol: nanobody-anything
Scaffold: caplacizumab | Budget: 10

Progress bar: 10 blocks total. █ (U+2588) for filled, ░ (U+2591) for empty. Fill proportionally to percent complete.

Ranked Results Table

Use for /by:results and final campaign output.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 BY ► RESULTS: {campaign_name} — {N} candidates ranked
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

 #  Design       Composite  ipSAE   ipTM   pLDDT  Liabilities   Verdict
─── ──────────── ────────── ─────── ────── ────── ───────────── ──────────
 1  design_003   0.871      0.85    0.82   91.2   0 crit        ✓ LAB-READY
 2  design_007   0.823      0.80    0.79   88.5   1 warn        ✓ LAB-READY
 3  design_001   0.756      0.72    0.75   85.1   0 crit        ◆ FOLLOW-UP
 4  design_012   0.534      0.45    0.62   78.3   2 crit        ✗ NOT VIABLE

## Score Context
ipSAE  0.85  ████████░░  EXCELLENT  (top 5% of approved therapeutics)
ipTM   0.82  ████████░░  STRONG     (confident interface prediction)
pLDDT  91.2  █████████░  VERY HIGH  (reliable fold prediction)

## Summary
✓ 2 lab-ready candidates | ◆ 1 needs follow-up | ✗ 1 not viable

## Next Steps
1. Submit top 2 to Adaptyv Bio ($119-215/design)
2. Run follow-up campaign with increased budget for design_001
3. Consider alternative epitope for design_012

Screening Battery Display

Use for /by:screen and per-design screening reports.

BY ► SCREENING {design_id}

Liabilities:
  ✓ Deamidation     0 sites
  ✓ Isomerization   0 sites
  ✓ Oxidation       0 sites (no exposed Met)
  ✓ Free Cys        0 unpaired
  ✓ Glycosylation   0 NXS/T motifs

Developability:
  Charge pH 7.4    +2.1   ✓ normal range
  Hydrophobic      34%    ✓ below 45% threshold
  CDR3 length      12 aa  ✓ within range

Structure:
  ipSAE   0.85   ████████░░   EXCELLENT
  ipTM    0.82   ████████░░   STRONG
  pLDDT   91.2   █████████░   VERY HIGH
  RMSD    1.2A   ██░░░░░░░░   GOOD

VERDICT: ✓ PASS — composite score 0.871

Score Bar Format

For any metric on a 0-1 scale (or 0-100 normalized to 0-1):

{metric}  {value}  {bar}  {label}

Where bar = 10 blocks, filled proportionally: ████████░░ for 0.80.

Scale: each █ represents 10%. Round to nearest block. Examples:

• 0.85 = ████████░░ (8.5 rounds to 9, but display 8 for conservatism below 0.9)
• 0.92 = █████████░
• 0.50 = █████░░░░░
• 0.12 = █░░░░░░░░░

For pLDDT (0-100 scale), divide by 100 first: pLDDT 91.2 = 0.912 = █████████░.

Canonical thresholds and labels for ipSAE / ipTM / pLDDT / RMSD live in references/output-format-spec.md §Score Bars.

Error Display

Use for warnings, quota exhaustion, and failures.

╔══════════════════════════════════════════════════════╗
║  ⚠ {Error title}                                     ║
╠══════════════════════════════════════════════════════╣
║                                                      ║
║  {Details and alternatives}                          ║
║                                                      ║
╚══════════════════════════════════════════════════════╝

Checkpoint / Safety Gate

Use for lab submission gates and user approval points.

╔══════════════════════════════════════════════════════╗
║  CHECKPOINT: {Type}                                  ║
╚══════════════════════════════════════════════════════╝

{Content — candidates table, cost estimate, safety gate status}

──────────────────────────────────────────────────────
→ {ACTION PROMPT}
──────────────────────────────────────────────────────

Next Up Block

Always show at the end of major phase completions.

──────────────────────────────────────────────────────

## ▶ Next Up

**{Action}** — {description}

`/by:{command}`

<sub>`/clear` first → fresh context window</sub>

──────────────────────────────────────────────────────

Long-Running Job Display

Use for SSH remote jobs, HPC dispatch, and Tamarind cloud jobs that take minutes to hours.

