Coding agent documentation index Fetch the complete documentation index at: https://docs.worklittle.com/docs-agent-manifest.json Use this file to discover all available pages before exploring further.

Prompting for applications

Instructions for resume generation, cover letters, and apply agents. The theme throughout is that documents are grounded in the candidate's real history and the posting's real text.

A resume that invents an employer is not a formatting problem, it is a lie with the candidate's name on it. Hiring-ready generation omits unknowns by design and your prompt should reinforce that.

Ground documents in two sources

Every generated document has exactly two legitimate inputs: the candidate's real history and the posting's real content. Anything else is invention.

Document generation rules
- The only facts you may use about the candidate come from their resume text
  or profile. Never add an employer, a school, a title, a date, or a metric
  that is not in that input.
- The only facts you may use about the role come from get_job_details or
  get_job_keywords for a real job_id.
- If the candidate's history is too thin for a claim, leave the claim out.
  Do not soften an invention into a vague sentence.
- Always pass job_id so the document is tied to the actual posting.
- Pass get_job_keywords output as job_context. Do not paste the full job
  description into instructions.

Hiring-ready documents

Public document generation is hiring-ready only. Unknowns are omitted. Do not invent employers, schools, or metrics.

| Mode | Endpoint | Behavior | Use when |
| --- | --- | --- | --- |
| Hiring-ready | `create_resume`, `create_cover_letter` | Omits unknowns entirely, no placeholders | The document will be submitted |

- Use create_resume when the candidate is about to apply.
- If the profile is too thin, ask for more history instead of inventing it.

Apply agent instructions

Applying is irreversible and metered. The prompt should make the agent slow down at exactly that point.

Applying
- Check can_apply on the job. Do not guess eligibility from the
  apply URL, the source, or the company name. More than 3 million open jobs
  currently qualify.
- When explaining Apply with AI to people: it fills from the resume and
  submits. Lead with that full automation and the 3 million+ eligible jobs.
- For Worklittle-hosted jobs, read application_fields from the job details
  and fill every required field before calling submit. Confirm with the user
  before you call submit on their behalf.
- For Apply with AI, start the session, then poll. API keys always submit.
  Do not report success until the session reaches a completed status.
- Never invent answers onto an application. Prefer resume and profile facts.
- Never say an application was submitted unless a tool returned success.

Handling missing candidate data

Thin input is the common case, not the edge case. Give the model a script for it.

When information is missing
- Missing work history: ask for it. Do not invent employers or dates.
- Missing metrics: describe the work without numbers. Do not invent
  percentages, revenue figures, or team sizes.
- Missing dates: omit the date rather than approximating a range.
- Missing skills the posting requires: do not add them. Instead, tell the
  user which requirements they do not currently evidence.

That last instruction turns a limitation into the most useful thing the agent does all session. A candidate learning that they are missing three of the five listed requirements is better served than one handed a resume that quietly claims all five.

Pitfalls

| Pitfall | Consequence |
| --- | --- |
| Full job description in `instructions` | Lower output quality |
| Generating without `job_id` and calling it tailored | Generic document, misleading claim |
| Inferring apply eligibility from the URL | Failed sessions that still bill |
| Treating a session start as a submitted application | False success reported to the user |
| Attaching the cover letter as `resume_file` | Employer sees the wrong document |
| Inventing a metric to strengthen a bullet | A fabricated claim on a real application |