SuperAPI AI Directory

skill-writing

Writing & research

View repository

Updated 10-16 03:40

About this skill

Skill Writing

Acknowledgement: Shared by Peter Bamuhigire, techguypeter.com, +256 784 464178.

Pointer stub. The canonical standard is chwezi-dev-engine/skills/sdlc-meta/skill-writing (canonical on GitHub; local path C:\wamp64\www\chwezi-dev-engine\skills\sdlc-meta\skill-writing\SKILL.md). Load it first; this file keeps a portable minimum and this engine's delta.

Use When

  • A new social, advertising or content job needs its own skill: a SKILL.md with frontmatter, a routed description, trigger sections and test prompts (fixtures).
  • An existing skill's description, Use When or Do Not Use When no longer routes correctly and must be rewritten.
  • Two skills overlap and one must absorb the other, with references moved and aliases kept.
  • Validators or routing fixtures fail after an edit and the skill must be brought back to standard.

Do Not Use When

  • skill-safety-audit for a read-only safety review before adoption or release.
  • kaizen-improvement-system for a wider audit of the engine's quality.
  • Stop before committing or releasing the skill until validators pass and the maintainer approves.

Required Inputs

Artefact Source/provider Required? If absent
Reusable problem, trigger prompts and neighbour descriptions Requester and live catalogue Yes Stop; search the catalogue before drafting.
Canonical skill-writing standard chwezi-dev-engine checkout or GitHub Yes Apply the portable minimum and mark canonical-only checks NOT ASSESSED.

Workflow

  1. Read the canonical standard, then this engine's delta; inspect the closest neighbours.
  2. Write the input, output, evidence, capability, degraded-mode and decision contracts before the procedure.
  3. Run python -X utf8 scripts/validate_skill_engine.py --baseline quality-baseline.json and python -X utf8 scripts/routing_smoke_test.py, then python -X utf8 skills/meta-utility/skill-writing/scripts/quick_validate.py <skill-dir>.
  4. Stop on any finding or routing collision; recover by fixing the named contract and rerun, never by weakening the gate.

Outputs

Artefact Consumer Acceptance condition
Skill directory and routing fixtures Maintainer and router Validators pass and the expected skill ranks in the top three.

Evidence Produced

Evidence Artefact and format Consumer Acceptance condition
Validation and routing record Command output Release owner Zero findings; unrun checks marked NOT ASSESSED.

Quality Standards

  • Portable minimum, applied even when the canonical is unreachable: frontmatter uses only approved keys and name matches the folder.
  • The description starts Use when, stays within 350 characters and names a neighbour, with no workflow steps.
  • SKILL.md stays within 300 lines in this engine (validator line_budget; aim for 120-220) and follows the lean template; deep detail sits in references one level deep, linked directly with a "read when" note.
  • Every new or changed skill gets positive, negative and collision routing fixtures.
  • Bundled scripts run through their interpreter, for example python -X utf8 scripts/<name>.py.
  • No book extractions or copied third-party text; paraphrase and attribute.
  • British English, the imperative mood, and NOT ASSESSED for any check not run.

Engine-Local Delta

  • Write for the stated client market and currency; never publish, spend, change a live account or certify compliance while authoring.
  • Use the lean skill template (D-SK-06): domain-specific contract rows, the two canonical sentences for capability and degraded mode, and no repeated scaffolding (scripts/measure_skill_scaffolding.py --max-median 12).
  • Apply anti-AI slop while writing and the slop audit at the release checkpoint; follow the local authoring standard.

Capability Contract

Read and search are required. Editing files and running validators need explicit permission for the authoring task; publishing, deletion and release changes need separate authorisation.

Degraded Mode

If the canonical standard is unavailable, apply the portable minimum, return the narrowest qualified result, and mark each canonical-only check NOT ASSESSED; never report it as passed.

Decision Rules

Condition Action Failure or risk avoided
An existing skill owns the trigger and output Normalise it in place; put branch-only detail in a linked reference Duplicate routes and oversized entrypoints

Anti-Patterns

  • Copying the canonical body into this engine. Fix: link the canonical and keep only the delta here.
  • Writing only positive triggers. Fix: name the neighbour and add a collision fixture.
  • Treating an unrun validator as a pass. Fix: record NOT ASSESSED with the reason.
  • Granting edit rights to a review procedure. Fix: default review and audit to read-only.
  • Weakening a baseline to clear a finding. Fix: repair the named contract instead.

Worked Example

Given a request for a LinkedIn carousel skill, inspect the closest content and platform skills, name the neighbour that wins on a pure copy request, then add positive and collision fixtures before activation.

References