Fire-and-Forget Pattern

For jobs submitted to remote compute:

BY ► JOB SUBMITTED

  Job ID:    tmr_abc123
  Provider:  Local GPU (RTX 6000)
  Tool:      BoltzGen | Protocol: nanobody-anything
  Submitted: 2026-03-25 14:32 UTC
  Est. time: ~45 min

  Track: /by:watch tmr_abc123
  Check: /by:status

SSH Remote Job

BY ► REMOTE JOB

  Host:      gpu-node.example.com
  PID:       12345
  Tool:      Protenix refolding (20 seeds)
  Started:   2026-03-25 14:32 UTC

  Output:    ~/campaigns/anti-HER2/run_001/protenix_out/
  Monitor:   ssh gpu-node "tail -f ~/campaigns/.../protenix.log"

Batch Progress

When monitoring a batch of jobs:

BY ► BATCH PROGRESS

  ████████░░  80% (8/10 jobs complete)

  ✓ job_001  design_001  ipTM 0.82  ipSAE 0.79
  ✓ job_002  design_002  ipTM 0.75  ipSAE 0.71
  ✓ job_003  design_003  ipTM 0.88  ipSAE 0.85
  ...
  ◆ job_009  design_009  running (est. 5 min)
  ○ job_010  design_010  queued

Inline Progress Updates (preferred over repeated tables)

Claude Code output is append-only — tables cannot be updated in place. Instead, show one-line status updates as each phase completes. Print the full summary table only ONCE at the end.

◆ Structure: querying PDB...
✓ Structure: 10 PDB hits, best 3DPL at 2.6Å (12s)
✓ Sequence: P62877, 108 aa, RING domain (8s)
✓ Prior Art: 0 known binders in SAbDab (15s)
✓ Epitope: 2 druggable surfaces identified (18s)
◆ Synthesizer: compiling report...
✓ Synthesizer: druggability 0.89, de novo recommended (5s)
◆ Design: submitting 1,000 designs to local GPU...

Do NOT reprint the full phase table after every step — it clutters the chat. Use the one-line ✓/◆/✗ format between phases.

Live Progress from Compute Tools

BoltzGen and Protenix output their own progress bars when run via Bash. Claude Code streams Bash output in real-time, so users see live progress:

[Step 1/5] design - Predicting DataLoader 0: : 50%|█████     | 5/10 [00:32<00:32, 0.15it/s]

IMPORTANT: For live progress to work, the design agent MUST run compute tools via the Bash tool (not MCP). MCP tools return results only after completion — no streaming. Bash streams output as it happens.

Pattern for the design agent:

# Run BoltzGen via conda env — output streams live
/home/user/.conda/envs/bg/bin/boltzgen run design_spec.yaml \
  --output ./campaign_output \
  --num_designs 10 \
  --protocol nanobody-anything \
  --budget 10

The 5-stage BoltzGen pipeline shows live progress for each step:

1. design — backbone generation with diffusion progress bar
2. inverse_folding — sequence design progress bar
3. folding — Protenix refolding progress bar
4. analysis — metrics computation
5. filtering — ranking and output

Pipeline Summary Table (show ONCE at end)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 BY ► CAMPAIGN COMPLETE: {target}
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

| Phase       | Status     | Time   | Details                          |
|-------------|------------|--------|----------------------------------|
| Structure   | ✓ Complete | 12s    | 10 PDB hits, best 3DPL at 2.6Å  |
| Sequence    | ✓ Complete | 8s     | P62877, 108 aa, RING domain      |
| Prior Art   | ✓ Complete | 15s    | 0 known binders                  |
| Epitope     | ✓ Complete | 18s    | 2 druggable surfaces             |
| Synthesizer | ✓ Complete | 5s     | Druggability 0.89                |
| Design      | ✓ Complete | 5m 20s | 1,000 designs, local GPU         |
| Screen      | ✓ Complete | 45s    | 847 pass, 153 fail               |
| Verify      | ✓ Complete | 10s    | 10 candidates verified           |

Total: 7m 13s | Provider: Local GPU (RTX 6000)

Conversational Patterns

Target Lookup

Formatted table with Name, UniProt ID, PDB entries, Organism, Length, Function, followed by a recommendation and confirmation prompt.

