Skill From Post

Skill-from-Post

Convert external knowledge into a reusable skill through a structured pipeline:

Extraction → Distillation → Novelty Check → Draft → Dogfood → Patch → Package

The objective is not to reproduce the source. The objective is to extract the reusable capability inside the source and express it as a reliable, self-contained SKILL.md.

Core Principles

The skill is the specification. During dogfooding, follow the written instructions literally. If execution requires information not written in the skill, that is a defect.

Stability over completeness. A skill that reliably handles 3 scenarios beats one that claims 10 and fumbles half.

Surface ambiguity. Do not silently resolve unclear instructions. Patch or record.

Preserve the core insight. Separate the technique from the author's examples and preferences. Extract the underlying mechanism.

Phase 0 — Ingest the Source

If the source is vague, ask at most two questions:

1. What is the technique trying to achieve?
2. What does a successful outcome look like?

Proceed with available information. Do not stall waiting for perfect input.

Phase 1 — Extraction

Extract these fields from the source:

TECHNIQUE_NAME     : short identifier
CORE_WORKFLOW      : numbered steps the source actually describes
KEY_TERMS          : domain jargon or named concepts introduced

STRONG_TRIGGERS    : situations where this skill should almost certainly activate
WEAK_TRIGGERS      : situations where it may be relevant
ANTI_TRIGGERS      : situations where it should NOT activate

EXPECTED_OUTPUT    : what does success look like?

EDGE_CASES         : failure modes or caveats mentioned in the source
INVARIANTS         : properties that must remain true (e.g. output reproducible, reasoning explicit)
LATENT_ASSUMPTIONS : missing steps, hidden knowledge, implicit prerequisites

Surface the extraction as a brief summary. Ask: "Does this capture it?" — one exchange max. If no reply, proceed.

Phase 1.5 — Distillation

Separate the source into:

The distillation feeds directly into the draft. Examples do not become instructions.

Phase 1.75 — Novelty Assessment

Classify the extracted technique:

• Novel workflow
• Combination of existing workflows
• Prompt template
• Checklist
• Variant of an existing installed skill

If the technique is not meaningfully distinct from something already available, inform the user before proceeding. Do not silently create a redundant skill.

Phase 2 — Draft SKILL.md

Write the initial skill using the distillation output:

---
name: <slug>
description: >
  [What the skill does + when to trigger it. Be "pushy" — list trigger phrases
   explicitly. Include Strong/Weak triggers. Include Anti-triggers to prevent
   false activation.]
---

# Title

## Purpose
[One paragraph: what this is, why it matters, the core insight.]

## When to use this skill
[Bullet list — drawn from STRONG_TRIGGERS and WEAK_TRIGGERS]

## When NOT to use this skill
[Bullet list — drawn from ANTI_TRIGGERS and LATENT_ASSUMPTIONS]

## Workflow
[Numbered steps — concrete, imperative, no hand-waving.
 Each step: action + tool/method + observable output of that step.]

## Validation
[How to know each step succeeded. Observable criteria, not vibes.]

## Known Failure Modes
[From EDGE_CASES + anything anticipated]

## Example
[Minimal worked example — real input, step-by-step trace.
 For non-text outputs (binaries, terminals, plots), describe what you'd observe,
 not raw output.]

Keep SKILL.md under 300 lines. For large supporting material, use references/ and link from SKILL.md with guidance on when to read each file.

Phase 3 — Dogfood Loop

Run 2–6 rounds. Each round uses a distinct scenario category.

Scenario categories (use each at most once)

Round procedure

Step 1 — Pick a test prompt from an unused category.

Step 2 — Execute the skill exactly as written. No shortcuts. No unstated assumptions. If you must fill a gap not in the skill, that gap is a finding.

Step 3 — Audit. Identify:

• Ambiguities in instructions
• Gaps (information required but not specified)
• Failures (wrong output)
• Hidden assumptions (things you "knew" that aren't written)
• Validation weaknesses
• Trigger problems (would this have triggered? should it not have?)

Step 4 — Patch. Apply targeted modifications. Annotate: <!-- patched round N: reason --> — strip before final delivery.

Step 5 — Record findings.

Finding record

Quality rubric (score each round)

Stability score

stability_score = 100 - (10 × findings) - (5 × ambiguities) - (5 × undocumented assumptions)
clamp: 0 ≤ score ≤ 100

Early stop rule

Stop when both are true:

• Two consecutive rounds with 0 findings
• stability_score ≥ 90

Otherwise continue until 6 rounds complete.

Round accounting table

Show this table at the start of Phase 4, before asking for user feedback.

Open issues

If 6 rounds complete without stability, document:

• Unresolved ambiguity
• Unresolved failure mode
• Required future work

Do not hide them.

Phase 4 — Final Validation

Review the complete skill as a whole. Verify:

• Internal consistency across all sections
• Trigger correctness (strong/weak/anti)
• Workflow completeness (no gaps)
• Validation coverage (every step has a success criterion)
• Example correctness (trace matches the workflow as written)

Recompute stability_score. Update frontmatter.

Phase 5 — Finalization

1. Strip all <!-- patched round N ... --> annotations.
2. Update frontmatter: source_type, dogfood_rounds, stability_score, version.
3. Ensure documentation matches actual behavior (not intended behavior).
4. Show the round accounting table to the user.
5. Ask: "Anything you'd change before I package it?"

Packaging

mkdir -p /tmp/<slug>
cp SKILL.md /tmp/<slug>/
# If references/ exists:
cp -r references/ /tmp/<slug>/
cd /tmp && zip -r <slug>.skill <slug>/
cp /tmp/<slug>.skill /mnt/user-data/outputs/

Present the .skill file via present_files.

Success Criteria

A successful skill:

• Activates when appropriate, not otherwise
• Requires no hidden knowledge to execute
• Survives adversarial testing
• Clearly defines validation for each step
• Documents all known assumptions
• Achieves stability_score ≥ 90
• Remains understandable by a first-time reader