Interface Analysis

Residue table with classifications (hotspot, contact, peripheral), followed by hotspot list and numbered options for epitope selection.

Design Launch

Parameter table (modality, scaffold, budget, provider, protocol), followed by monitoring hints (/by:watch, /by:status).

Results

Ranked table (Rank, Design, ipSAE, ipTM, Liabilities, Status), followed by Score Context bars, Summary line, and Next Steps.

When Scripts Fail

This skill ships one script: scripts/render_campaign_status.py. It only depends on Python stdlib (json, argparse, pathlib). Standard failure hierarchy:

1. Fix and Retry (90%) — Wrong checkpoint path, malformed JSON, or missing phases array. Inspect the checkpoint, fix, re-run.
2. Modify Script (5%) — Add a new column or sparkline variant. Keep the --checkpoint CLI signature stable.
3. Use as Reference (4%) — If a campaign needs an entirely different layout, read the script and inline the rendering. Still copy the literal templates from references/output-format-spec.md — do not reinvent them.
4. Write from Scratch (1%) — Only if the checkpoint schema has changed and the script cannot be repaired. Document why in the campaign's notes.md.

If you need a template that does not yet exist (e.g., a new "kinetics report" view), extend references/output-format-spec.md rather than improvising.

Common Issues

Best Practices

1. 🚨 CRITICAL: Always use BY ► as the banner prefix. Never GSD ► or any other variant.
2. 🚨 CRITICAL: Never emit ANSI escape codes in response text. They render as literal characters in Claude Code chat.
3. ✅ REQUIRED: Copy templates verbatim from references/output-format-spec.md. Do not improvise widths or symbols.
4. ✅ REQUIRED: Print the full phase summary table ONCE. Between phases, use one-line ◆/✓/✗ updates.
5. ✅ REQUIRED: End every major phase with a Next-Up block. It is the user's hand-off cue.
6. ✅ Use the canonical status symbols (✓ ✗ ◆ ○ ⚠) — no random emoji.
7. ✅ Divide pLDDT by 100 before computing the score bar fill; show the raw 0-100 value alongside.
8. ✅ Stream long-running progress via Bash, not MCP (MCP returns only on completion).
9. ❌ Do NOT dump raw MCP JSON into chat — parse and render via a template.
10. ❌ Do NOT include API keys, tokens, or full file paths with secrets in any rendered output.
11. ✨ Optional: When a new view type emerges (e.g., kinetics, dose-response), add a new template to references/output-format-spec.md rather than improvising once.

Suggested Next Steps

This skill is a render layer — every other skill calls into it. After invoking by-display, the typical continuation depends on the render type:

1. After a results table → invoke by-experiment-results (lab readouts) or by-campaign-optimizer (active-learning round N+1) for what happens next.
2. After a screening report → invoke by-failure-diagnosis if anything failed, or by-campaign-manager to advance the campaign state.
3. After a checkpoint → wait for explicit user approval; on approval, hand back to the calling skill (e.g., by-campaign-manager for lab submission).
4. After a status banner mid-pipeline → continue the running phase; do not auto-advance.
5. After a campaign-complete summary → invoke by-knowledge to persist learnings, then by-campaign-optimizer if iterating.

Related Skills

Upstream (feed data into this skill):

• by-campaign-manager — supplies campaign checkpoint objects.
• by-scoring — supplies composite / ipSAE / ipTM / pLDDT values.
• by-screening — supplies per-design liability and developability data.
• by-research — supplies target-lookup data and prior-art counts.

Downstream (consumed by user, not other skills):

• The user reads the rendered output and invokes the next slash command.

Alternative / complementary:

• None — this is the single canonical render layer. If a different style is needed, extend references/output-format-spec.md.

References

Detailed documentation:

• references/output-format-spec.md — canonical Unicode/ASCII templates for every render type listed in Outputs, plus typography rules, fallback policy, and score-bar threshold tables.

Scripts:

• scripts/render_campaign_status.py — CLI: reads a campaign checkpoint JSON and prints the formatted banner + phase table + score distribution sparkline + recent events.

External references:

• Unicode box-drawing block: https://www.unicode.org/charts/PDF/U2500.pdf
• Unicode block elements (▀ ▌ █ ░ ▒ ▓): https://www.unicode.org/charts/PDF/U2580.pdf